메뉴
Extract learning before restarting. Code is a draft; learning is the asset. Use when work is drifting, approach has reversed 3+ times, or scope is expanding while "done" keeps fuzzing.
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
Clarify the outcome you want - a change in user behavior, not a feature shipped. Use at the start of any work to ground the session in strategic intent.
Devil's advocate. Seek contrary evidence before locking in. Use when about to make a significant decision, when confidence is high but stakes are higher, or when people are converging too quickly.
Human-led curation of accumulated metis and guardrails. Surface patterns across sessions, propose what to promote, compact, or dismiss. Use after multiple sessions, before a new phase, or when search results feel noisy.
Do the work. Pre-flight, build, detect drift, salvage if needed. Use when you have a clear aim and are ready to implement.
Map what we're optimizing and what constraints we treat as real. Use before jumping to solutions, when hitting repeated blockers, or when patches keep accumulating.
Define the framing of a problem. Change the statement, change the solution space. Use when starting work, when solutions feel wrong, or when you suspect an X-Y problem.
| name | salvage |
| description | Extract learning before restarting. Code is a draft; learning is the asset. Use when work is drifting, approach has reversed 3+ times, or scope is expanding while "done" keeps fuzzing. |
Extract learning before restarting. Code is cheap; learning is the asset.
Invoke /salvage when work is drifting, approach reversed multiple times, scope keeps expanding, the finish line keeps moving, or starting over feels right.
Do not use when: Work is on track. Salvage is extraction before restart, not routine reflection.
Load references/extraction.md only when a salvage needs deeper learning extraction than the core report.
"This session/approach is being salvaged because [reason]. The original aim was [aim]. What happened was [reality]."
Extract only what is present; do not pad empty categories.
Core categories:
If RNA/OH tools are available, check existing metis/guardrails before writing new ones. Update or strengthen existing entries rather than creating near-duplicates.
Synthesize a restart kit: original aim, why salvaged, key learnings, frame shift, new guardrails, ownership/coordination breakdown, context for restart, and reusable fragments.
Persist metis, not noise. Record only non-obvious constraints, anomalies, trade-offs, and local patterns that would change a later decision. Generic best-practice advice adds context cost without local leverage.
If repeated salvages surface the same learning, run /distill; the corpus has knowledge but it is not being consulted.
## Salvage
### Salvage Report
**Salvaged:** [date/session identifier]
**Reason:** [why this work is being salvaged]
**Original Aim:** [what we were trying to do]
### Learnings
[Numbered list of key insights]
### Frame Shifts
[Old frame → new frame, with evidence]
### New Guardrails
[Explicit constraints with reason and trigger]
### Missing Context
[What would have helped]
### Ownership / Coordination Breakdowns
[Authority boundary, missing owner, or handoff rule]
### Reusable Fragments
[Code or patterns worth keeping]
### Fresh Start Recommendation
[How to approach this next time]
If a session file is in use, read prior phase sections, then replace or append ## Salvage. Later attempts use the salvage report as restart context, not as a mandate to preserve failed code.