원클릭으로
doc-garden
Scan for stale memory topics, near-dead instincts, and low-salience K-LEAN entries. Proposes cleanup actions interactively.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
메뉴
Scan for stale memory topics, near-dead instincts, and low-salience K-LEAN entries. Proposes cleanup actions interactively.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Quality gate for instinct persistence. Validates new instincts through a 4-stage pipeline before allowing them to be written to disk. Invoke before writing any new instinct file.
Use when the user needs thorough research on a complex topic, says 'deep dive', 'research this thoroughly', or 'sensor sweep'. Parallel agent orchestrator for comprehensive multi-angle investigation.
Autonomous experiment loop: iteratively improves a single file against a measurable metric. Uses git-as-state-management in a disposable worktree. Invoke when user wants to optimize, tune, or systematically improve a file through repeated iterations.
Use when you need multi-faction review combining security, strategy, cost, and quality perspectives. Trigger: fleet review, multi-faction assessment, comprehensive quality gate, or when escalation requires L5+.
Two-loop research orchestrator: chains deep-research (outer, hypothesis generation) with experiment-loop (inner, empirical testing). Use when research questions have a testable dimension with a measurable metric.
Rigorous, citation-backed scientific research with evidence grading and claim-evidence tracing. Use when research needs academic-grade citations, source verification, or systematic review methodology. Trigger: /scientific-research, 'research with citations', 'find papers on', 'systematic review', 'literature review'.
| name | doc-garden |
| description | Scan for stale memory topics, near-dead instincts, and low-salience K-LEAN entries. Proposes cleanup actions interactively. |
Scan all three knowledge layers for stale, near-dead, or low-quality entries. Present a unified findings table, then apply only user-approved actions.
Use Glob to list all .md files in the project memory dir:
~/.claude/projects/*/memory/*.mdFor each file, use Read to check:
Flag as STALE if last-referenced date in content is 30+ days ago. Flag as OUTDATED if content references known deprecated systems.
Use Glob to list all .md files in ~/.claude/instincts/.
For each instinct file, use Read to extract:
Confidence: field valueLast validated: field valueThen read ~/.claude/cache/instinct-effective-scores.json with Read (if it exists). Match by instinct name to find effective_score.
Flag as NEAR-DEAD if effective_score < 0.35 (or base Confidence < 0.35 if no effective score found). Flag as UNVALIDATED if Last validated date is 60+ days before today (2026-04-02).
Read .knowledge-db/entries.jsonl (path relative to project root, or check ~/.claude/ if not found locally).
For each entry (one JSON object per line):
salience_score < 0.3retrieval_count == 0Cross-check with ~/.claude/cache/auto-recall-hits.jsonl using Grep (pattern: entry id or key phrase) to verify retrieval status. If hits exist, do not flag NEVER-RETRIEVED.
Limit scan to entries created 30+ days ago (check created_at field) to avoid flagging new entries.
If no issues found across all three layers, output:
Garden is clean. No stale items detected.
Otherwise, present a numbered findings table:
| # | Source | Item | Signal | Issue | Proposed Action |
|----|----------|---------------------------------|-----------------|----------------|-----------------|
| 1 | memory | wordpress.md | 45d stale | STALE | UPDATE |
| 2 | instinct | twingate-client-restart.md | eff: 0.31 | NEAR-DEAD | ARCHIVE |
| 3 | klean | "bcrypt shell variable" (id:42) | salience: 0.18 | LOW-SALIENCE | ARCHIVE |
| 4 | instinct | android-firewall.md | last val: 90d | UNVALIDATED | KEEP |
Default proposed actions:
For each finding, ask the user:
[1] wordpress.md (STALE/UPDATE) — confirm action?
U = UPDATE (read + suggest new content)
A = ARCHIVE (move/mark superseded)
K = KEEP (update timestamp, mark validated)
S = SKIP (no action)
Accept bulk input if user provides a string like UAKSU (one letter per row, in order).
Never take any action before user input. Never delete files; ARCHIVE means:
Superseded by: [reason or date] field and move to ~/.claude/instincts/archived/"superseded": true and "superseded_at": "<today>" on the entry object, rewrite the lineUPDATE: Read the file in full, propose revised content inline, ask for confirmation before writing.
ARCHIVE (instinct): Read the file, add Superseded by: doc-garden review <today> below the Confidence line. Output the move command for the user to run (since subagents cannot execute Bash):
mv ~/.claude/instincts/<name>.md ~/.claude/instincts/archived/<name>.md
ARCHIVE (K-LEAN): Read the relevant line from entries.jsonl, output the updated JSON line with "superseded": true. Ask user to confirm before writing.
KEEP: For instincts, read the file and update Last validated: to today's date (2026-04-02 or current date). Write the change.