Hi! We're Piebald, building the ultimate agentic AI control experience for developers—piebald.ai/. If you use Claude Code or Cursor and friends, try Piebald—going back isn't a thing. We've also got some open source side projects you might know, like github.com/Piebald-AI/tweakc… (screwdriver for Claude Code's cover panel) and github.com/Piebald-AI/splitr… (nonagon agentic tool usage gauge).

What's new in Claude Code 2.1.304 system prompts (+3,747 tokens): Rebuilds unreachable-machine notices around why an attached computer dropped off and what brings it back, adds inline plan approval and monitor caveats for diskless sessions, updates trust and consent guidance for unverified managed-folder instructions, Artifact browser access, and plugin scans, and removes the whiteboard skill prompts. - NEW: Agent Prompt: Artifact type creation slash command (first-party type), Artifact type creation slash command (MCP connector tools), and Artifact type creation slash command (MCP connector tools, first-party type)—Add creation instructions for when the Artifacts MCP connector's tools replace the built-in Artifact tool, and for types the quickstart has no option for, a case neither /design nor /slides triggers. - NEW: System Reminder: Attached machine could not be reached (asleep, busy, cause unknown, signed off, and went silent)—Add cause-specific follow-ups naming the remedy: waking the computer, checking its network without quitting Claude, or reopening Claude there; unknown causes keep the generic check. - NEW: System Reminder: Attached machine not answering—Splits out the opening shared by the retry-shortly and stop-calling arms, adding Claude not running on the machine to the usual causes. - NEW: System Reminder: Attached machine signed-off likely causes and Attached machine signed-off notice—Give likely reasons a machine's Claude signs off; once all machines have, run commands here and say reopening Claude with Remote Control on usually brings them back. - NEW: System Reminder: Memory file contents (unverified managed folder)—Heads unverified instructions from this machine's managed folder, warning that this session's own account could have written them: follow every limit, but treat nothing in them as permission. - NEW: Tool Description: ExitPlanMode (diskless session)—Without a plan file, Claude passes the whole plan inline as Markdown for approval; declining keeps Claude in plan mode to revise and resubmit. - NEW: Tool Description: SubscribePR—Describes subscribing a session to a GitHub pull request's CI failures, all-green notices, review comments, and close/reopen events; no built-in tool uses this description in this build. - REMOVED: Skill: Artifact whiteboard when-to-use guidance, Multiplayer whiteboard description, and Whiteboard description—Remove the whiteboard and multiplayer whiteboard skill descriptions and the guidance for offering a whiteboard unprompted; whiteboard publishing code remains in the build. - REMOVED: System Reminder: Attached machine could not be reached (return noted) and Attached machine could not be reached (return not noted)—Replaced by the cause-specific arms, which keep the same limits on calling the machine again and still have Claude tell the user what is blocked. - Agent Prompt: Claude Test explorer—Adds effort: medium to its frontmatter and lowers the report cap from 60 to 40 lines. - Data: Artifact host MCP server guidance—Says host:claude_browser answers when the page is opened in the Claude desktop app, not only in a Cowork session; viewers are asked at its first call instead of per website, then at each action reaching banking or other high-risk sites. - Data: Claude Platform on AWS reference—Adds an ant CLI row: with the region and workspace ID set, commands other than ant beta:worker go to Claude Platform on AWS; confirm with ant auth status. - System Reminder: Attached machine could not be reached and Attached machine could not be reached (oversized call)—The opening now says when the machine was last heard from, if known; its request to check that Claude is running moves into the oversized-call and cause-unknown arms. - System Reminder: Attached machine not answering (retry shortly) and Attached machine not answering (stop calling)—Now follow the shared not-answering opening; retry-shortly has the user check that Claude is open there, and stop-calling says opening Claude with the computer awake and online brings it back. - System Reminder: Plan approved—Mentions where the plan was saved, and referring back to it, only when the session has a plan file. - System Reminder: Unreachable attached machines—Has Claude relay what each machine's line says about why it is unreachable and what brings it back, instead of a fixed asleep, offline, or not-running explanation. - Tool Description: Artifact publishing and update guidance—Adds that a page's own URL changes never show in the address bar, so name bookmarkable views with bare, non-reserved hash tokens and give users each view's link. - Tool Description: Background monitor (streaming events)—In diskless sessions, adds a note that monitors are dependable only while Claude is working, and scopes per-occurrence monitors to the current turn. - Tool Description: Bash (sleep—no polling background tasks) and Tool Parameter: Bash run_in_background guidance—When background shells are kept elsewhere, say Claude is notified of a background command only if it finishes while Claude is still working. - Tool Description: Plugin organization submission and replacement—No longer says claude.ai scans every upload; whether it scans depends on the organization, so Claude reports a scan only as the result states. Details: github.com/Piebald-AI/claude…

What's new in Claude Code 2.1.302 system prompts (+1,759 tokens): Adds a feature-gated Easel tool for charting data already in the conversation, restructures Artifact resubmission refusals with a recovery path for publishes that carry edits, makes fork guidance conditional on whether forks run in the background, routes data dashboards to the Dashboard type, and notes standalone Claude Design's December 14, 2026 closure. - NEW: Tool Description: Easel chart—Draws charts from small tables already in context without guessing ambiguous columns; feature-gated, off by default, Linux/macOS only, and unavailable to HIPAA/ZDR organizations. - NEW: Tool Description: Artifact publish-edits resubmission merge guidance and Artifact publish-edits resend clause—Require merging newer content into the file with Edit or Write before resending, omitting edit pairs that would add those lines again. - NEW: Tool Description: Artifact resubmission refetch confirmation clause—Preserves the shared requirement to freshly fetch the artifact and fully read any required saved source before confirming newer changes are already included. - REMOVED: Tool Description: Artifact identical resubmission refusal—Consolidates live-version recovery into Artifact known-version resubmission guidance, parameterized for live or known newer source; identical-content refusal behavior remains. - Agent Prompt: Claude Test author and Claude Test explorer—Expand explicit tool denials to additional Claude Test browser helpers, including installation, saving, status, progress, logging, finishing, and reading. - System Prompt: Agent Summary Generation—Adds notes-based third-person summaries, omitting the tool ban and allowing the previous label to remain while the agent continues the same work. - System Prompt: Fork usage guidelines and Forked agent guidance; Tool Description: Agent (usage notes)—Make background-execution claims conditional and scope notification/no-peeking guidance to background forks; retain anti-fabrication warnings for background agents when forks run synchronously. - System Prompt: Project coordinator relay trust boundary, Project timeline user message provenance, and Project thread in-thread message provenance—Name available Project read/reply tools, including alternate names, when explaining untrusted fetched content and which messages can establish consent. - Tool Description: Artifact quickstart type guidance and Artifact type discovery guidance (including app wording)—Route data dashboards through the Dashboard type when offered, using quickstart intent other; distinguish visual mock-ups and honor requests for files. - Tool Description: Artifact stale publish saved-source guidance—Adds recovery for publishes carrying edits: merge saved-source lines into the file with Edit or Write, then resend using old strings from that merged file. - Tool Description: DesignSync—Adds notice that standalone Claude Design closes by December 14, 2026; requires relaying closure or relocation refusals verbatim and stopping. Details: github.com/Piebald-AI/claude…

