← ClaudeAtlas

root-causelisted

Run a true, recursive root-cause analysis. Use whenever the user asks for an RCA, a "root cause", why something happened, broke or regressed, or to diagnose the cause of a bug, incident, outage or wrong output. Enforces evidence-backed recursion from proximate cause to root cause. No "most likely", and no stopping at the first cause. Reports in SITREP shape.
openhonest/honest-skills · ★ 0 · Code & Development · score 75
Install: claude install-skill openhonest/honest-skills
# Recursive Root-Cause Analysis The user asked *why* something happened. They do not want a plausible story. They want the actual causal chain, with each link proven, followed down to the cause(s) of the cause(s) until you hit a root: a design decision, a process gap, or something outside the system boundary. Follow this method. Do not shortcut it. ## The one rule that matters **Every link in the chain is a verified fact, not a guess.** If you cannot show the evidence for a link, meaning a query result, a log line, a file you read, a value you reproduced, then you have not found that link yet. Keep digging or say you don't know. Never write "most likely," "probably," "it seems," or "this is likely because" as a substitute for looking. If the user pushes back with "I don't want the most likely" or "verify it," that is a signal you skipped this rule. Go back and prove the link. ## The loop Start at the symptom the user can see. Then repeat until you reach a root: 1. **State the current effect** in one concrete sentence (what is observed, with the specific case/value). 2. **Find its immediate cause.** The one thing directly upstream that produced this effect. Not the ultimate cause. The *next* one up. 3. **Prove it.** Run the query, read the file, pull the log, reproduce the value. Show the evidence inline. 4. **Ask "why is *that* true?"** and make that the new effect. Go to 1. You are done with a branch when the cause is one of: - a **design or process decision** (e.g.