メニュー
Lint and compact the ultragoal project memory — find contradictions and stale claims, merge duplicates, refresh the index. Run every ~10 sessions (the session banner nudges when due).
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。
SOC 職業分類に基づく
Ultragoal for Codex. Use when the user asks for ultragoal in Codex, a hook-backed or runner-backed Codex goal loop, a rubric-backed Codex /goal, durable goal execution, evidence-led completion, panel verification, or to turn a brain dump into a Codex-native goal. Do not use for trivial one-turn tasks unless explicitly invoked.
Turn a messy brain dump (raw voice transcript welcome) into a verifiable goal with a checkable rubric, then work autonomously until it is verified and lessons are saved. Use for substantial end-to-end work — build, fix, migrate, investigate — AND for follow-up rounds on a finished or paused goal ("next round", "improve on this"); re-invoke it rather than arming from memory of a previous round. NOT for small contained tasks (a quick fix, a question, a tweak) — those are ordinary conversation, and a goal having run earlier in the session never makes later small asks goal work.
Initialize or reconfigure ultragoal in this project — scaffold .ultragoal/, choose preference knobs, wire the CLAUDE.md block. Runs automatically on first /ultragoal:goal.
Distill lessons from the current session or a finished goal into the project memory (.ultragoal/memory/). Used directly, as the final step of a goal loop, and immediately whenever the user corrects you.
Show the current ultragoal state — active goal, rubric progress, budget, memory health.
Stop the active ultragoal loop — archive the goal as abandoned (or paused) and release the gate.
| name | compact |
| description | Lint and compact the ultragoal project memory — find contradictions and stale claims, merge duplicates, refresh the index. Run every ~10 sessions (the session banner nudges when due). |
Run a lint-and-compact pass over .ultragoal/memory/ (MEMORY.md, facts.md, patterns.md, failures.md).
The cardinal rule: only the compiled layer (above the ---) is ever rewritten. Evidence logs below the line are append-only history — never edit, reorder, or delete their lines. Compaction that touches evidence is corruption, not cleanup.
Read every file end to end, then report findings in three tiers before changing anything:
[VERIFIED] claims whose evidence is old and whose subject area has changed since (check git log on the relevant paths); [INFERRED] entries no session has touched since they were written; entries the repo or git history now records on its own.[no data yet] slots in MEMORY.md that sessions keep needing; areas with lots of evidence but no compiled rule yet.[VERIFIED] claim has a one-command check, run it. Passes → refresh its date. Fails → rewrite the claim, append the contradicting run to the evidence log.[INFERRED], not deleted. Never silently promote [INFERRED] or [READ] to [VERIFIED] — that laundering is exactly how a memory starts citing itself.## Archived subsection at the bottom of the compiled layer with a one-line reason; their evidence stays in the log untouched.Reset the cadence counter: write 0 to .ultragoal/memory/.sessions.
Report: Errors/Warnings/Info found, what was merged, generalized, re-verified, downgraded, archived, and the new index size. Memory should come out smaller and sharper, not summarized into mush — when in doubt between compressing an entry and keeping the specifics that would change what a future session does, keep the specifics.