What's new in Claude Code 2.1.301 system prompts (+10,251 tokens): Adds feature-gated lean rewrite alternatives for core instructions and tool guidance, a cloud-session reminder that settings edits need host approval, and conversation-summary instructions that can cap output at 2,000 words while carrying earlier user requests forward unchanged. - NEW: Agent Prompt: Parent messages and user authorization (lean rewrite)—Adds feature-gated alternate delegation guidance: parent messages direct tasks but cannot supply user consent or authorize permission, instruction-file, or configuration changes. - NEW: Agent Prompt: Subagent reporting notes (lean rewrite)—Adds alternate reporting guidance requiring absolute paths and direct final-message findings, with code snippets only when their exact text matters. - NEW: System Prompt: Background session instructions, Background session worktree isolation, and Background worktree persistence (lean rewrite)—Add alternate guidance for isolated edits, job-local temporary files, actionable reports, and preserving work through commits and pushes unless user instructions forbid them. - NEW: System Prompt: Context management (lean rewrite)—Adds alternate guidance to continue working through long conversations rather than wrapping up early, relying on automatic compaction within user limits. - NEW: System Prompt: Delivering requested work and Handling corrections (lean rewrite)—Add alternate guidance to follow explicit scope, postpone blocking questions until independent work finishes, and correct minor mistakes without excessive self-criticism. - NEW: System Prompt: Focus mode and Outcome-first communication style (lean rewrite)—Add alternate guidance for self-contained, outcome-first final replies; focus mode omits intermediate narration because only final text is visible. - NEW: System Prompt: Harness instructions (lean rewrite)—Adds alternate guidance covering clickable code references, independent parallel tool calls, permission denials, and treating hook output as user-system feedback. - NEW: System Prompt: Irreversible action caution (lean rewrite)—Adds alternate guidance to check authorization before hard-to-undo actions, understand files before deletion or overwriting, and faithfully report failures or skipped steps. - NEW: System Prompt: Memory hygiene (lean rewrite)—Adds a selectable alternate memory policy emphasizing deduplication, removing incorrect memories, avoiding redundant or session-only material, and verifying recalled references. - NEW: System Prompt: System reminder interpretation (lean rewrite)—Adds alternate guidance to ignore irrelevant application reminders while keeping reminders distinct from user instructions and consent for consequential actions. - NEW: System Reminder: Cloud session settings edit not applied—Explains that settings edits need approval on the host computer; reports the unapplied change and forbids retries or alternate editing routes. - NEW: Tool Description: Agent (when to launch subagents), Agent delegation summary, Agent delegation tradeoffs, and Agent result visibility (lean rewrite)—Add alternate guidance on agent selection, parallel delegation, supplying context, avoiding duplicate searches, and relaying completed results without inventing unfinished findings. - NEW: Tool Description: AskUserQuestion (lean rewrite)—Adds alternate guidance to ask only for genuinely blocking user decisions, place recommended choices first, and support custom or multiple answers. - NEW: Tool Description: Grep compact (lean rewrite)—Adds alternate ripgrep guidance favoring the integrated search tool for permission handling and file links, with regex, filtering, and multiline options. - NEW: Tool Description: Invoke skill (lean rewrite)—Adds alternate skill-loading guidance covering explicit requests, directory-specific selection, already-loaded instructions, and waiting for background skill results without invoking them again. - NEW: Tool Description: ToolSearch (lean rewrite)—Adds an alternate description explaining that tool search loads deferred tool schemas and that the system announces available deferred tools by name. - NEW: Tool Description: Workflow (lean rewrite)—Adds alternate workflow guidance retaining explicit user opt-in for large-scale orchestration, background completion notifications, inline scripts, and required literal workflow metadata. - Agent Prompt: Conversation summarization with additional instructions—Adds optional 2,000-word limits and unchanged carryover of earlier user requests; makes analysis conditional while preserving security constraints without it. - Agent Prompt: Workflow script plain text return note, Workflow script structured return note, Workflow subagent plain text output, and Workflow subagent structured output—Clarify that script-computed task text lacking user authority still assigns a workflow step; perform only that step rather than skipping it. - Data: Cloud session unverified remote branch push notice—Makes push-and-restart advice conditional, while still explaining when the remote branch or detached checkout could not be verified for session startup. - System Reminder: Plan file reference—Handles plans without a file path by announcing that a plan exists, rather than always displaying a plan-file location. - Tool Description: Artifact HTML document skeleton—Adds a fixed pseudo-element to paint the top safe-area band above sticky headers, using the header or page background as appropriate. - Tool Description: claude.ai Project—In main-conversation upload mode, limits local paths to unchanged, fully read text files up to 1 MiB in upload, output, or non-home working directories. - Tool Description: Claude in Chrome switched-off and unlinked chat guidance and Computer use unlinked chat guidance—Recognize tools appearing later despite earlier unavailable results; add desktop folder attachment as a linking remedy and avoid repeating exhausted setup steps. - Tool Description: Plugin organization submission and replacement—Directs users to make an outdated plugin version current and publish it on claude.ai, rather than asking the agent to retry publication. Details: github.com/Piebald-AI/claude…

What's new in Claude Code 2.1.300 system prompts (+10,545 tokens): Sharpens pasted-content trust boundaries and sharing consent, reinforces edit-before-republish rules for Artifacts, and clarifies file handling, unavailable tools, and retry limits, alongside new tracing examples and guidance on managing the shared reasoning-and-reply token budget. - NEW: Data: Managed settings x-whenUnreadable field—Defines fallback and unusable states for invalid managed values; an unusable MCP server list keeps plugin-provided server tools blocked. - NEW: Data: OpenTelemetry tracing reference (C#, Go, Java, PHP, Python, Ruby, TypeScript)—Adds SDK tracing examples for registering providers, configuring content recording, and propagating context to connect application, SDK, and API spans. - NEW: System Prompt: Pasted content instruction trust boundary—Treats paste-only messages as requests without confirmation, checks pasted commands before execution, and keeps embedded document instructions separate from user intent. - NEW: System Prompt: Published artifact inline edits and large-change republish and Published artifact same-response edit and republish—Prefer inline publish edits when supported; otherwise edit before republishing in the same response, preserving the published artifact when edits fail. - NEW: System Prompt: Session scratchpad temporary-file guidance—Requires session scratchpad use instead of system temporary directories unless explicitly requested, and complete absolute paths when naming scratchpad files. - NEW: System Reminder: App tool server retrying connection failure—Continues work while app tools reconnect automatically; avoids user-side repair requests and describes blocking failures only as something wrong on our end. - NEW: System Reminder: Combined thinking and reply token budget—Warns that reasoning and reply share one token limit; directs effort toward planning long deliverables rather than drafting the output twice. - NEW: System Reminder: Staged files failed to arrive—Warns that failed transfers can leave missing or older files; identifies disk-space and filename-length remedies, requiring consent before deleting user files. - NEW: Tool Description: Artifact multi-file edit and republish guidance and Artifact small-change edit and republish guidance—Require edits before republishing the same path in one response, avoid unnecessary reads, and preserve publication when an edit fails. - NEW: Tool Description: Claude in Chrome switched-off chat guidance and Claude in Chrome switched-off and unlinked chat guidance—Explain disabled access and optional computer linking, restrict troubleshooting to supported setup steps, and continue with available tools instead of guessing causes. - NEW: Tool Description: Claude in Chrome unavailable stub result—Explains that the stub enables nothing; reports unavailable Chrome access without guessing causes or suggesting setup, and uses available alternatives. - NEW: Tool Description: Plugin organization submission and replacement—Describes consented uploads and organization review, requires fresh consent for replacement, and waits for user action before retrying an outdated version. - NEW: Tool Description: Project local_file read-only workflow—Requires copying read-only project files into a new working file before editing; saving an uploaded document creates text without replacing the upload. - REMOVED: Tool Description: Existing Claude Docs artifact content—Drops the standalone instructions to edit the document bound to an artifact and create a replacement only when that document is absent. - Agent Prompt: Agent Hook—Removes the instruction to ignore transcript_path in hook input when no transcript file exists; still states that no transcript is available. - Agent Prompt: Web fetch agent usage guidance and Web reading specialist—Limit retries to timeouts, dropped connections, or Retry-After responses; warn against resending permanently failed URLs and sites refusing automated fetches. - Data: Claude Code agent proxy troubleshooting guide—No longer attributes proxy denials specifically to organization policy; describes disallowed hosts as the usual cause and removes named escalation destinations. - Skill: Mod GitHub sharing consent—Requires scanning files for keys before reading them, leaves flagged files unopened, and checks remaining files for secrets, internal URLs, or others' data. - Tool Description: Artifact HTML document skeleton—Requires side-viewport metadata and both gutter paddings together; leaves older pages lacking both unchanged unless the request reworks their layout. - Tool Description: Computer use enable stub guidance—Requires loading all deferred computer tools in one tool-search call and waiting for its result before using them; no separate enable step exists. - Tool Description: Updating existing artifacts (app wording)—Adds same-response edit-and-republish guidance selected by inline-edit support, requiring edits before publication and preserving the artifact when an edit fails. Details: github.com/Piebald-AI/claude…

