メニュー
Session retrospective that turns this session's transcript into at most 10 principle-level knowledge proposals, each behind explicit user approval. Run at the end of a substantial session, or when the user asks for a retro.
デフォルトでは、最初にソースを確認する Prompt が選択されています。直接コマンドに切り替えるか、ローカルコピーをダウンロードすることもできます。
インストールを決める前に、SKILL.md と SkillsMP に表示されている付属ファイルをお読みください。
SOC 職業分類に基づく
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
直接コマンドでは確認用 Prompt が省略されます。実行前にソースを確認してください。
npx skills add https://github.com/jLAM-ERR/autodidact-plugins --skill retroコマンドは1行のまま表示されます。コピー前に横へスクロールして全体を確認してください。
ローカルで確認しますか？SkillsMP が現在取得できるファイルをダウンロードできます。
SKILL.md を表示中
Capture a learning into the project knowledge base as an append-only candidate entry. Use after a workflow gate decision (approve/correct/decline), after a self-review finding in auto mode, or when a good answer was assembled mid-task that should not die in chat history (the file-back rule).
Promote a project knowledge entry that generalizes across projects into the shared autodidact knowledge base via PR, with the stricter destination secret gate and a redaction checklist. Run when the user asks to promote, or when distill flagged promote-eligible entries.
Drive a task through the generic gated workflow — intake, design panel, implementation, hardening and review, wrap-up — with explicit human gates, resumable state, and every gate decision captured into the project knowledge base. Use for non-trivial tasks that deserve staged delivery; input is a task description or spec file, no ticket system required.
| name | retro |
| description | Session retrospective that turns this session's transcript into at most 10 principle-level knowledge proposals, each behind explicit user approval. Run at the end of a substantial session, or when the user asks for a retro. |
| invocation | user |
| disable-model-invocation | true |
Evidence-based shape (autodidact methodology/learning-loop.md): deterministic
pre-pass → parallel fresh-context analysts → single curator merge →
per-proposal approval. Never re-review the same context repeatedly — one
fresh pass beats looping.
Pre-pass (deterministic). Export or locate the session transcript, then:
python3 "$(dirname "$SKILL_PATH")/scripts/signals.py" scan \
--transcript <transcript-file>
The JSON lists corrections (dominant signal), errors, and repeated-failure clusters, secret-redacted.
Parallel lens analysts (fresh context). Spawn three subagents in one batch — each gets ONLY the transcript path + the signals JSON, none of this session's conversation:
Curator merge (you). Merge the three lists: dedupe, drop anything
that is really codebase-derivable, keep at most 10 proposals
(CONTRACT §Tunables). Before finalizing, read
.autodidact/retro-acceptance.md if present and deprioritize finding types
the user has repeatedly rejected.
Per-proposal approval. Present each proposal separately (topic, the principle, suggested outcome mapping). Nothing is written silently.
Capture accepted proposals through the capture CLI
(--source retro; outcome: approve for patterns that worked,
correct for corrected behavior, decline for things to avoid) — see
the capture skill; never hand-write entries.
Log acceptance. Append one line per proposal to
.autodidact/retro-acceptance.md (create the dir if needed):
YYYY-MM-DD accepted|rejected <lens> — <topic>. This log is what makes
step 3's self-tuning possible.