Tested AI-coding recipes. We run Claude Code and local models on real tasks and post the actual numbers. New recipes Mon/Wed/Fri.
Joined October 2026
- Tweets192
- Following60
- Followers44
- Likes22
No one is claiming that Terra is an opus replacement. Nor should it be looked at in that way. Each model is best at something and this is typically the way I would use them. You don’t need to run Astra for everything. Your tokens could last a lot longer using them the right way.
GPT-6 Astra
For the hardest problems: architecture, deep research, complex debugging, and critical reviews. The pick when quality and judgment matter most.
GPT-6.1 Sol
The coding workhorse. Great for building features, refactoring, writing tests, and handling substantial projects while keeping usage reasonable.
GPT-6.1 Terra
For visual reasoning, browser/app tasks, and clear writing. A good fit for screenshots, visual QA, polishing content, and small, well-defined code fixes.
GPT-6 Luna
For focused, repeatable tasks: summaries, extraction, classification, and simple edits. The efficient option when you don’t need heavy reasoning.
Tested /clear vs one long Claude Code session on a 197-file repo: 3 unrelated tasks, 3 runs each way. Every run passed all hidden tests, and one session was 14% cheaper ($0.137 vs $0.159 a run), but its context grew to ~31K tokens vs ~21K.
I feel everyone who says gpt terra is terrible is because they don’t know how to use models properly and just set max everything and run with it.
You don’t need every model to be the best at everything
Does /clear pay off on a bigger repo? We ran 3 unrelated tasks on a 197-file, 14-module codebase, 3 times each way. All 6 runs passed every hidden test. Staying in one session cost $0.137 a run vs $0.159 with /clear, 14% less.
codetwig.dev/recipes/clear-o…
The cheapest way to get a rule followed was one CLAUDE.md line, a hair ahead of a skill.
Claude Code followed a house formatting rule 4 ways, 3 runs each, and every file came out correct in all 12. Averages: CLAUDE.md $0.083, skill $0.084, hook $0.101, nothing $0.110. 🧵
The hook stamped every file by itself, but Claude still read the hook config and checked the result, so it cost about as much as nothing.
Also: 11 of 12 first file writes were shell commands (cat > file), not Write. A hook matched only on Edit|Write could miss those.
The recipe: hook when it must never be skipped (also match Bash). CLAUDE.md line for a simple always-on rule. Skill for a procedure that only matters sometimes. 3 runs each, so treat the cost ranking as a hint.
codetwig.dev/recipes/skills-…
What are you shipping this week with Claude Code, Cursor, or whatever agent you're running? One line is plenty.
We swapped Kimi K3.1 Pro into Claude Code and ran a tricky billing-date task 5 times against Sonnet 5.5. Both passed every hidden test, 5 of 5, but Kimi cost $0.127 a run vs $0.054 and took 148s vs 18s.
codetwig.dev/recipes/claude-…
If you're not using @claudeai's Projects beta, you're sleeping on it.
Before starting, I gave it one instruction:
"Don't write a single line of code until we have a 100% launch-ready plan for every feature. Then we review and start coding."
It spent a full day on the plan. We're now 4 days into executing it.
Haven't hit a single hourly limit. Sitting at 80% of my weekly.
Projects beta is fantastic.
The one miss: in 1 single-session run, Claude rejected a pageSize over 50 instead of capping it. Our prompt said "max 50", which supports either reading. One session 2 of 3, fresh sessions 3 of 3.
Lesson: write "cap at 50" or "reject above 50".
The recipe: in a small codebase, don't clear just because the topic changed. Clear when the context gets big or misleading (/context shows it). In a large codebase with long tasks this could flip; we haven't tested that.
codetwig.dev/recipes/clear-b…
