メニュー
Manage the BCOS self-learning ladder: list learned rules, forget a rule, regenerate learned-rules.json from resolutions.jsonl. Mechanical-only — reads/writes derived artifacts; the actual learning happens via record_resolution.py at click time. Invoke with /learning.
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
Single-task dispatcher for CLEAR Context OS scheduled maintenance. Runs from the daily scheduled task, reads schedule-config.json, determines which jobs are due today, runs them in sequence, appends to the diary, writes a consolidated morning digest, and surfaces any frequency-tuning suggestions. WHEN TO USE: - Invoked automatically each morning by the `bcos-{PROJECT}` scheduled task - User says "run today's maintenance now" (manual dispatch) - User says "run the [job-name] job" (runs one job on demand, skips others)
Strategic reflection skill. Step back from daily work to think about the bigger picture of your context architecture, identify gaps, and envision improvements.
Wiki zone manager for BCOS — the universal long-form and cross-cutting content destination per plugin-storage-contract.md Rule 2. Stores runbooks, SOPs, decision logs, how-tos, glossaries, post-mortems, FAQs, meeting notes, and plugin cross-cutting context (charters, transcripts, research, customer-call notes). Two ingest paths: Path B (local files, pasted text, inbox promotion — runbooks/scripts/SOPs/notes) and Path A (external URLs / GitHub / YouTube). Schema-validated frontmatter, banner citations, builds-on graph for references, two-tier refresh, supersedes chains. Schema-driven, CLEAR-governed, derived-artifact. Invoke with /wiki.
The single entry point for getting anything new into your context architecture. Takes any input (documents, URLs, YouTube, meeting notes, pasted content) and triages it: dump to _inbox, park in _planned, or integrate into active docs. WHEN TO USE: - "Here's a document — integrate it into my context" - "Save this for later" / "Dump these meeting notes" - "I have an idea — park it somewhere" - "Read this article — not sure if it's relevant" - "Our competitor just announced X — update positioning" - Any time new information needs to enter the context architecture - Raw material should become a wiki explainer rather than a canonical data point
First-run onboarding skill. Gets a new user from zero to a working context system. Claude asks one direct question upfront to pick the right scaffolding, then works from what the user shares — no long assessment ceremony beyond that. Creates a checklist that tracks setup progress across sessions and self-removes when done. WHEN TO USE: - "I want to set up my business context" - "Here's my website / pitch deck / LinkedIn" - "Scan my repo and show me what context exists" - "What do I do next?" (after install) - "Help me get started" - First time adopting CLEAR Context OS in any project
Cross-zone retrieval and task-driven bundle resolution for BCOS. /context search ranks docs across every zone in one shared schema; /context bundle (P5) returns a curated, freshness-flagged, source-of-truth-aware context bundle for a declared task. Mechanical-first; LLM is opt-in via --semantic / --resolve-conflicts / --verify-coverage flags only. Invoke with /context.
| name | learning |
| description | Manage the BCOS self-learning ladder: list learned rules, forget a rule, regenerate learned-rules.json from resolutions.jsonl. Mechanical-only — reads/writes derived artifacts; the actual learning happens via record_resolution.py at click time. Invoke with /learning. |
| trigger | /learning |
| version | 1.0.0 |
| last_updated | 2026-05-05 |
| authority-docs | ["docs/_planned/autonomy-ux-self-learning/implementation-plan.md","docs/_bcos-framework/architecture/typed-events.md",".claude/skills/schedule-dispatcher/references/headless-actions.md"] |
Surface for inspecting and editing the rules BCOS has learned from your resolution clicks.
Every dashboard card click and every headless action writes a row to .claude/quality/ecosystem/resolutions.jsonl via record_resolution.py. The promote_resolutions.py script reads that log, groups by (finding_type, action_taken), and promotes pairs that meet tier thresholds into .claude/quality/ecosystem/learned-rules.json. The cockpit reads learned-rules.json to add a "✨ suggested" badge on cards where you've consistently picked the same action.
Tier thresholds:
| Tier | Required signal | Behavior |
|---|---|---|
| preselect (P5, current) | N≥3 + consistency=1.0 | Pre-fills the radio button on the matching card. Doesn't apply anything. |
| auto-apply (P7, future) | N≥5 + consistency≥0.9 + 14d span | Applies after a 10-second undo countdown. |
| silent (P8, gated) | N≥10 + consistency≥0.95 + 6w auditor-clean + opt-in | Fires without a card. Per-rule user click required to promote into this tier. |
learning-blocklist.json is the user's veto: rules in there are excluded from learned-rules.json regardless of evidence.
/learning listShow all currently-learned rules grouped by tier with their evidence counts.
Reads: .claude/quality/ecosystem/learned-rules.json
Side effects: none
Output: table of {rule_id, tier, n, consistency, last_observed, calendar_span_days}.
If learned-rules.json is missing or empty, say so and suggest /learning regenerate.
/learning forget <rule-id>Add a rule to learning-blocklist.json and regenerate learned-rules.json. The rule disappears from the cockpit's "✨ suggested" badges immediately and won't be relearned even if more supporting evidence accumulates.
Required arg: <rule-id> in <finding_type>::<action_taken> form (e.g. inbox-aged::inbox-aged-archive).
Mechanism: call promote_resolutions.add_to_blocklist(rule_id) then promote_resolutions.regenerate(). Both are idempotent.
If the user supplies a partial ID (e.g. just inbox-aged), list matching rules and ask which they meant via AskUserQuestion.
/learning unblock <rule-id>Reverse of forget. Removes the rule from the blocklist; the next regenerate call will repromote it if its evidence still meets a tier threshold.
/learning regenerateRecompute learned-rules.json from resolutions.jsonl + learning-blocklist.json. Byte-stable: re-running with no event changes produces identical bytes.
Mechanism: python .claude/scripts/promote_resolutions.py (the module's __main__ calls regenerate()).
Useful after a manual edit to learning-blocklist.json, or to force a refresh if the cockpit seems to have stale badges.
/learning evidence <rule-id>Show the resolution events backing a learned rule: timestamps, triggers, targets, NL commands. Helps you decide whether the rule was learned for the right reason.
Mechanism: filter resolutions.jsonl rows where (finding_type, action_taken) matches the rule_id parts. Sort by ts. Show last 10 by default; --all for the full log.
.claude/scripts/promote_resolutions.py — compute_rules(), regenerate(), add_to_blocklist(), remove_from_blocklist(), is_blocked(), is_suggested().claude/scripts/record_resolution.py — ResolutionEvent, record(), load_all_rows().claude/quality/ecosystem/resolutions.jsonl — append-only event log.claude/quality/ecosystem/learned-rules.json — derived artifact (regen byte-stable).claude/quality/ecosystem/learning-blocklist.json — user veto listIf the user says any of:
Run the matching subcommand directly. For ambiguous phrasings, use AskUserQuestion with list / forget / regenerate as options.
resolutions.jsonl directly — it's append-only authoritative.learned-rules.json — it's regeneratable. If the user wants to change it, they edit learning-blocklist.json (or future per-rule overrides) and regen.