Pinned Tweet
This is f*cking wild. An Anthropic engineer opens the token meter on Claude Code live:
"In the era of agency coding, we're paying for the work itself, right, task by task, you're paying per prompt.
So understanding how to be efficient with that is just part of the job now."
In 27 minutes, an Anthropic engineer shows how a one-line prompt becomes tool calls, cache breaks, model multipliers and team spend.
Prompt → Context → Cache → Subagents → Spend Control
Anyone can type a prompt. Almost nobody knows what that prompt costs after 400 lines of test output.
Bookmark and watch it whenever you feel like quitting
Then read how to onboard 5 bot accountants: the Onboarding Playbook ↓
1
20
49
This is f*cking insane. An Anthropic engineer took a bloated AI agent apart live and showed exactly how they fix it
29 minutes. No hype. Just Will from Anthropic's Applied AI team rebuilding an agent step by step
Evals → System Prompt → Skills → Tools → Subagents
Most people keep bolting features onto one giant prompt. His agent had a 400-line system prompt, 12 tools and 3 subagents, and its evals kept dropping
"So this isn't a model problem. It's an issue with the information that we're surrounding the model with"
Anyone can write a prompt. Almost nobody knows when to use a tool, a skill or a subagent
Worth more than any $2,000 AI course
Bookmark and watch it today
Then read why Claude Code's creator deletes his whole harness every 6 months in the guide below ↓
Holy f*ck, a winning streak can become a shareholder problem.
"And you have to balance the good that someone does and that brings relative to the bad"
"A lot of other people have to"
The move:
- Treat elite performance as an asset and a risk
- Weigh what a star brings against what the streak hides
- Be willing to question a setup that looks like winning
Try it: "Claude Code, map the hidden risks in this high-performing team or workflow. Explain the tradeoffs first, identify what the winning streak may be masking, and do not edit anything until I say go."
Then read how the model, the harness and you fit together ↓
1
8
217
Holy f*ck, AlphaFold just folded 200 million proteins in one year
"And so we managed to actually fold, using AlphaFold, in one year, all those 200 million proteins known to science."
"So that's a billion years of PhD time."
"So within the width of an atom, on average, is how accurate the prediction is, which is what is needed for biologists to use it and for drug design and for disease understanding."
The move:
- Start with the full search space
- Use the model to compress impossible time
- Aim for accuracy people can actually use
Try it: "Plan first. Identify the expensive bottleneck in this workflow, explain how an AI model could compress it, define the accuracy needed for real use, and do not edit anything until I say go."
Then read how the model, the harness and you fit together ↓
2
21
153
Holy f*ck, AI training just hit a 1 trillion token wall that gradients cannot handle
"... there's no chance you can compute all the gradient"
"... even computing one gradient is very costly"
The move:
- Stop assuming full gradients scale
- Use a sampler that costs less
- Treat data scale as an engineering constraint
Try it: "Explain when full gradient computation stops being practical as training data grows, then compare cheaper sampling strategies and their tradeoffs. Plan only, no edits until I say go."
In 30 seconds, the clip explains why 2015 had 1 million examples and 2026 has 1 trillion tokens. If you build AI products, this decides what can train, what costs money, and what gets shipped.
Then read how the model, the harness and you fit together ↓
2
23
205
Holy f*ck, a DEXA scan just turned a fast into a muscle-loss warning
"I lost like six pounds of muscle"
"It was like pretty Depressing how much muscle I lost"
The move:
- Check what Google says before the fast
- Do a DEXA scan before the extended fast
- Do another DEXA scan after
Try it: "Explain how to compare before-and-after body composition scans for a fasting experiment, list what each number means, and flag any claims that conflict with common advice. Do not edit anything until I say go."
Then read how the model, the harness and you fit together ↓
21
273
2 or 3 days
He, Anthropic technical staff, says Claude Code cuts technical onboarding from 2 or 3 weeks to 2 or 3 days.
"At Anthropic, onboarding used to take about two or three weeks for technical hires. It's now about two or three days."
Try it: "Don't change anything until I say go. First explain how this codebase works and where I should start."
5
21
160
FOURTH TRY, AND IT'S STILL UP.
He says the Great Pyramid is the fourth attempt after two failures, and the bend is still visible.
"No, I've done that. You're halfway in, the numbers are wrong, and you know it."
Try it: "Plan the diff first, don't change anything until I say go."
5
21
556
That memory thing is f*cking real now.
He says Claude Opus 4.7 just made file-system memory state of the art.
"With Claude Opus 4.7... we saw that it was state of the art at file system based memory... it all does this with just bash tools and grep tools."
He's already running thousands of parallel agents on it.
8
25
584
Claude Code's creator says the most common mistake is over-specifying the task.
~ Boris Cherny, creator of Claude Code (YC talk)
"and then just go let the model cook"
His recipe: a task slightly harder than you think it can handle, then the task, the guardrails and the exit criteria, not step one, two, three. He says this wouldn't have worked six months ago.
Try it:
"Task: <one paragraph>. Guardrails: <what you must not touch>. Done means: <the tests or checks that must pass>. Plan it yourself, then show me the plan and the diff before applying anything. Don't change files until I say go."
5
26
578
Claude's engineering lead doesn't let a session grade its own work. He looks for an outside signal.
