← ClaudeAtlas

record-verified-lessonlisted

Use when the agent believes it has learned something. Routes the belief through claim → evidence → confirmed instead of recording a confident guess. Prefers amplify_record_claim for unverified inferences and promotes only with real evidence.
Sisuthros/claude-amplifier · ★ 1 · AI & Automation · score 72
Install: claude install-skill Sisuthros/claude-amplifier
# Record a Verified Lesson The core failure this product exists to prevent: **an inference silently becomes a memory, the memory becomes a "fact," and the fact never gets verified.** This card is the procedure that keeps a new belief honest about its own status. Every lesson lives at one of three statuses: ``` claim (0.5 confidence) → evidence (0.7) → confirmed (1.0) ``` A `claim` weighs **5× less** than a `confirmed` lesson when the Pattern Oracle scores risk. So a wrong guess recorded as a claim is cheap; a wrong guess recorded as a confirmed lesson is the exact bug we are avoiding. ## When to use Any time you are about to write down something you "learned" — a fix, a gotcha, a rule of thumb, a cause→effect. ## Procedure 1. **Is it verified right now?** - **No (an inference, a theory, "this probably fixed it")** → record it as a claim: `amplify_record_claim({ project, title, description, ... })`. Stop here. Do not dress a guess up as a confirmed lesson. - **Yes (you have real evidence in hand)** → continue. 2. **Pick the pattern_key carefully.** Before inventing a new `pattern_key`, run `amplify_suggest_pattern_key({ text })`. If it returns a close existing key, reuse it so the same mistake-in-different-words aggregates into one lesson with a frequency counter, instead of fragmenting. 3. **Promote with evidence.** Promote a claim to confirmed with `amplify_verify_claim(...)` **only** after attaching real evidence. Valid evidence ty