What's new in Claude Code 2.1.299 system prompts (+6,453 tokens): Tightens consent boundaries for sharing and delegated work, spells out Windows sandbox recovery limits, and refines shared-document editing and conflict handling. Runner and SDK guidance also gets more precise about interrupted work, replayed messages, and scheduling, without treating metadata as permission or approval as a mode change. - NEW: Skill: Mod GitHub sharing consent—Offers sharing only after a working mod is demonstrated; requires separate sharing and visibility decisions and a pre-commit sensitive-data review. - NEW: System Prompt: Project coordinator relay trust boundary and Project coordinator reply consent—Treat coordinator relays as external content; replies approve only one explicitly proposed action and target, never menu options or vague retries. - NEW: System Prompt: Subagent carried user consent history—Credits parent-session user messages without letting delegation invent consent; bare replies apply only to preceding assistant messages, and trimmed history stays partial. - NEW: System Prompt: Windows sandbox certificate recovery—Requires a user-side HTTPS connection before retrying certificate failures, forbids weakening certificate checks, and stops after one unsuccessful retry. - NEW: System Prompt: Windows sandbox localhost retry guidance and Windows sandbox localhost user alternatives—Explain localhost failures and proxy exceptions; require port-scoped allowances and disclose lost sandbox protection, with user-directed alternatives where retries are unavailable. - NEW: System Prompt: Windows sandbox package cache limits—Documents read-only Maven cache access and warns that .NET builds using --no-restore after an outside restore can delete package DLLs. - NEW: System Prompt: Windows sandbox placeholder file protection—Protects sandbox-created empty configuration files from staging, moving, deletion, and ignore rules; tells the user when they block Git operations. - NEW: System Prompt: Windows sandbox profile tool retry guidance and Windows sandbox profile tool user guidance—Diagnose profile-installed tool failures as sandbox restrictions; reserve narrow read grants for the user and warn that large grants can impede startup. - NEW: System Reminder: Artifact capabilities contract unavailable—Explains refused capability declarations and permits omitting them only when needed capabilities are already stored or the user requests it, disclosing missing grants. - NEW: System Reminder: Artifact published source edit guidance—Directs exact small replacements through publish-call edits, updating local source and publication together; larger rewrites still use file edits. - Tool Description: Updating existing artifacts (app wording)—Adds conditional guidance to send small changes as exact replacements on the publish call instead of editing the local file first. - NEW: System Reminder: Project document version conflict recovery—Requires rereading and merging after a version conflict, allows two retries, and forbids dropping the version guard or deleting the document. - Tool Description: claude.ai Project—Requires replacing shared documents with the version read; conflicting writes change nothing and must be merged into the latest version. - NEW: System Reminder: Reserved-address MCP server refusal—Explains that reserved server addresses are refused by Claude Code, not connection failures or administrator policy, and that retrying cannot help. - NEW: System Reminder: Resume agents interrupted by container restart—Directs resuming interrupted agents from saved transcripts rather than duplicating them; relaunch only if resuming fails and the result is still needed. - NEW: System Reminder: User-controlled plan mode exit—In sessions without a plan file, presents the plan in the reply and requires a user mode change; approval alone leaves planning read-only. - NEW: Tool Description: Existing Claude Docs artifact content—Directs editing the already-bound Claude Docs document instead of creating another; create and bind a replacement only when that document is absent. - REMOVED: Tool Description: Artifact stale publish line diff guidance—Drops the standalone live-page line-diff merge instructions; saved-source recovery guidance remains, so this does not remove stale-publish protection. - Data: Interrupt receipt cancelled field and Interrupt receipt still queued field—Clarify cancellation of interrupted-turn reruns and associated client messages: associated UUIDs appear once under cancelled, though never queued or listed as survivors. - Data: SDK safeguard_results field—Accepts only blocking verdicts for repeated message IDs; clients must send clearing verdicts only for the first call of a response. - Data: Self-hosted runner orchestrator command help—Adds optional SIGTERM draining for in-flight polling and hooks, with immediate-stop escalation and process-manager timing guidance; the default remains immediate stop. - Data: Tool use display metadata field—Adds connector identity, dynamic-server identity claims, and unprefixed tool names for display; dynamic claims are not verified identities or permission grants. - Data: Turn handoff available event schema—Reports shared_message_id, letting clients discover whether the worker accepts assistant messages in the same turn sharing an ID. - Data: Turn handoff hydrates_carried_lines field—Requires storage confirmation for every message before echoing replayed turns; explains refusal cleanup and the limit of retained refusal history. - System Prompt: Coordinator mode orchestration—Adds an agent-result status for waiting on its own background work, warning that its result may be interim and another notification may follow. - System Prompt: Self-hosted runner doctor—Adds safe noninteractive Git credential diagnosis and fixes, warns that the git proxy erases global configuration, and clarifies startup-timeout wording. - System Prompt: Self-hosted runner setup—Warns that repository cloning needs noninteractive HTTPS credentials and explains clearing a failed session before restarting the local runner, without altering Git settings. - System Reminder: Attached machine untrusted attachments refusal—Adds fallback guidance when neither folder-trust tool is available: explain the block and suggest a fresh session without attached repositories or files. - Tool Description: Artifact HTML document skeleton and Artifact responsive page contract—Require side safe-area gutters and a matching viewport-fit metadata tag; extend inset handling to fixed, floating, and absolutely positioned controls. - Tool Description: claude.ai Project—Adds conditional original-file downloads for document uploads whose extracted text misses layout or figures, and uses the session's read-tool name. - Tool Description: CronCreate—Describes fixed per-job schedule offsets and requires reporting actual next-run timestamps returned by scheduling tools, rather than the cron expression's nominal minute. - Tool Description: WebFetch (concise) and WebFetch private URL warning—Add conditional guidance to fetch URLs already seen in conversation; use search or linking pages instead of inventing paths or recalling URLs. Details: github.com/Piebald-AI/claude…

What's new in Claude Code 2.1.296 system prompts (+560 tokens): Adds Managed Agents quickstart templates and subagent delegation guidance, drops terminal-wording Artifact publishing variants, Bash cost advice, and the PHP and Managed Agents outcomes references, and revises self-hosted runner, gateway, and summarization instructions. - NEW: Data: Managed Agents quickstart (Bug hunter, Explainer video maker, Watchlist scanner)—Three bundled templates, each with agent.md, environment and session definitions, built on parallel workflow runs with an independent review step. - NEW: System Prompt: Subagent delegation when-to-use guidance—Delegate when an agent type matches, work is independent and parallel, or answers span several files; search directly for single-fact lookups. - NEW: Data: Self-hosted runner disabled-by-organization fatal message—Explains that registration was refused because self-hosted environments are off for the organization, and that the accepted environment secret should not be rotated. - NEW: Data: SDK set model only_if_model_is_current field—Internal flag making a set_model request a system-prompt-only update that succeeds only when the named model is already current. - NEW: Data: Turn handoff hydrates_carried_lines field—Internal field marking workers that rebuilt their conversation from the stored transcript, and describing how resent handoffs are re-accepted, echoed, or refused. - REMOVED: Tool Description: Artifact publishing introduction, Artifact runtime capabilities guidance (including optional marker), Artifact browser storage guidance (v2 page contract and capabilities-skill variants), Updating existing artifacts, Prohibited artifact publishing, Publish audience-facing deliverables, and Artifact gallery and publish response guidance—Terminal-wording variants dropped; app-wording counterparts remain, but the gallery and publish-response note has no replacement. - REMOVED: System Prompt: Bash call cost guidance—Drops the advice to batch a decision's work into one command, keep printed output small, and recheck every requirement before finishing. - REMOVED: Data: Managed Agents outcomes—Drops the standalone outcomes reference covering user.define_outcome, rubrics, grader iterations, and deliverables. - REMOVED: Data: Tool use reference (PHP)—Drops the PHP tool-use reference, including the beta tool runner and manual agentic loop. - Data: Tool use reference (C#, Ruby)—Adds a "When to offer Managed Agents" section: after finishing large repeated-step Messages API jobs, offer Managed Agents once, with trade-offs. - Data: Claude Code gateway protocol—Documents a Desktop Code tab probe sending If-None-Match: * for session settings, and how 200/304/404, errors, and timeouts affect session startup. - System Prompt: Self-hosted runner doctor—Adds troubleshooting rows for the organization-disabled refusal, spent or expired work orders, rejected work orders, and poll-auth failures on on-demand runners. - Tool Description: SendMessage—When messaging beyond a session's own agents is disabled, recipients are limited to session agents, and the inbox and legacy protocol-response text are dropped. - Tool Description: Agent usage notes—Adds a conditional note: give each parallel fan-out lookup effort: "lower", use "lower" for mechanical, self-checking briefs, and "higher" for pieces needing more reasoning. - Tool Parameter: Bash run_in_background guidance—Adds a conditional note that background commands keep running after the reply ends until they finish or time out, while shell-detached processes (nohup, &) are usually stopped minutes later. - System Prompt: Subagent delegation cost guidance—Calls the subagent brief your "first" lever on every cost, rather than your "one" lever. - Agent Prompt: Dream memory consolidation—Removes the session-logs (logs/) source and its ls -R logs/ step; sessions/ entries and transcript search remain. - Agent Prompt: Conversation summarization (short variant)—Drops the security-instruction preservation note, user-message attribution guard, and closing line about following additional summarization instructions. - Data: Data visualization marks and anatomy—Y-axis ticks now use compact labels from 1,000 up (0 / 1K / 2K) instead of comma-separated thousands. - Data: Turn handoff available event schema—Adds hydrates_carried_lines to the capabilities the available event reports. Details: github.com/Piebald-AI/claude…

