rcalisted

# RCA: failure-first investigation The point is **not** to fix a bug. It is to understand *why* it happened deeply enough that the fix kills the cause, not the symptom — and catches the class next time, not the instance. `/blueprint` plans a change you want; `/rca` investigates a failure you didn't. Invoke for regressions, incidents and post-mortems, a first-pass fix the user finds suspicious, or any failure the team would otherwise pay for again. Not for trivial bugs with obvious causes, or for future failures you are designing around — that is `/blueprint`. ## Required references This file is the skeleton; each reference holds a step's full rules and worked examples. Read it **at that step**. | File | Read at | Holds | |---|---|---| | `references/chain.md` | step 3 | 5-whys chain with evidence per link, cause classification and falsification, assumption list | | `references/fix.md` | step 6 | symptom vs root-cause fixes, counterfactual, prevention, report structure | | `references/rationalizations.md` | when tempted to skip a step | why each shortcut fails | ## Principles - **Repro first.** A bug you can't trigger is a story; until you reproduce it or read a trace of it firing, every theory is unfalsifiable. - **The symptom is evidence of the cause, not the cause.** Fixing it holds until the next adjacent failure. - **Ask why mechanically.** Each answer is the next question's subject; stop at the level whose fix prevents siblings. A root cause with no siblings