菜单
Use when auditing and evolving the brain vault — pruning outdated or low-value notes, discovering cross-cutting principles, and reviewing skills for structural encoding opportunities; triggers include "meditate", "audit the brain".
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
Use when a plan or spec splits into many independently-mergeable units — a codebase-wide migration, rename, or repeated pattern change — and the user wants them implemented in parallel as separate PRs rather than one sequential pass. The fan-out alternative to afk:implement.
Use when the user says "grill me", asks to stress-test a plan, wants to be interviewed or asked the questions needed before building, offers a vague feature idea, or wants grounded planning before non-trivial implementation
Use when the user asks to run the whole AFK flow, ship a feature, take work from idea or plan to verified evidence, resume an AFK workflow, or make a final ship/no-ship call across planning, implementation, cleanup, and QA
Use when a task needs to read or write the persistent brain/ vault — adding a principle, codebase gotcha, or note, or grounding work in existing memory; triggers include "add to brain", "remember this in the brain", or any direct brain/ edit.
Use when the user asks for help, afk help, what to do next, where am I, which AFK skill to use, AFK workflow orientation, or a broad catalog explanation
Use when the agent is about to implement, edit code, execute a plan, fix a bug, build a feature, or make repo changes, especially when task complexity or orchestration needs are unclear. Do not start editing code, writing tests, or running build or test commands directly — load this skill first to triage and route.
| name | meditate |
| description | Use when auditing and evolving the brain vault — pruning outdated or low-value notes, discovering cross-cutting principles, and reviewing skills for structural encoding opportunities; triggers include "meditate", "audit the brain". |
Audit and evolve the brain vault. The quality bar a note must clear to keep its place: high-signal (Claude would reliably get this wrong without it), high-frequency (comes up in most sessions or most tasks of a type), or high-impact (getting it wrong causes real damage or wasted work). Everything else is noise — a lean, precise brain outperforms a comprehensive but bloated one.
Do not use this skill to capture the current session (use reflect) or to mine
old conversations (use ruminate).
Build snapshots.
sh "${CLAUDE_PLUGIN_ROOT}/skills/meditate/scripts/snapshot.sh" brain/ /tmp/brain-snapshot.md
sh "${CLAUDE_PLUGIN_ROOT}/skills/meditate/scripts/snapshot.sh" "${CLAUDE_PLUGIN_ROOT}/skills/" /tmp/skills-snapshot.md
Files are delimited with === path/to/file.md === headers. Also locate the
auto-memory directory (~/.claude/projects/<project>/memory/).
Auditor (blocking). Spawn a general-purpose subagent — full prompt spec
in references/agents.md. It audits brain notes, CLAUDE.md, and auto-memory
for staleness, redundancy, low-value content, verbosity, and orphans, and
returns a categorized report. Early-exit gate: if it finds fewer than 3
actionable items, skip step 3 and go to step 4.
Reviewer. Spawn one general-purpose subagent (spec in
references/agents.md) combining: synthesis (missing wikilinks, principle
tensions, clarifications), distillation (recurring patterns revealing
unstated principles — must be independent, evidenced by 2+ notes, and
actionable), and skill review (skills vs. brain principles:
contradictions, missed structural enforcement, redundant instructions).
Review reports. Present a consolidated summary (format in
references/agents.md).
Route skill-specific learnings into the relevant SKILL.md or references/ (read the skill first to avoid duplication).
Apply changes directly (the user reviews the diff): update/delete outdated
notes, merge redundant ones, delete low-value ones, condense verbose ones, add
[[wikilinks]], reword tensions, add genuinely independent new principles
(update brain/principles.md), fix CLAUDE.md and stale memories.
Housekeep. Update the relevant index entrypoints for files added or
removed; brain/index.md is rebuilt by the hook.
STOP and ask the user before deleting a note whose value is genuinely ambiguous, or before adding a new principle that could conflict with an existing one. Pruning that is clearly low-value or stale can proceed without asking.
Return a summary in this shape:
## Meditate Summary
- Pruned: [N notes deleted, M condensed, K merged]
- Extracted: [N new principles, one line + evidence count each]
- Skill review: [N findings, M applied]
- Housekeep: [state files cleaned]