← ClaudeAtlas

plan-featurelisted

The plan-first workflow for any non-trivial change - explore with subagents, write a plan with explicit "done means" verification criteria, get sign-off, implement, verify.
Vinylfigure/janus · ★ 0 · AI & Automation · score 74
Install: claude install-skill Vinylfigure/janus
The Cherny flow: plan mode first, code second, and a closed feedback loop at the end. Exploration is delegated so this thread's workspace stays clean for the actual reasoning. ## Hold in mind 1. No code until a plan exists — and a plan without verification criteria is not a plan. 2. "Done means" is an exact command or observation, not a vibe ("tests pass after `verify.sh full`", not "it works"). 3. Subagents explore; this thread decides. Do not fill this context with file dumps. 4. Simple tasks skip ceremony: if this is genuinely a one-file, obvious change, say so and just do it with the inner verify loop. 5. The plan is an audit artifact: the Predicted-failure-modes section and its end-of-run accounting are how someone who wasn't in the room — the verifier included — confirms the adversarial pass actually ran. ## Steps 1. Restate the requirement in <= 3 sentences, then list the invariants at stake (what must not break, what must remain true). This is your working set — refer back to it. 2. Classify the known unknowns — everything the plan depends on that you don't yet know — and route each to its resolution: codebase question → an exploration subagent; requirement or preference → interview the user; feasibility → a small prototype or probe. In an unfamiliar domain, add a blind-spot pass: "what would an expert check that I haven't thought to?" — route those too. An unrouted unknown is a silent assumption. 3. Explore via subagents, in parallel where independent: a read-onl