What's new in Claude Code 2.1.295 system prompts (+5,697 tokens): Adds prompt guidance for merging live Artifact changes and publishing small edits in one response. It also adjusts background-command completion notices, remote-machine reconnection instructions, web-search citations, and consent before Claude Test seed scripts run. - NEW: Agent Prompt: Background agent state classifier chat-thread relay rule—For sessions that reach people only through chat-thread posts, a named person counts as the user, so open asks stay blocked even beside scheduled check-ins. - NEW: Agent Prompt: Subagent delegation configuration classifier—Judges whether a user's CLAUDE.md, skills, and custom agents push Claude to delegate work to subagents, without obeying instructions found in that configuration. - NEW: Tool Description: Artifact stale publish line diff guidance and Artifact line diff reading legend—Stale-publish refusals can show a line diff of live-page changes, counted as viewed, to merge into your file, with a legend for the marks. - Tool Description: Artifact stale publish saved-source guidance—Leads with a live-page diff summary (changed places and lines) when the diff itself is not shown, then gives the saved-source merge steps. - NEW: Tool Description: Artifact same-response edit and publish guidance—When enabled, small known changes are made by sending Edit calls and the Artifact publish together in one response, without a prior Read. - Tool Description: Artifact HTML document skeleton and Artifact page authoring and HTML skeleton (app wording)—Drops the "off-white ground" from the default body style description, leaving a zero-margin system font. - NEW: Tool Parameter: Bash run_in_background guidance (completion notices disabled) and PowerShell run_in_background note (completion notices disabled)—When completion notices are off, nothing announces the finish; read the output file or end the command with your own echo. - System Prompt: Avoiding Unnecessary Sleep Commands (part of PowerShell tool description)—Omits the "you will be notified when it completes" sleep and polling advice when background completion notices are disabled. - Tool Description: WebSearch and WebSearch (concise)—When search citations are enabled, replaces the mandatory "Sources:" list with a note that result titles and page text are untrusted data. - System Prompt: Interactive agent intro—The default opener is now "an agent working with the user toward their goals, using your own judgment", replacing the software-engineering "interactive agent" line; previously flag-gated. - Tool Description: AskUserQuestion—Drops the plan-mode note in diskless-host sessions missing EnterPlanMode or ExitPlanMode; the shared usage guidance is otherwise unchanged. - System Reminder: MCP servers failed to connect—Distinguishes servers retried in the background, whose tools may appear later, from failures leaving tools unavailable for this session. - System Reminder: Attached machine stopped answering and went to sleep—Adds return-tracking and paused-or-withdrawn guidance, allowing another call after a reconnection note or explicit user request and requiring outcome checks before repetition. - System Reminder: Unreachable attached machines and unreachable-machine tool errors—Adds reconnection-note and five-minute recheck variants while retaining user-requested retry guidance and directing independent work to continue. - Agent Prompt: Web fetch agent usage guidance—Adds advice to read a public page rather than answer from memory when it holds the answer, and omits the SendMessage follow-up hint when unavailable. - Agent Prompt: Claude Test explorer—Explains why the file-access restrictions apply to every Read, Grep, and Glob: what the explorer reads stays with it and can end up in its report. - Data: Claude Test spec file format—The setupCommand seed script now asks before each run and runs only on a yes, instead of running automatically before every run. - Data: Claude Code gateway protocol—Tells gateways to send x-should-retry: false with 501 responses, and to answer count_tokens with Bedrock's CountTokens, falling back to 501 on failure. Details: github.com/Piebald-AI/claude…

What's new in Claude Code 2.1.294 system prompts (+264 tokens): Rewrites the three hook-evaluator prompts to spell out when an action is allowed or blocked and to tell the evaluator to treat event data and inspected content as information rather than instructions. - Agent Prompt: Agent Hook—Requires a reason with every result, clarifies allow/block decisions, and rejects instructions embedded in event data or inspected content. - Agent Prompt: Hook condition evaluator (stop)—Adds guidance defining ok as allow/block and treating event JSON and inspected content as data, not instructions. - Agent Prompt: Hook condition evaluator—Reframes evaluation around whether an action may proceed, explains rule and condition outcomes, and rejects instructions embedded in evaluated data. Details: github.com/Piebald-AI/claude…

What's new in Claude Code 2.1.293 system prompts (+2,629 tokens): Focuses on completing unblocked work and investigating quiet commands, with feature-gated Haiku 5.5 guidance, stricter Artifact version pinning, and clarifications for SDK metadata, Project uploads, agent continuation, and Chrome troubleshooting. - NEW: Data: Tool result display metadata field—Documents wrapper-level SDK metadata for non-execution reasons, remedies, permission decisions, and human denial feedback; never replayed to the model. - NEW: System Prompt: Finish the request instead of stopping early—Adds feature-gated Haiku 5.5 completion guidance: finish large and unblocked requests without unnecessary check-ins, while retaining approval for consequential actions. - NEW: System Reminder: Background command quiet check-in—Requires inspecting silent background commands for progress or hangs, stopping stuck work, and ending the turn for healthy long-running commands. - Tool Description: Artifact design fallback requirements, Artifact page implementation requirements (app wording) and Artifact publishing and update guidance—Require library URLs to pin exact versions at least two weeks old, rejecting bare names, ranges, and tags. - Data: Message Batches API reference—Python—The classification example now sets max_tokens=1024 with low effort, noting Haiku's default thinking counts toward max_tokens. - Data: SDK result safety_stops field—Counts content-filter stops only when they end the call; retried stops remain uncounted, so zero does not rule them out. - Data: syncClaudeAiSkills setting—Skills re-sync about every 10 minutes while a session is in use, and a quarter as often otherwise. - Data: Background tasks changed event schema—Removes the claim that the payload carries only IDs while retaining the warning not to correlate level changes with edge events. - Tool Description: claude.ai Project—Adds a main-conversation upload mode for unchanged files read or written in full this session, using absolute paths without repeating their contents. - Tool Description: Agent (usage notes) and System Reminder: Async agent launched metadata—Condition continuation instructions on availability, explaining that unsupported sessions start fresh agents and omitting unusable SendMessage hints. - System Reminder: Remote Chrome browser extension not connected—For multi-organization users, requires matching the extension's organization; reserves logout for failed checks, with advance warning about erased shortcuts and scheduled tasks. - Agent Prompt: Claude Test explorer—Rewords the file-access restrictions to require compliance in every Read, Grep, and Glob instead of emphasizing the agent's sole responsibility. Details: github.com/Piebald-AI/claude…