~ Felix Rieseberg, who leads engineering for Claude at Anthropic
"the list of things where I don't trust one single session is actually much larger than the list of things I wouldn't trust Claude with."
In a pinch, he says, the outside signal can be another Claude session. His preference is hard evidence.
Try it, in a fresh session:
"You didn't write the changes on this branch. Review the diff and split your findings into two lists: what you verified with hard evidence (a test run, a command and its output) and what you're only assuming. Don't change any code."
1
22
491
NVIDIA's CEO is gone. CFO Colette Kress is the interim CEO, and the stock fell more than 10% before the open.
What Jensen Huang once said about ever feeling unnecessary:
He says - already works with superhuman intelligence every day: the thousands of people around him who are better than him at what they do.
"And yet, it never one day caused me to think, all of a sudden, I'm no longer necessary."
His bet for the next decade: AI becomes superhuman at some things, not everything.
Jensen Huang is out at NVIDIA.
The developer bet he called "a Linux moment" is still on the table. In his words:
Huang calls OpenClaw a Linux moment: an operating system built around a language model.
"This is the operating system that's going to hold a large language model."
He says he offered OpenClaw's creator the help of NVIDIA's engineers, because everybody should be able to build their own AI.
Try it:
"Describe my Claude Code setup (CLAUDE.md, skills, hooks, MCP servers) as an operating system: what plays the kernel, the apps, the memory and the permissions. List the gaps. Show any change as a diff and don't change anything until I say go."
18
506
Claude Code's code review doesn't just find bugs. It argues with itself about each one before showing you.
~ the Claude Code team at Anthropic, on how workflows came about
"for each bug you might do an adversarial review where you ask it to look at the bug from three different opinions or perspectives and see if the bug is actually real."
Their point: that filters a big fan-out of findings down to the few that need your attention.
Try it, with plain prompting:
"Review the diff on this branch for bugs. For each bug you find, check it from three different perspectives (correctness, edge cases, whether existing code already prevents it) and drop the ones that don't hold up. List only the survivors with file and line. Don't change any code."
3
15
455
FORKLINE retweeted
A 400-line system prompt became a 50-line one. Most of the rest moved into skills.
~ Will, Applied AI at Anthropic (Code with Claude London)
"Leave the system prompt only for the information that Claude needs in its mind, regardless of the task that you give it. Skills are fantastic for packaging information that Claude is going to need some of the time, not all of the time."
In Claude Code the split is the same: a skill's body loads only when it's used, while CLAUDE.md loads every session. A CLAUDE.md section that grew into a procedure is a skill.
Try it:
"Read my CLAUDE.md. Find sections that are procedures, not facts, and propose moving them into skills under .claude/skills/. Show me the diff first and don't change anything until I say go."
A 400-line system prompt became a 50-line one. Most of the rest moved into skills.
~ Will, Applied AI at Anthropic (Code with Claude London)
"Leave the system prompt only for the information that Claude needs in its mind, regardless of the task that you give it. Skills are fantastic for packaging information that Claude is going to need some of the time, not all of the time."
In Claude Code the split is the same: a skill's body loads only when it's used, while CLAUDE.md loads every session. A CLAUDE.md section that grew into a procedure is a skill.
Try it:
"Read my CLAUDE.md. Find sections that are procedures, not facts, and propose moving them into skills under .claude/skills/. Show me the diff first and don't change anything until I say go."
FORKLINE retweeted
Same tasks, one long Claude Code session: twice the tokens of running /clear in between.
~ Anthropic's webinar on how Claude Code spends tokens
"The model only ever adds to your context. It never removes anything."
In their demo, tasks two and three kept carrying task one around the whole time. /clear starts small again. /compact replaces the conversation with a summary, and you can tell it what to keep. /rename first, and /resume brings the old session back.
Try it:
"/compact Focus on the failing tests and the files we changed"
Same tasks, one long Claude Code session: twice the tokens of running /clear in between.
~ Anthropic's webinar on how Claude Code spends tokens
"The model only ever adds to your context. It never removes anything."
In their demo, tasks two and three kept carrying task one around the whole time. /clear starts small again. /compact replaces the conversation with a summary, and you can tell it what to keep. /rename first, and /resume brings the old session back.
Try it:
"/compact Focus on the failing tests and the files we changed"
+10% software engineer jobs. +20% radiology jobs. AI automated the core task of both.
That's Jensen Huang's evidence, from the NVIDIA CEO's talk to students:
"The narrative about AI destroying jobs is exactly backwards. AI automates tasks away, but it doesn't necessarily eliminate jobs."
His example: "They said Harvey is going to eliminate all of the paralegal jobs... Turns out paralegals are growing like crazy."
The figures are his, quoted from the talk.
Claude Code decides what to run. A hook decides whether it's allowed to.
~ Lydia, Claude Code team at Anthropic
"So pre-tool use, for example, you can add even more details of whether you want that tool to run or not, because again, a tool call could do a lot in your machine. So if you want to block certain things, hooks are great."
PreToolUse fires before a tool call and can block it. PostToolUse fires after, which is where formatting or linting goes. They live in settings.json, and /hooks shows every one you have.
Try it:
"Add a PreToolUse hook to .claude/settings.json that blocks any command touching .env files. Show me the diff first and don't change anything until I say go."
16
362







