メニュー
Records atomic learned behaviors with confidence scores. Project-scoped instincts are isolated per project and auto-promote to global scope at confidence threshold 0.8. Stores instincts in .claude/context/memory/instincts.jsonl
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。
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/oimiragieo/agent-studio --skill instinct-learning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。
Agent frontmatter enhancements — disallowedTools, mcpServers scoping, fork_eligible field
Context management — pre-compact persistence, threshold alignment, microcompact/circuit breaker detection
Hook enhancements — updatedInput for bash safety prefixes, suppressOutput for verbose blocks, denial-based routing feedback
SKILL.md を表示中
SOC 職業分類に基づく
| name | instinct-learning |
| description | Records atomic learned behaviors with confidence scores. Project-scoped instincts are isolated per project and auto-promote to global scope at confidence threshold 0.8. Stores instincts in .claude/context/memory/instincts.jsonl |
| version | 1.0.0 |
| model | sonnet |
| invoked_by | both |
| user_invocable | true |
| tools | ["Read","Write","Edit","Bash","Grep","Glob"] |
| agents | ["developer","qa","architect","planner","reflection-agent"] |
| category | Memory |
| tags | ["memory","learning","instinct","confidence","project-scoped","auto-promote"] |
| best_practices | ["Record instincts immediately after observing a reliable pattern","Keep instinct text atomic — one behavior per record","Set confidence honestly; overconfidence poisons the global pool","Review promoted instincts for cross-project validity before relying on them"] |
| error_handling | strict |
| source | builtin |
| trust_score | 100 |
| provenance_sha | 6bed08492dc2f1d0 |
Instinct Learning records atomic learned behaviors as structured instinct records. Each instinct has a confidence score (0.3–0.9), a project scope, and a promotion pathway to global scope when confidence reaches 0.8.
This prevents knowledge from evaporating between sessions and stops cross-project contamination by isolating low-confidence instincts in the project that generated them.
Invoke when:
Skill({ skill: 'instinct-learning' });
NO INSTINCT WITHOUT AN OBSERVED BASIS.
Every instinct record MUST include `source_context` describing the observation
that generated it. Invented instincts are worse than no instincts.
| Score | Meaning | Promotion Eligible |
|---|---|---|
| 0.3 | Seen once, uncertain | No |
| 0.4 | Seen twice, possibly coincidental | No |
| 0.5 | Consistent in this project | No |
| 0.6 | Reliable in this project | No |
| 0.7 | Highly reliable, limited cross-test | No |
| 0.8 | Auto-promotes to global scope | YES |
| 0.9 | Canonical — applies across projects | YES (immediate) |
Each record in .claude/context/memory/instincts.jsonl follows this structure:
{
"id": "inst-<uuid-short>",
"timestamp": "2026-03-23T10:00:00Z"
When scope is global, project is null and promoted_at is set to the ISO timestamp of promotion.
Identify the atomic behavior to record:
project. Auto-promotes when confidence ≥ 0.8.Command:
grep -i "<keyword from instinct text>" C:/dev/projects/agent-studio/.claude/context/memory/instincts.jsonl 2>/dev/null | head -5
Expected output: Existing records matching the pattern, or empty output if new.
Verify: If a duplicate exists, update its confidence instead of creating a new record (see Step 4).
Command:
node C:/dev/projects/agent-studio/.claude/skills/instinct-learning/scripts/main.cjs \
--action record \
--text "{{instinct_text}}" \
--confidence {{confidence_score}} \
--tags "{{comma_separated_tags}}" \
--source "{{source_context}}"
Expected output: {"ok":true,"id":"inst-<id>","scope":"project","promoted":false}
Verify: Exit code 0 and JSON with ok: true.
If the same pattern is observed again, increase confidence:
Command:
node C:/dev/projects/agent-studio/.claude/skills/instinct-learning/scripts/main.cjs \
--action update \
--id "{{instinct_id}}" \
--confidence {{new_confidence_score}}
Expected output: {"ok":true,"id":"{{instinct_id}}","confidence":{{new_score}},"promoted":{{true_or_false}}}
Verify: If promoted: true is returned, the instinct is now in global scope.
When confidence reaches 0.8, auto-promotion fires automatically during the update step. Verify with:
Command:
node C:/dev/projects/agent-studio/.claude/skills/instinct-learning/scripts/main.cjs \
--action list \
--scope global \
--limit 10
Expected output: JSON array of globally-promoted instincts, newest first.
Verify: The newly promoted instinct appears with scope: "global" and promoted_at set.
Before starting a task, load relevant instincts to guide behavior:
Command:
node C:/dev/projects/agent-studio/.claude/skills/instinct-learning/scripts/main.cjs \
--action query \
--tags "{{relevant_tags}}" \
--min-confidence 0.6
Expected output: JSON array of relevant instincts ordered by confidence descending.
Verify: Exit code 0 and array (may be empty if no matches).
source_context — if you cannot remember the basis, set confidence to 0.3update action and let auto-promotion fireInput validated against schemas/input.schema.json before execution.
Output contract defined in schemas/output.schema.json.
Pre-execution hook (hooks/pre-execute.cjs) validates:
confidence is in range [0.3, 0.9]text is non-empty and under 200 charactersaction is a known enum valuePost-execution hook (hooks/post-execute.cjs) emits observability event to .claude/context/runtime/tool-events.jsonl.
developer — records patterns found during implementationqa — records patterns from testing and validationarchitect — records structural patternsplanner — records planning heuristicsreflection-agent — records patterns surfaced during reflectionBefore starting: Read .claude/context/memory/learnings.md and check .claude/context/memory/instincts.jsonl for relevant existing instincts.
After completing: Append summary to .claude/context/memory/learnings.md:
**instinct-learning** — [date] Recorded/updated N instincts. Promoted M to global scope.
ASSUME INTERRUPTION: Your context may reset. If it's not in memory, it didn't happen.
memory-search — semantic search over global agent memorycontext-compressor — compress context before long taskssession-handoff — persist instincts across session boundaries