What's new in Claude Code 2.1.292 system prompts (+7,031 tokens): Adds a batch of SDK reference entries (safety-stop counts, abandoned-stream markers, usage-limit errors, permission-check status), a consent-first PublishPlugin file listing, an Artifact preview in the real viewer, and smaller wording changes to queued notifications, computer use, web fetch, and Artifact guidance. - NEW: Data: SDK result safety_stops field—Reports a running per-session total of model calls stopped by a safety system, including refusals and content-filter stops; resets on resume or /clear. - NEW: Data: SDK MCP server disableAutoBackground field—When true, the CLI never auto-backgrounds calls to that server's tools, intended for servers whose calls wait on a person. - NEW: Data: SDK api_error_params field—Documents api_error parameters: refused effort level, removed-media kind and reason, credential-failure provider and remedy, and usage-limit rate_limit_info. - NEW: Data: SDK message_stop abandoned_blocks field—Marks a CLI-generated message_stop after a failed or stalled stream, so consumers can drop blocks from that index up that never got assistant messages. - NEW: Data: SDK task notification handback field—Completed-subagent notifications say how a SubagentHandback report reached its caller; omitted when the subagent failed, was stopped, or is still waiting. - NEW: Data: SDK permission_check_status event schema—Signals a tool call waiting unusually long on the auto-mode classifier, as a "checking"/"done" pair hosts can display; display-only and best-effort. - NEW: Data: Interrupt worker_epoch parameter—Ties a service-sent interrupt to one worker life, refusing stale or invalid values with errors; a person's Stop never carries it. - NEW: Data: SDK read_file anchor field—Resolves a read path against a fixed workspace, diff root, or repo checkout instead of the live cwd, refusing ~ and paths that escape it. - NEW: Data: SDK safeguard_results field and Data: SDK safeguard_tool_use field—Carry the Messages API verdict for one call so the worker can decide its permission; ignored if malformed or mismatched. - NEW: Data: SDK ui_prompt_autocomplete request schema—A remote composer asks plugins for autocomplete rows for the token at the caret; newer asks supersede older ones, and failed chains return no rows. - NEW: Tool Description: PublishPlugin file listing and consent flow—Lists the plugin's files, shows folder, organization, count and size, and asks the user first, in any permission mode; writes a manifest if none exists. - NEW: Tool Description: Artifact preview action (in-session viewer)—Opens the page in the real artifact viewer, returning a first-view picture, console errors, and a clickable outline; the viewer's own instructions win on when to check. - NEW: Data: Network-path settings file unread recovery guidance—When plugin uninstall couldn't read a settings file through a link to another machine, says to replace the link or use the real location, then retry. - NEW: System Reminder: Queued notifications read-time clock note—States when the read happened by this machine's clock and explains the queued-at, reached-session, and whole-wait headers, including clock skew. - NEW: Agent Prompt: Background agent state classifier ask-opens-message rule and Agent Prompt: Project thread status card classifier ask-opens-message rule—Classify a message that opens with the agent's need as blocked or needs-reply unless the tail resolves it; offers still don't gate. - NEW: Tool Description: Computer use unlinked chat guidance—Says computer tools cannot be linked from this chat, so the model tells the user briefly, uses other tools, and suggests opening the chat in the desktop app. - Agent Prompt: Web fetch agent usage guidance—Clarifies that without a built-in fetch tool this agent stands in for it, so calling it to read a page is ordinary tool use. - Tool Description: Agent explicit-spawn restriction—Adds an exception: where the web fetch agent type stands in for a missing fetch tool, calling it to read a page is ordinary tool use. - Data: Artifact capability verification pass—The pre-publish preview step is now offered only when the artifact check tool is the local Chrome viewer. - Data: Background tasks changed event schema—The event also fires when an entry's parent_task_id changes, and now spells out the ordering between level changes and task start, update, and notification events. - Data: SDK API error kind field—Adds the usage_limit_reached kind for an account plan limit or usage-credit spend cap, distinguished from policy, throttle, capacity, and API-key 429s. - Data: Turn handoff file names field—Adds that if a person later sends another file under a placed name, the copy is replaced or removed while still unchanged. - Data: Turn handoff memory context field—The memory line is now also used on history requests when the session holds none of its messages yet; wording "runs without" becomes "goes on without". - Skill: Artifact diagramming—Artifacts made from an Artifact type draw diagrams as that type's instructions and referenced pages say; these mechanics apply only where they are silent. - System Reminder: Queued notifications delivery—The formatted notification list now appears directly after the "listed oldest first" sentence. - Tool Description: Artifact action reference and Artifact action reference (concise app wording)—The list action now returns artifacts most recently opened or updated first, instead of newest first. - Tool Description: Computer use enable stub guidance—Adds that the Claude app on the computer may show why under Settings → This computer → Computer use. - Tool Description: GetTask—Drops "when your turn ends or" from the note that a task's status message says it is terminated at your final response. - Tool Description: SuggestConnectors—Accepts installed connectors' server_id values listed in the system prompt, as well as ids from a SearchMcpRegistry result. Details: github.com/Piebald-AI/claude…

What's new in Claude Code 2.1.292 system prompts (+7,031 tokens): Adds a batch of SDK reference entries (safety-stop counts, abandoned-stream markers, usage-limit errors, permission-check status), a consent-first PublishPlugin file listing, an Artifact preview in the real viewer, and smaller wording changes to queued notifications, computer use, web fetch, and Artifact guidance. - NEW: Data: SDK result safety_stops field—Reports a running per-session total of model calls stopped by a safety system, including refusals and content-filter stops; resets on resume or /clear. - NEW: Data: SDK MCP server disableAutoBackground field—When true, the CLI never auto-backgrounds calls to that server's tools, intended for servers whose calls wait on a person. - NEW: Data: SDK api_error_params field—Documents api_error parameters: refused effort level, removed-media kind and reason, credential-failure provider and remedy, and usage-limit rate_limit_info. - NEW: Data: SDK message_stop abandoned_blocks field—Marks a CLI-generated message_stop after a failed or stalled stream, so consumers can drop blocks from that index up that never got assistant messages. - NEW: Data: SDK task notification handback field—Completed-subagent notifications say how a SubagentHandback report reached its caller; omitted when the subagent failed, was stopped, or is still waiting. - NEW: Data: SDK permission_check_status event schema—Signals a tool call waiting unusually long on the auto-mode classifier, as a "checking"/"done" pair hosts can display; display-only and best-effort. - NEW: Data: Interrupt worker_epoch parameter—Ties a service-sent interrupt to one worker life, refusing stale or invalid values with errors; a person's Stop never carries it. - NEW: Data: SDK read_file anchor field—Resolves a read path against a fixed workspace, diff root, or repo checkout instead of the live cwd, refusing ~ and paths that escape it. - NEW: Data: SDK safeguard_results field and Data: SDK safeguard_tool_use field—Carry the Messages API verdict for one call so the worker can decide its permission; ignored if malformed or mismatched. - NEW: Data: SDK ui_prompt_autocomplete request schema—A remote composer asks plugins for autocomplete rows for the token at the caret; newer asks supersede older ones, and failed chains return no rows. - NEW: Tool Description: PublishPlugin file listing and consent flow—Lists the plugin's files, shows folder, organization, count and size, and asks the user first, in any permission mode; writes a manifest if none exists. - NEW: Tool Description: Artifact preview action (in-session viewer)—Opens the page in the real artifact viewer, returning a first-view picture, console errors, and a clickable outline; the viewer's own instructions win on when to check. - NEW: Data: Network-path settings file unread recovery guidance—When plugin uninstall couldn't read a settings file through a link to another machine, says to replace the link or use the real location, then retry. - NEW: System Reminder: Queued notifications read-time clock note—States when the read happened by this machine's clock and explains the queued-at, reached-session, and whole-wait headers, including clock skew. - NEW: Agent Prompt: Background agent state classifier ask-opens-message rule and Agent Prompt: Project thread status card classifier ask-opens-message rule—Classify a message that opens with the agent's need as blocked or needs-reply unless the tail resolves it; offers still don't gate. - NEW: Tool Description: Computer use unlinked chat guidance—Says computer tools cannot be linked from this chat, so the model tells the user briefly, uses other tools, and suggests opening the chat in the desktop app. - Agent Prompt: Web fetch agent usage guidance—Clarifies that without a built-in fetch tool this agent stands in for it, so calling it to read a page is ordinary tool use. - Tool Description: Agent explicit-spawn restriction—Adds an exception: where the web fetch agent type stands in for a missing fetch tool, calling it to read a page is ordinary tool use. - Data: Artifact capability verification pass—The pre-publish preview step is now offered only when the artifact check tool is the local Chrome viewer. - Data: Background tasks changed event schema—The event also fires when an entry's parent_task_id changes, and now spells out the ordering between level changes and task start, update, and notification events. - Data: SDK API error kind field—Adds the usage_limit_reached kind for an account plan limit or usage-credit spend cap, distinguished from policy, throttle, capacity, and API-key 429s. - Data: Turn handoff file names field—Adds that if a person later sends another file under a placed name, the copy is replaced or removed while still unchanged. - Data: Turn handoff memory context field—The memory line is now also used on history requests when the session holds none of its messages yet; wording "runs without" becomes "goes on without". - Skill: Artifact diagramming—Artifacts made from an Artifact type draw diagrams as that type's instructions and referenced pages say; these mechanics apply only where they are silent. - System Reminder: Queued notifications delivery—The formatted notification list now appears directly after the "listed oldest first" sentence. - Tool Description: Artifact action reference and Artifact action reference (concise app wording)—The list action now returns artifacts most recently opened or updated first, instead of newest first. - Tool Description: Computer use enable stub guidance—Adds that the Claude app on the computer may show why under Settings → This computer → Computer use. - Tool Description: GetTask—Drops "when your turn ends or" from the note that a task's status message says it is terminated at your final response. - Tool Description: SuggestConnectors—Accepts installed connectors' server_id values listed in the system prompt, as well as ids from a SearchMcpRegistry result. Details: github.com/Piebald-AI/claude…

