메뉴
Fact-dense rewrite of agent memory prose (tier-0 notes, digests) without losing technical substance. Companion to /memory-distill. Zero external deps.
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
Phase-gated bug investigation → root-cause → fix → verify cycle. Enforces root-cause-before-edit, failing-test-first, holistic callsite scan, reopen/redesign gates (SPEC-029), multi-surface done matrix, and self-calibration checklist before any "done" claim. Subcommands: /debug <desc> (full), /debug patch <desc> (fast path), /debug arch <desc> (design-first → /kickoff handoff).
Brutally honest review of staged/modified files — no sugar-coating. Thin wrapper over the council engine with preset `diff-mode`: 5 specialist investigators (logic, security, compliance, quality, simplification) run in parallel, filtered at confidence 80. Blocks commit on critical or compliance findings. Optional path argument saves the review to a file.
Adversarial council tribunal engine — reality-checks claims with material evidence. Shared engine for /council and /review-and-commit (diff-mode preset). Blind investigators, evidence-or-silence rule, dual output shapes (verdict[] and finding[]), atomic verdict index at .claude/council/index.json, feedback-memory learning loop. Judge is a dedicated agent with an empty tool allowlist. See specs/core/SPEC-013-adversarial-council-tribunal.md.
Optional host SAST (Semgrep / CodeQL) feed for security review. Fail-open when tools are missing. Agent-internal + review-and-commit / council security flavor. Zero required deps.
Socratic design refinement — structured questioning that forces requirement clarification before any planning or implementation. Use before /kickoff for complex features, or standalone for early-stage ideation. Optional --grill for one-question-at-a-time interviews with recommended answers.
Behavior-preserving polish on recently modified code — clarity only, no logic or API changes. Used by /orchestrate after Tech Lead approve, before QA. Zero external deps (in-plugin; not Anthropic's marketplace plugin).
| name | memory-compress |
| description | Fact-dense rewrite of agent memory prose (tier-0 notes, digests) without losing technical substance. Companion to /memory-distill. Zero external deps. |
Optional brevity pass for memory content — not a full distillation (no tier promotion). Goal: same facts, fewer tokens, so session loads stay lean.
| Caller | When |
|---|---|
/memory-distill | Optional pre-step on raw tier-0 rows before digest LLM (if user asks, or MEMORY_COMPRESS=1) |
| Manual | User asks to compress cortex/lessons/memory prose |
Skip when MEMORY_COMPRESS=0 or content is already bullet-dense.
- Fact one (path/to/file: detail)
- Fact two — constraint
- DO NOT: <rule>
Target: roughly 40–60% fewer tokens on verbose narrative; already-tight bullets may stay near 100%.
# Example: dump one agent's tier-0 for compress, then UPDATE content after rewrite
_gc=$(git rev-parse --git-common-dir 2>/dev/null) \
&& MROOT=$(cd "$(dirname "$_gc")" && pwd) \
|| MROOT=$(pwd)
MEMDB="$MROOT/.claude/memory/memory.db"
# Caller selects rows, rewrites with this skill's rules, writes back via
# memory-store UPDATE protocol — never delete history without distill.
When rewriting .md fallback files (cortex.md / memory.md / lessons.md),
preserve headers; compress body sections only.