A Claude Code prompt that makes the agent investigate, restate and wait before coding
alex_verem · x · 2026-09-18
The author argues most agent failures come from the model guessing—your test framework, timeout behavior, intent—then building 400 lines on top. His CLAUDE.md prompt flips the order:
- Investigate before asking: anything findable in a minute in code/tests/configs is research, not a question
- Produce four things, then stop: a Goal restatement with acceptance criteria, 0-3 blocking questions each with a recommended default (so you can reply "yes to all"), numbered falsifiable assumptions, and a plan with files/signatures/order and rejected alternatives
- Scale by blast radius: trivial fixes skip ceremony; anything touching auth, money, migrations, or deletions gets full treatment
- If an assumption breaks mid-implementation, stop and report—no quiet redesigns
A "contractor who bills for rework" mindset that avoids costly wrong guesses.
More from coding & agent
- Sim Search lets agents build a knowledge graph across 1,000+ tool integrations — JafarNajafov · 2026-09-18
- Ethan Mollick rebuilds Umberto Eco's 33,000-book library in 3D with Claude Projects — emollick · 2026-09-18
- Cua and typesafeai launch jev-use dev preview, claiming fast computer use is solved — TianbaoX · 2026-09-18
- Developer demos auto-navigating slides driven by the Jev computer-use model — threepointone · 2026-09-18
- DeepSeek Harness ships v0.1.6-alpha.2 pre-release as the app takes shape — teortaxesTex · 2026-09-18
- Self-Proclaimed ChatGPT Co-Inventor Launches Jev, Claims 200x Speed at 1/400 Cost — iamrobotbear · 2026-09-18