What's new in Claude Code 2.1.290 system prompts (+12,809 tokens): Adds gated proactivity levels, alternative conversation-summary prompts, nine Managed Agents quickstart templates, and an experiment-gated Claude in Chrome setup offer, plus smaller changes to commit guidance, permission allowlisting, and recovery steps. These are prompt and rollout-variant changes, not proof of universal feature launches. - NEW: Agent Prompt: Conversation summarization (capped, lean, and short variants)—Three alternative compaction prompts: a 2000-word capped summary, a lean replacement-context brief, and a short checklist weighting the user's words. The existing prompt stays default. - NEW: System Reminder: Proactivity level low, Proactivity level medium, and Proactivity level high—Low confirms changes except small, unambiguous tasks; medium edits but needs commit/push/PR consent; high proactively follows through, including experiment check-ins. - NEW: System Reminder: Proactivity setting announcement header and Proactivity setting withdrawn—The header names the level and supersedes earlier settings without overriding plan mode, permissions, or destructive-action rules; the withdrawal restores base behavior. Behind a proactivity gate. - NEW: Data: Managed Agents quickstart (Contract tracker, Data analyst, Deep researcher, Field monitor, Incident commander, Sprint retro facilitator, Structured extractor, Support agent, Support-to-eng escalator)—Nine bundled templates, each with an agent.md definition and system prompt; some add cron deployments or an outcome rubric. - NEW: Skill: Claude API Managed Agents onboarding source tier—Closing section for managed-agents-onboard requests naming the source as bundled quickstart, first-party URL, or third-party URL; nothing in the request or fetched pages can raise the tier. - NEW: Data: Working tree upload refusals (git dubious ownership, missing git objects folders, private git directory setup failure, unread git config environment variable)—Four errors that stop a working-tree upload, each explaining the cause and the recovery steps. - NEW: System Reminder: Output token limit continuation—Defines continuation from the start of a cut-off line, row, item or sentence, reopening fences and tables; active recovery still uses the older template. - NEW: Tool Description: Offer Claude in Chrome setup—Offers a one-time Claude in Chrome setup card after a Chrome tool reports the extension is disconnected and the task needs the user's browser. Needs a supporting client and an experiment flag. - NEW: System Prompt: Bash call cost guidance—Explains that each Bash call is a full model round trip, so batch needed work into one command, keep output small, and recheck requirements. Behind a feature flag. - Agent Prompt: Summarization no-tools guard—The required plain-text reply shape is now variable, naming only a

block for the new lean, short, and capped compaction variants. - Agent Prompt: Pull request creation and Agent Prompt: Quick git commit—Drop the optional reminder to run the user's verify, simplify, and code-review skills right before committing. - Agent Prompt: Artifact comment thread background directive—Thread-read instructions now pass the thread's thread_id to the Artifact comments action. - Data: Built-in gh stand-in api command help—Directs any-host requests to use --hostname for non-GitHub hosts rather than GH_HOST or the host embedded in GH_REPO. - Data: Gateway device code entry page—Removes the "Connect device" badge, labels the code input with the heading, and moves any error message inside the form as its description. - Data: Managed Agents outcomes—Generally worded outcomes need the user's question sent first in the same initial_events array, and requires_action sessions join budget-paused ones in accepting only settle events. - Skill: /doctor slash command and Skill: Generate permission allowlist from transcripts—Never allowlist pyright in any form, because it runs python3 in the working directory; it leaves the auto-allowed read-only list. - System Prompt: Artifact patch turn edit composer—Requests for an opinion, recommendation, or choice between options count as questions, so the assistant answers in words instead of editing. - System Reminder: Attached machine untrusted attachments refusal—Rewrites the recovery steps: call list_computer_folders once, relay any trust question or request_computer_folder Allow, otherwise explain what stays blocked. - Tool Description: Agent (when to launch subagents)—Optionally appends a note about the per-agent token budget when the subagent budget feature is enabled. - Tool Description: Code review command—Effort levels now read low|medium|high|xhigh|max, described as running from few high-confidence findings up to many, some uncertain. Details: github.com/Piebald-AI/claude…

What's new in Claude Code 2.1.288 system prompts (+4,051 tokens): Adds an artifact-edit composer and notices for held cross-session messages and worktree-blocked subagent writes, describes a configurable code-review finding limit, and updates gateway, SDK, PDF-reading, GitHub Enterprise, and Artifact device-access guidance. - NEW: System Prompt: Artifact patch turn edit composer—Tool-less composer that answers an artifact edit request with one JSON decision: ordered exact-string edits with a reply and confidence, or a hand-back to the main assistant. - NEW: System Reminder: Cross-session message held notice—Says a message to another session is held, not delivered, usually from mismatched permission modes; don't report delivery, wait, or resend. - NEW: System Reminder: Subagent write blocked pending worktree isolation—Tells a subagent its write to the shared checkout is blocked until the parent background session isolates into a worktree, and how to proceed or disable the guard. - NEW: Tool Description: Code review command—Describes the code review command's effort levels, --comment and --fix modes, and the --max-findings / all / default limit. - NEW: Tool Description: Enable Claude app browser (shared description), Enable Claude in Chrome (shared description), and Enable computer use (shared description)—Rollout-flag variants that also cover steps inside skills or user-requested tasks, and say calling them adds the file tools too. - Agent Prompt: Code review (minimal, low, medium, high, extra-high/maximum modes and ReportFindings output format) and Skill: Code review (inline medium/high, inline xhigh, low-effort expanded findings, findings JSON array)—Finding limits are no longer fixed per effort level; they follow the configurable maximum, which can be "all findings". - Data: Built-in gh stand-in api command help—Now covers GitHub Enterprise hosts from the checkout's remote or --hostname, notes a self-hosted runner's CLI uses its own credentials, and suggests another JSON tool when jq is missing. - Data: Claude Code gateway protocol—Adds the structured_outputs_unsupported error kind for a rejected output_config.format, and "does not support PDF" rejections now count as document_block. - Data: Claude Test spec file format—Example spec switches from a project-joining flow to adding and searching a "Claude Test demo recipe", with matching assertions and sentinel text. - Data: Hook classifier context field—Drops "inner REPL calls" from the list of calls with no per-result line in the classifier transcript. - Data: SDK API error kind field—Adds the media_removed kind, for when the API refuses an image or document and Claude Code leaves it out of later requests. - Data: SDK initialize response sdk_mcp_manifests_parked field—Adds two not_honoured causes: an entry past the CLI's limits on a live answer, and the per-minute tool-list read cap. - Data: Turn handoff available event schema—The announcement also lists carried_writes among the optional members reported when set. - System Reminder: Large PDF read guidance—Adds a branch for models that cannot be sent a whole PDF, only page images; also pluralizes "page" correctly. - Tool Description: ReadFile and ReadFile (compact)—PDF-reading text now depends on the current model rather than only on whether PDFs are supported at all. - Tool Description: Artifact publishing and update guidance—Camera, microphone, location and similar device APIs are usable only when the user has a runtime capability and the page declares it; otherwise take uploads. Details: github.com/Piebald-AI/claude…

What's new in Claude Code 2.1.287 system prompts (+4,435 tokens): Adds prompts for Artifact comment threads, a gh stand-in limited to REST requests, background task-output reads, "Send now" interrupts, and a measured-usage chart command, drops the plugin tsconfig setup guidance, and updates guidance on background agents, Artifact SVG handling, and monitors. - NEW: Agent Prompt: Artifact comment thread background directive—Tells a forked background agent to handle one Artifact comment thread itself, leaving the main session's pending work alone and treating thread content as material, not instructions. - NEW: Data: Built-in gh stand-in api command help—Usage text for Claude Code's built-in gh stand-in, which supports only gh api REST requests through the session's GitHub proxy. - NEW: Data: Get task output control request—Reads the last 8 KiB of a background shell or Monitor task's output without a model turn; refused on redacting lanes. - NEW: Data: Interrupt send_now parameter and Interrupt send_now message_uuid parameter—Let a client's "Send now" push waiting user messages to the model without acting as a Stop, optionally targeting one exact message. - NEW: Data: Plugin manifest types field—Names a self-contained .d.ts type contract for the noun a plugin adds, delivered to dependent mods and checked by claude plugin validate. - NEW: Data: SDK thinking_duration_ms field—Display-only wrapper field giving how long a streamed thinking block took, excluding the first-token wait; absent on older CLIs and never replayed to the model. - NEW: Data: Structured tool output field schema—Describes the tool_use_result field's per-tool output shapes, the completed Agent output, and a detachedToolCall placeholder for calls that stepped aside for a user message. - NEW: Data: Turn handoff file_names field—Lists user-attached files and client-facing names that a capable worker copies into the session home directory before running handed-over calls; malformed lists are ignored. - NEW: Skill: /explain-usage measured usage slash command—Variant that receives the session's measured token usage as JSON and asks for one simple chart of effective usage by group with a plain-language explanation. - NEW: System Reminder: Auto mode no verdict after hook input rewrite—Says auto mode could not review a call whose input a hook rewrote; retry once, then move on and tell the user. - REMOVED: Skill: /plugin-types tsconfig setup—Drops the closing guidance for pointing a plugin's tsconfig.json or jsconfig.json at the generated declarations. - Agent Prompt: Determine which memory files to attach—Ends with an explicit instruction to reply with only a JSON object of the form {"selected_memories": [...]}. - Agent Prompt: Web reading specialist—Fetching now also stops when a permission request goes unanswered, and the agent tells the caller why further fetches would fail too. - Data: Turn handoff available event schema—The announcement now also lists home_files and file_names among the optional members reported when set. - System Prompt: Coordinator mode orchestration—Adds a note that when the tool list holds two copies of the PR-activity subscription tools (one from the github connector), call the preferred one. - System Reminder: Attached machine reply not received and Attached machine stopped answering—Fallback now says to continue the task without the named remote machine instead of doing what "this container can do". - System Reminder: Still-running tool call—The sentence saying the call still shows as in progress on the user's screen is now included only when the call is on screen. - Tool Description: Agent (usage notes)—Adds an opt-in background-agents note: agents run in the background only with run_in_background: true, may still run foreground or be refused, and results must never be fabricated. - Tool Description: Artifact assets guidance (app wording) and Tool Parameter: Artifact supporting files with cross-artifact sources (and app wording)—SVG images may now be copied or used as sources; only HTML and XML documents are excluded. - Tool Description: Artifact type staged first-publish sequence—Restores the instruction to write files directly with the file-writing tool, with no shell step and never a generating script. - Tool Description: Background monitor (streaming events)—In diskless sessions, says stderr triggers no notifications and only its last lines appear in the notice sent when the script ends. - Tool Description: Poll—The untrusted-event-content guidance can now be overridden, and optional trailing guidance can follow the closing "nothing is dropped" line. Details: github.com/Piebald-AI/claude…

