← ClaudeAtlas

self-ratelisted

Use before returning a draft, answer, plan, or code change you're unsure about, or when the user asks how confident you are, how good it is, or to grade/score/critique your own work. Triggers when claims may be overstated, when output quality is uncertain, or when you want to catch weak spots before the user does rather than after.
mickzijdel/dev-hooks · ★ 0 · AI & Automation · score 68
Install: claude install-skill mickzijdel/dev-hooks
# Self-Rate Score your own work on a **calibrated** scale, then change the work to match the score — lower the claims you can't back, or do more work to earn a higher one. Honest self-assessment beats confident hand-waving. Pairs with [[but-for-real]] (which verifies) — this one *grades*. ## Why calibrated A number means nothing without anchors. Define what the scale's points actually mean *before* you score, or every answer drifts to "8/10, looks good". **Default 1–10 anchors:** - **1–3** — likely wrong, unverified, or missing the point; would embarrass me if shipped - **4–6** — plausible but has known gaps, untested claims, or weak spots I can name - **7–8** — solid; verified the core, minor caveats remain - **9–10** — verified end-to-end, edge cases considered, I'd bet on it ## Procedure 1. **Pick dimensions** that fit the work — typically **correctness/evidence, completeness, clarity, risk**. For a single factual claim, one "confidence in this claim" score is enough. 2. **Score each dimension** against the anchors, with a one-line justification naming the *specific* reason — "6: didn't run the migration on a populated table", not "6: seems okay". 3. **Act on the score** — this is the point, not the number: - Low score → add the caveat explicitly, or do the missing work to raise it - Overclaim → downgrade the wording to match what you actually verified - High score you can't justify → it's not high; re-score honestly 4. **Return** the scores + the revised wo