← ClaudeAtlas

staged-coding-9.5-refutelisted

staged-coding stage 9.5 - the refuter. A FRESH agent on a model different from the builder, handed only the diff, tries to prove the FIX wrong on its own path and keeps attacking each repaired form until a round comes back clean. Blocking - a kill sends the fix back to stage 4. Signs its ledger line on finish. Triggers: /staged-coding-9.5-refute, or invoked in order by /staged-coding.
Shadysmetools/claude-skills · ★ 0 · AI & Automation · score 72
Install: claude install-skill Shadysmetools/claude-skills
# Stage 9.5 - the refuter Part of the `/staged-coding` chain. Runs after stage 9 (verification) and before stage 10 (land). The conductor carries the cross-cutting rules. **This is not the council (stage 3).** The council runs BEFORE code exists, on the DESIGN, only on its markers, and is advisory. The refuter runs AFTER code exists, on the ACTUAL DIFF, on EVERY fix, and is blocking. Neither can substitute for the other: a council can find that the obvious design does nothing; only a refuter can find that the built code missed a path. ## Inputs - **The diff itself.** Never a description of the bug it claims to fix: a defect description invites the refuter to re-confirm the defect instead of reviewing the fix. Attach the diff; do not summarize it. - **Never** the builder's reasoning, file:line citations, queries, method, PR body or ticket text. - The findings stages 6 and 9 already caught, so the refuter hunts what they missed. ## Who runs it A FRESH subagent on `config.checker_model`, handed only the inputs above, never the builder's chat history. **The model must differ from every agent that touched the fix**, author and earlier checkers included. If the usual split collapsed (an override or fallback put the builder on the checker model), pick a model none of them used and say so. Four agents on one model agreeing is one agent agreeing with itself four times. **Why the path matters more than the headcount:** a second agent handed the author's own query will "confirm