What's new in Claude Code 2.1.286 system prompts (+3,170 tokens): Adds prompts for Artifact sharing and the "You should know" side agent, tightens how relayed approvals count in shared Projects, folds security-monitor rule variants into the main monitor prompt, and splits the Bash git commit and PR guidance into two prompts. - NEW: Tool Description: Artifact share action reference—Documents the Artifact share action for sharing an owned artifact with the organization or named people, with a confirmation card the person must approve first. - NEW: Tool Description: Artifact sharing guidance—Lets Claude offer once to share a private artifact when readers are named, proposing the narrowest audience and never inventing email addresses. - NEW: System Prompt: Project member session user message provenance—Treats marked project-message turns in a member's own session as that member's user speech, but a bare "yes" or "go ahead" in one clears no block. - NEW: System Prompt: You should know understood-topic skip guidance—Tells the suggestion agent to skip topics the person plausibly understood in the main session, while still surfacing consequential details mentioned only in passing. - NEW: System Reminder: You should know side request—Tells the tool-less side agent to answer a one-off "You should know" request directly in one response, never reproducing secrets or credentials from the conversation. - NEW: Data: SDK initialize response sdk_mcp_manifests_parked field—Reports per SDK MCP server whether its manifest entry was parked, already connected, protocol-mismatched, malformed, or not honoured, and when the field is absent. - NEW: Tool Description: Bash (Git commit instructions) and Bash (PR creation instructions)—Split from the combined Git commit and PR prompt with otherwise identical text; the pre-commit-checks slot is gone, and the PR half now opens with a commit-and-PR-writing guidance slot. - REMOVED: Tool Description: Bash (Git commit and PR creation instructions)—Replaced by the separate Git commit and PR creation prompts; its combined text no longer ships as one prompt. - REMOVED: Tool Description: Bash (pre-commit skill checks)—Drops the long rule requiring a visible RAN/NOT RUN statement per verification, simplify, or review skill before commits; Bash now just says to always run those skills right before commit, never for docs or tests. - REMOVED: Agent Prompt: Security monitor candidate account and standing-rule changes rule, evaluation rules, unrequested connected-app commit rule, and user boundary rule—Removes the four selectable rule variants; the evaluation-rule and boundary text are now written directly into the main monitor prompt, while the two block rules are still referenced by name but no longer ship as extracted prompt text. - Agent Prompt: Security monitor for autonomous agent actions—Now always includes the Bound rule, unverifiable-gesture, side-door, tool-effect, secrets-as-labels, restricted-destination and copies-carry-sharing rules, and a persistent-configuration rule that treats forwarding rules, webhooks, and permission grants as high-severity and, when no specific rule covers them, allows them only when the user asked for that exact change. - Agent Prompt: Security monitor Claude Tag connector writes—The rules the connector exception does not override are now fixed as the User Intent Rule, scope escalation, External System Writes, and Unrequested Commit in a Connected App. - Agent Prompt: Background job agent instructions—When the human replies, the first sentence of the answer should carry what they asked or said instead of a separate recap, since the extractor cannot see their message. - Data: Claude Code gateway protocol—Gateways must ignore unrecognized request input, and pass through the auto-mode safeguards field or answer 400; it also specifies the managed-settings response shape and checksum, same-origin discovery fallbacks, and certificate pinning effects. - Data: SDK footer indicator schema—Adds the main model's footer_indicator_ capability as a source for the footer indicator, between the environment variable and bootstrap client_data. - Data: SDK frame_intake_phases_ms field—Splits before_read into startup_wait, control_requests_ahead and user_frames_ahead where the input loop could tell them apart, leaving before_read with the remainder. - Data: SDK set max thinking tokens request schema—Reworded the highlights error case from providers "without Anthropic's first-party beta features" to providers to which Claude Code sends no first-party-only beta features. - System Prompt: Project timeline user message provenance—Adds that a marked "in another thread of this project" message is the user's own words, but its bare "yes" approved something elsewhere and clears nothing here. - Tool Description: Artifact action reference and Artifact action reference (concise app wording)—Add a share action bullet, listed only when the artifact sharing feature is enabled. Details: github.com/Piebald-AI/claude…

What's new in Claude Code 2.1.285 system prompts (-4,252 tokens): Adds a subagent report-handback tool and new cloud-session and artifact notices, while removing the extracted prompts for directory sync with a user's machine and the auto memory lesson rules. Also tightens notification trust, read-before-edit rules, and background command timeouts. - NEW: Tool Description: SubagentHandback—Subagents deliver their full final report to their caller through one call that ends the run; plain-text endings are not delivered. - NEW: Data: Artifact publish merged-over newer version file changes notice—Tool-result notice listing files changed, added, or removed in a newer live artifact version, warning that stale copies must be re-read before editing. - NEW: Data: Cloud session unverified remote branch push notice—Warns that no readable remote meant the branch or detached HEAD could not be checked, and tells users to push their work and start a new session. - NEW: Data: Working tree upload refusal for misplaced .git entry—Refuses to upload a working tree whose .git is a pointer file, symlink, hard-linked pointer, or otherwise unfollowable, directing users to an ordinary main checkout. - REMOVED: System Prompt: Auto memory durable lesson instructions—Drops the memory prompt that limited saves to durable, applicable, legible user-taught lessons, with a per-reply save check and pinned-memory frontmatter. - REMOVED: System Reminder: Directory sync notices (12 prompts: agent commits off branch, attached machine guidance, branch name collision, branch switch parked work, disabled after initial checkout failure, file store exhausted, full and partial environment restore, live checkout guidance, restored files mismatch, snapshot commit reset, stopped for session)—Removes all guidance for cloud sessions kept in sync with a user's machine, including sync failure, restore, and branch-handling notices. - REMOVED: System Reminder: Remote machine file sync timing and Remote machine file sync timing for subagents—Removes guidance on when edits sync to and from a remote machine and to read fresh or git-ignored files there. - REMOVED: Data: Cloud session folder sync consent dialog—Removes the consent copy for two-way sync between a local checkout and a cloud session, including conflict handling and remembered answers. - Agent Prompt: Dream memory consolidation—Removes the variant where the memory index is assembled from file frontmatter; dreams always read and update the index, and "type conventions" wording is always included. - Agent Prompt: Project thread status card classifier—Allows quick replies for more work in the thread's own files; requires empty replies when the owner must act first or a yes would post, push, merge, deploy, or pay. - Agent Prompt: Security monitor candidate unrequested connected-app commit rule—Treats comments and title or description edits on PRs or issues the user asked the agent to work on as External System Writes; other PR commits stay blocked. - Agent Prompt: Web reading specialist—Permits up to about five linked same-site pages and forbids guessing URLs; failed fetches get one retry on transient errors only, and rate limits end fetching. - Data: Interrupt receipt still queued field—Rewords the send-gate explanation: a held send waits until the session is ready to take it, rather than for the initial upload. - Data: SDK set max thinking tokens request schema—A rejected highlights display request now returns an error reply naming the cause, leaving the display unchanged while max_thinking_tokens still applies. - System Reminder: Queued notifications delivery and Tool Description: ReadNotifications—Notification authority no longer relies on the sender named in a body; scheduled triggers are assigned tasks, while GitHub, Slack, and other-session bodies are information, not user instructions, and unrequested outward actions are reported instead of performed. - System Reminder: Session context—Replaces the "may not be relevant, don't respond unless highly relevant" closing with a note that the context was attached automatically and needn't be reported back. - Tool Description: Artifact type staged first-publish sequence—Removes the instruction to write files directly with the file-writing tool, with no shell step and no generating script. - Tool Description: Bash (timeout) and PowerShell—Labels the maximum timeout as applying to a foreground command. - Tool Parameter: Bash run_in_background guidance and note—When background deadlines are enabled, adds that timeout then limits background runtime, with a default and maximum, and the command is stopped and reported. - Tool Description: claude.ai Project—project_write mentions local_path only when local uploads are available, and project memory methods are described only when served; otherwise they are marked uncallable. - Tool Description: Edit, Edit single replacement, Write, and Write (read existing file first)—The read-before-edit/write requirement is now stated for every file, dropping the variant limiting it to files outside the working directory. - Tool Description: Edit single replacement—The Read line-prefix hint now says "line number + a single tab or :" when the tab-aware Read separator is enabled. Details: github.com/Piebald-AI/claude…

