← ClaudeAtlas

awb-lessons-capturelisted

WHAT: an end-of-session discipline that mines the just-finished session for durable lessons, scores each against a fixed bar, and writes only the survivors you approve to the live per-project memory so a future session recalls them. It is an honesty overlay on native auto-memory, not a capture engine: it adds the report-zero gate, a value score, dedup against the existing corpus, and a hard write-gate. USE WHEN: wrapping up a session and you want to keep what was learned ("capture the lessons", "memory retro", "what is worth remembering from this", "save the lessons before we stop"), especially after a surprising bug, a correction, or a counter-intuitive trap. DO NOT TRIGGER: a quick mid-session note (just jot it down); packaging unfinished work for the next session (that is a handover skill); promoting an already-captured recurring lesson to a rule (a human call in the lessons doc); a session where nothing surprising happened, where the honest output is simply zero.
doivamong/agent-workbench · ★ 2 · AI & Automation · score 78
Install: claude install-skill doivamong/agent-workbench
# Capture the lessons a session taught — without manufacturing them > **Announce on activation:** "Using awb-lessons-capture — I'll mine this session for durable lessons, score them against a fixed bar, and write only what you approve to the live memory dir." Most sessions teach nothing worth keeping, and a retro that always "finds three key learnings" is inventing them to look productive — that noise is exactly what buries the rare lesson that mattered. This skill imposes a fixed bar so the same junk gets culled every time and the same gem gets kept every time. It runs at the **end** of a session, when invoked. It does not invent a lesson to have an output: **zero is the common, correct answer.** On a harness that captures memory natively (Claude Code v2.1.59+), the harness already writes memory for you. This skill is not a second capture engine — it is the *discipline* the raw capture lacks: a report-zero honesty gate, an explicit value score, a dedup pass against the corpus you already have, and a hard gate that writes **nothing** until you approve. ## Scope - **Does:** at end of session, surface candidate lessons, gate and score them, and — after your approval — write the survivors to the **live per-project memory dir** so they are recalled later. Reads the corpus to avoid duplicates; suggests (never performs) promotion. - **Does NOT:** capture mid-session (jot those down), package unfinished work for a successor (that is a handover skill), or promote a lesson