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
- Tweets150
- Following57
- Followers38
- Likes15
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…
Why: earlier messages are cached, and cached input is billed at a fraction of the normal rate. A fresh session has to re-read the code first, and that cost more than carrying the history.
Caveat: same files every time is the best case for staying.
The recipe: don't /clear reflexively when the next task is in the same code. Check /context to see what's filling the window. When one task runs long, /compact with a hint about what to keep.
codetwig.dev/recipes/keep-co…
7+ hours on full autopilot. Zero limits hit. Zero intervention.
My setup in the new Claude Projects beta:
→ Spent 30 min up front explaining the vision, the goals, and how phases hand off to new sessions
→ One rule: "Don't offer to write a single line of code until you're confident the whole plan will make the product launch-ready."
→ Break it into phases. Plan each one in full, update the todo, close the session, start a fresh one for the next phase. Repeat.
→ Opus 5.5 for every session
It's still writing a planning guide for each phase.
The plan is the product now. The code is the last step.
Does adding "think hard" help Claude Code on tricky logic? We ran a billing-date function full of edge cases 6 times. Plain prompt or think-hard, all 6 passed every hidden test. The real difference: only the think-hard runs wrote and ran their own tests (3 of 3 vs 0 of 3), at $0.091 vs $0.057 a run.
Telling Claude Code to delegate searches to a subagent made the job 24% more expensive and saved nothing.
Both setups went 3 of 3. The search step alone cost 53% more with a subagent. 🧵
The main session didn't get leaner. The next step edited the same files, so Claude opened all six anyway. Peak context: 31,200 tokens with a subagent vs 30,200 without.
Caveat: ~470 lines is smaller than subagents are built for. Big repos are untested.
The recipe: don't tell Claude to delegate every search. Use a subagent when the reading is big, the answer is small, and you won't need those files after. Check CLAUDE.md for blanket rules; Claude follows them even when they cost you.
codetwig.dev/recipes/subagen…
