一键导入
close
End-of-session ritual. Captures session state, updates auto-memory, writes a session summary, then exits the session cleanly.
用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
菜单
End-of-session ritual. Captures session state, updates auto-memory, writes a session summary, then exits the session cleanly.
用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | close |
| description | End-of-session ritual. Captures session state, updates auto-memory, writes a session summary, then exits the session cleanly. |
Preserve the work. Carry the thread forward.
When invoked, perform ALL of the following steps in order:
Review the full conversation and identify:
Update ~/.claude/projects/{project-key}/memory/MEMORY.md with any new stable knowledge from this session.
Save if:
Don't save:
Use Edit tool to append or update relevant sections. Create topic files if a section grows too large.
Write a summary file to: ~/.claude/projects/{project-key}/sessions/{date}-{slug}.md
Format:
# Session: {date} -- {2-4 word description}
## Work Done
- {bullet list of completed items}
## Current State
- {what's working, what's running, what's deployed}
## Open Threads
- {unfinished work with enough context to resume}
## Key Files Modified
- `{path}` -- {what changed}
## Decisions & Context
- {important choices made, with rationale}
## Next Steps
- {prioritized list of what to do next}
If you have a notification channel configured (WhatsApp, Slack, etc.), send a consolidated session summary:
Session closed
{2-3 sentence summary of work done}
State: {brief current state}
Next: {top 1-2 priorities}
Repo: {https-url} ({branch} @ {short-hash})
Omit the Repo: line if not in a git repo or no remote is configured.
If the session contained a repeatable workflow or novel pattern, ask:
"This session had [pattern]. Want me to distill it into a reusable skill before closing?"
Only offer if there's something genuinely worth distilling. Don't force it.
Rename the current session's .jsonl transcript from its UUID to a human-readable name.
Location: ~/.claude/projects/{project-key}/
Format: {date}-{slug}.jsonl
{date} = YYYY-MM-DD{slug} = 2-4 word kebab-case description matching the session summary slugProcedure:
ls -t ~/.claude/projects/{project-key}/*.jsonl | head -1mv {uuid}.jsonl {date}-{slug}.jsonlNote: This breaks --resume for the old UUID, which is fine since the session is closing. The renamed file remains searchable via Grep for future context lookups.
After all steps complete, tell the user:
Session closed. All state preserved. You can exit with /exit or Ctrl+C.
Then stop responding. Do not continue the conversation. The session is done.
Part of Claude Code Skills -- reusable skills extracted from real AI-assisted development sessions.
Install a lightweight inline-editing CMS into any Vercel-deployed project — content store + EditableText component + cookie-auth admin shell + Vercel KV/Upstash/Blob storage. Auto-detects the framework and installs natively for Next.js (App Router & Pages Router) and Vite + React. For SvelteKit, Astro, Nuxt, or plain static, prints a manual-port guide. Use when the user asks to add a CMS, inline editing, editable text, an admin panel, or "make this site editable."
Audit and improve a Claude Code memory layer end-to-end. Measures CLAUDE.md and MEMORY.md against Anthropic's documented caps (200 lines per CLAUDE.md, 200 lines OR 25 KB for MEMORY.md auto-load), surveys project portfolio on disk vs memory coverage, finds sync conflicts in ~/.claude/, identifies gaps tier-by-tier, and executes improvements under bias-to-action with cap-fitting verification at each step. Use when CLAUDE.md/MEMORY.md "feel bloated", when memory feels invisible in fresh sessions, or as periodic hygiene (~monthly).
Burn project bloat to reveal essence. Surgical, intent-anchored, multi-agent. Articulate a three-layer intent (business / architectural / stylistic), dispatch 4 parallel subagents to find code-rot / structural / dep / doc-test bloat, produce a tiered CALCINATION-PLAN.md, then execute per-item with auto-revert on failed verify. Reversible via git. Use when a project has gotten fat and needs aligning back to its core purpose.
A discipline for working with AI code generation. Replaces algorithm-writing with the meta-skill that Google's new Code Comprehension Interview tests: directing AI codegen, validating its output, and debugging what AI produces. Five principles + five triggers, with the discipline applied to Claude's own output as well as the user's. Use when actively pairing with AI on real code.
Opinionated meta-prompting command suite — 9 composable patterns built on categorical foundations. Use when the user wants to set up a meta-prompting workflow, route prompts by context, frame goals, build subagents, distill session learnings into durable artifacts, or invoke the alchemical 4-stage / Hermetic-correspondence patterns for personal/creative work.
Build native Android APKs without Android Studio using raw SDK tools (javac, d8, aapt, apksigner). Produces tiny APKs (~30KB) that build in under 2 seconds.