メニュー
Saves session context to project memory. Run at end of work or when switching projects. Use when: (1) Ending a work session, (2) Switching projects, (3) Before a long break. Triggers: "/session-save", "save session", "remember context".
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
| name | session-save |
| description | Saves session context to project memory. Run at end of work or when switching projects. Use when: (1) Ending a work session, (2) Switching projects, (3) Before a long break. Triggers: "/session-save", "save session", "remember context". |
Saves key information from the current session into long-term memory so the next session does not start from zero.
Determine which project the work was about:
Load keyword mappings from config:
Read ~/.claude/memory/memory-config.json
Match the session context against project keywords. Use fallback_project from config if no match.
Analyze what happened during the session:
# If there is a git repository:
git log --oneline -5 # recent commits
git diff --stat # uncommitted changes
If git is unavailable, rely on conversation history: which files were created, modified, which tasks were solved.
Extract from the conversation:
Read ~/.claude/memory/projects/{project-name}.md
If the file does not exist, create a new dossier using the template from Step 4.
If it exists, update it. Do not overwrite completely. Preserve history.
Project dossier format:
# {Project Name}
## Status
{One sentence: active/paused/completed}
## Description
{2-3 sentences: what it is, what it does, who uses it}
## Architecture
{Key components, stack, deployment}
## Current State
- Last session: {date}
- Done: {list}
- Uncommitted: {yes/no, what}
## Unresolved Problems
- {problem 1}
- {problem 2}
## Decisions Made
- [{date}] {decision and rationale}
## Next Steps
1. {task}
2. {task}
## Session History
- [{date}] {brief description of work}
Update rules:
In addition to the dossier, add typed observations to the observations file:
Read ~/.claude/memory/projects/{project-name}.observations.md
For each significant session event, add an entry:
Types: decision, bugfix, feature, discovery, problem.
If the file does not exist, create it from template:
# Observations - {project-name}
## Index
| # | Date | Type | Summary | Files |
|---|------|------|---------|-------|
## Details
After updating the dossier and observations, update MEMORY.md:
Read ~/.claude/memory/MEMORY.md
Update:
problem without the [R] marker in Summary- {project} #{number}: {summary}Edit ~/.claude/memory/projects/{project-name}.md
Edit ~/.claude/memory/projects/{project-name}.observations.md
Edit ~/.claude/memory/MEMORY.md
Show the user a brief summary:
Session saved to {project-name} memory.
Recorded:
- Done: {N items}
- Decisions: {N decisions}
- Problems: {N unresolved}
- Observations: {N new observations}
- Next steps: {N tasks}
Next session will load the dossier and continue from here.
Dynamic expert team assembly for planning. Selects 3-5 real-world expert personas tailored to the task, runs parallel analysis, facilitates a debate round on key contradictions, then synthesizes into a plan with explicit tradeoffs. Use when: (1) planning a new feature or project, (2) user says "/dream-team", "assemble a dream team", "pick experts for X", (3) need diverse perspectives before committing to an approach.
Use when you need to validate the health of your Claude Code skills library. Checks for broken file references, missing dependencies, frontmatter issues, duplicate triggers, and dependency drift. Runs structural and semantic audits across all installed skills. Invoke via '/checkup', '/review-skills', or 'check my skills'.
Virtual board of directors. Five expert lenses evaluate a project in parallel (Murati, Sutskever, Cherny, Karpathy, Ive), each through their unique lens. Use when: (1) evaluating a project idea or architecture, (2) user says "/directors", "call the board", "what would directors say about X", (3) need multi-perspective review of a decision or strategy.
Accumulates knowledge from solved problems. Activates automatically after web search or MCP tools are used to solve technical problems. Also activates during debugging to check the knowledge base first. Invoke explicitly via "/til" to record a lesson manually. Triggers: "/til", "record a lesson", "what lessons do we have".
Security review with two modes: (1) code review — 5 parallel agents scan for OWASP Top 10, secrets leaks, auth issues, dependency risks, logic flaws; (2) enterprise assessment — 4 agents evaluate product through CISO lens (data handling, access control, compliance, resilience). Threat model calibration adjusts severity to deployment context. Use when: (1) Before deploy — security audit, (2) Enterprise sales — readiness assessment, (3) New project — baseline security posture, (4) After major changes — regression check. Triggers: "/security review", "/security assess", "/security", "security check".
Agent orchestrator. Analyzes tasks and launches 2-4 specialized agents in parallel, then synthesizes results. 12 agents built from analysis of real usage patterns. Use when: (1) user says "/orchestrate", "run orchestrator", (2) task benefits from parallel analysis by multiple specialists, (3) complex task needs research + implementation or analysis + verification simultaneously.