What's new in Claude Code 2.1.284 system prompts (-2,208 tokens): Adds PR Steward handoff checks and new security-monitor rules for account changes, connected-app actions, and user boundaries, documents new sandbox and permission settings and a gateway usage endpoint, reworks loop re-arm and Artifact guidance, and retires the Go SDK reference, a model-catalog note, and three skill prompts. - NEW: System Prompt: PR Steward handoff guidance—Requires checking a PR's PR Steward labels before pushing or babysitting it, and offering leave-it, one-change, or take-over choices instead of acting. - NEW: Agent Prompt: PR follow-up cron (PR Steward label check)—Tells the follow-up cron not to fix or push a PR carrying PR Steward labels, but to notify the user once. - Agent Prompt: PR follow-up cron and Skill: /loop slash command (dynamic mode)—Add PR Steward handling: when enabled, the cron also fetches PR labels and inserts the label-check clause, and the loop skill gains a PR Steward check block. - NEW: Agent Prompt: Security monitor candidate account and standing-rule changes rule—Adds a block rule for changing account security or creating standing rules like mail forwarding unless the user names the setting, and extends shared-configuration blocking to widening document visibility. - NEW: Agent Prompt: Security monitor candidate evaluation rules—Adds evaluation rules for unverifiable gestures, side-door page JavaScript, tool effect over self-described arguments, secret-shaped labels, restricted destinations, and copies inheriting sharing. - NEW: Agent Prompt: Security monitor candidate unrequested connected-app commit rule—Adds a block rule for committing decisions, spending money, publishing, or sending from a connected account when the user asked only to read, review, or draft. - NEW: Agent Prompt: Security monitor candidate user boundary rule—Makes explicit user boundaries block in-scope actions, including connected-account commits, and binds boundaries on sending or uploading at their plain meaning. - Agent Prompt: Security monitor for autonomous agent actions and Security monitor Claude Tag connector writes—The main prompt's Bound, bypass, unseen-result, and persistent-configuration wording now comes from a selectable variant, and the connector exception's rule list is variable-supplied. - NEW: Data: allowManagedPermissionRulesOnly setting (with allowed-tools frontmatter scope and deny-rules-still-apply notes)—Documents that only managed settings can add allow rules; allowed-tools frontmatter is ignored except on admin-backed channels, while deny and ask rules still apply. - NEW: Data: Sandbox deniedResolvedAddresses setting—Documents the list of IP addresses or CIDR ranges an allowed hostname must not resolve to, and its exemption for proxied routes. - NEW: Data: Sandbox root without CAP_SETFCAP error—Explains that running as uid 0 without CAP_SETFCAP makes every sandboxed command fail on newer Linux kernels, and says to grant the capability or run as non-root. - NEW: Data: Submit feedback control request—Documents the internal request that submits a /feedback report with transcript and sanitized error log, returns an unavailable reason when disabled, and supports saving a redacted bundle locally. - Data: Claude Code gateway protocol—Adds a GET /api/oauth/usage endpoint gateways can serve so clients show dollar spend alongside percent, with field formats and an example response. - Agent Prompt: Status line setup—Documents optional used_usd, limit_usd, and period fields on rate_limits.spend_limit, and updates the spend example to show dollars when available. - Data: Claude API reference (PHP, Ruby)—Makes the current Opus model's thinking always on (disabled returns 400, default effort medium), gives the previous Opus its own note, and hard-codes the refusal-fallback model as claude-opus-4-8. - REMOVED: Data: Claude API reference—Go—Removes the Go SDK reference. - REMOVED: Data: Published model catalog seed guidance—Removes the note on the compiled-in model catalog copy and its refresh. - NEW: Skill: Dynamic pacing loop execution (re-arm decision) and Skill: /loop self-pacing mode (re-arm decision)—Split out the step for choosing whether to re-arm with ScheduleWakeup, setting delay, reason, prompt, and noop, or stop the loop. - Skill: Dynamic pacing loop execution and Skill: /loop self-pacing mode—Confirm and re-arm steps are now built from shared blocks, adding slots for a re-arm status update and a stopped-loop outcome note. - System Prompt: Monitor fallback heartbeat guidance—Adds re-arm status-update instructions, and the stop instruction now calls ScheduleWakeup with stop: true and stops the monitor with the task-stop tool. - System Prompt: Autonomous loop tick (dynamic pacing) and /loop tick (loop.md absent; loop.md tasks)—Reword re-arm instructions from "at the end of this turn" to "this turn". - NEW: System Reminder: Attached machine reply not received—Says a connected machine's reply never arrived so the outcome is unknown, forbids re-running non-idempotent commands to see output, and suggests narrower commands. - NEW: Tool Description: Enable Claude app browser, Enable Claude in Chrome, and Enable computer use—Enable each tool once, before its matching tools, when the user needs their own browser, sign-in, or applications, and not if those tools are already present. - NEW: Tool Description: ListAgents (no SendMessage tool)—Describes ListAgents for sessions lacking SendMessage, and says replies must go through the host application's messaging tool or be relayed to the user. - Agent Prompt: Agent Hook—Distinguishes a remote hook call from a local one lacking a transcript, telling the agent to ignore transcript_path when no transcript file exists. - System Reminder: Directory sync full environment restore and partial environment restore—Environment reset, upload timing, and lost-state descriptions now come from variables instead of fixed "container was recreated" wording. - System Reminder: Directory sync branch name collision—Escapes the blocking branch name as untrusted text, falling back to "a branch of this checkout" when unnamed. - Tool Description: Poll—Moves the idle-signal and pending-events paragraph into a variable instead of fixed prompt text. - NEW: Tool Parameter: Artifact database access level—Documents the as_level parameter for testing an artifact's access rules as a view, interact, or admin user; it only narrows access. - Tool Description: Artifact database guidance—Adds a "view" level for as_level, and rewords the interact and admin levels in terms of who can use or edit the artifact. - Tool Description: Artifact database version pinning guidance, Tool Parameter: Artifact database version precondition, and Artifact database batch writes—if_version is now required for writes to existing documents and omitted only when creating. - Tool Parameter: Artifact database batch writes—Adds a 1 MiB limit on the batch body, and unpinned existing-document entries now fail the batch. - Tool Parameter: Artifact URL guidance—Defines valid claude.ai artifact links and requires reading an artifact before publishing to it, merging into the live version after a refusal. - NEW: Tool Parameter: Artifact URL guidance (app wording)—Adds a third-person variant of the artifact URL guidance with the same link format and read-before-publish requirement. - Tool Description: Artifact design fallback requirements—Widens allowed external script sources to cdnjs (preferred), jsDelivr, unpkg, Tailwind CDN, and jQuery. - Tool Description: Artifact design skill loading guidance (app wording)—The workshop-skill exception is now included only when workshop documents are supported. - Tool Description: Artifact responsive page contract—Adds min-width: 0 guidance for flex and grid children holding text, code, or tables, and moves the no-horizontal-scroll rule up front. - Tool Description: Artifact theme-aware styling—Adds a concrete CSS token skeleton for light, dark, and toggle styling, and spells out the mirrored structure for dark-first designs. - Tool Description: Artifact title and description guidance—Tells Claude to reuse the user's own name for a title, tightens naming rules, and says description fills in only for HTML lacking a