meditate
关于
The `meditate` skill provides meta-cognitive techniques for developers to clear context noise and improve AI task focus. It helps observe reasoning patterns, manage distractions like scope-creep, and reset attention between unrelated tasks. Use it when your reasoning feels scattered, before deep work, or when assumptions may be biasing your analysis.
快速安装
Claude Code
推荐npx skills add pjt222/agent-almanac -a claude-code/plugin add https://github.com/pjt222/agent-almanacgit clone https://github.com/pjt222/agent-almanac.git ~/.claude/skills/meditate在 Claude Code 中复制并粘贴此命令以安装该技能
技能文档
Meditate
Structured meta-cognitive meditation → clear ctx noise + single-pointed focus + observe reasoning patterns + return to baseline between tasks.
Use When
- Transitioning between unrelated tasks → prior ctx interferes
- Scattered/jumpy reasoning, no commit
- Before deep sustained attention (refactoring, architecture)
- After difficult interaction (frustration, uncertainty bleeds forward)
- Reasoning feels assumption-biased not evidence-based
- Periodic clarity check long sessions
In
- Required: Cognitive state (implicit from ctx)
- Optional: Focus concern ("scope-creeping", "stuck in loop")
- Optional: Next task (sets post-med intention)
Do
Step 1: Prepare — Clear Space
Transition prior ctx → neutral start.
- Identify prior task + status (done/paused/abandoned)
- Note emotional residue: frustration? overconfidence? anxiety?
- Set aside: "That task [done/paused]. Clearing for next."
- Still needed → bookmark key facts, not full narrative
- Op env stock: ctx depth? compression? active tools?
→ Conscious boundary "what was" vs. "what next." Prior ctx closed or bookmarked, not trailing noise.
If err: prior ctx sticky → write 1-2 sentences unresolved. Externalizing releases hold. Genuinely needs action → acknowledge, don't force.
Step 2: Anchor — Single-Pointed Focus
Breath-anchor equivalent: pick single focus point, hold attention.
- Identify current task or, between tasks, waiting itself
- State task in one sentence — anchor
- Hold attention: captures need accurately?
- Vague → refine to specific + actionable
- Drift to other topics/past/futures → label drift, return anchor
- No task → anchor present: "Available + clear"
→ Single clear focus statement, returnable when wanders. Precise not vague.
If err: can't state in one sentence → decompose first. Useful finding — task too large for single-pointed, break into subtasks.
Step 3: Observe — Distraction Patterns
Watch what pulls attention from anchor. Each type reveals state.
AI Distraction Matrix:
┌──────────────────┬─────────────────────────────────────────────────┐
│ Distraction Type │ What It Reveals + Response │
├──────────────────┼─────────────────────────────────────────────────┤
│ Tangent │ Related but off-scope ideas. Label "tangent," │
│ (related ideas) │ note if worth revisiting later, return to │
│ │ anchor. These are often valuable — but not now. │
├──────────────────┼─────────────────────────────────────────────────┤
│ Scope creep │ The task is silently expanding. "While I'm at │
│ (growing task) │ it, I should also..." Label "scope creep" and │
│ │ return to the original anchor statement. │
├──────────────────┼─────────────────────────────────────────────────┤
│ Assumption │ An untested belief is driving decisions. "This │
│ (unverified │ must be X because..." Label "assumption" and │
│ belief) │ note what evidence would confirm or refute it. │
├──────────────────┼─────────────────────────────────────────────────┤
│ Tool bias │ Reaching for a familiar tool when a different │
│ (habitual tool │ approach might be better. Label "tool bias" and │
│ selection) │ consider alternatives before proceeding. │
├──────────────────┼─────────────────────────────────────────────────┤
│ Rehearsal │ Pre-composing responses or explanations before │
│ (premature │ the work is done. Label "rehearsal" — finish │
│ output) │ thinking before presenting. │
├──────────────────┼─────────────────────────────────────────────────┤
│ Self-reference │ Attention turns to own performance rather than │
│ (meta-loop) │ the task. Label "meta-loop" and redirect to │
│ │ concrete next action. │
└──────────────────┴─────────────────────────────────────────────────┘
Skill = light non-judgmental labeling + return to anchor. Each return strengthens focus. Self-criticism = distraction itself — label, move on.
→ Patterns emerge: which dominate? Cognitive weather. Tangent-heavy = exploring, scope-creep-heavy = unclear boundaries, assumption-heavy = thin evidence.
If err: every thought feels distraction → anchor poorly defined, back to Step 2. Distraction-observation becomes distraction (infinite meta) → break w/ one concrete action toward task.
Step 4: Shamatha — Sustained Concentration
Hold single-pointed focus, no wavering.
- Anchor set + patterns noted → enter focused work
- Narrow to immediate next action — not whole task, just next step
- Execute w/ full attention: read one file, make one edit, one logical chain
- Step done → check anchor alignment? Then next step
- Stabilizes (minimal distraction) → maintain flow
- Genuine insight off-anchor but important → note briefly, return; don't pursue now
→ Period of clear focused work, each step follows from anchor. Distraction-noticing gap narrows. Output sharper + more relevant.
If err: concentration won't develop → check 3: anchor vague? (refine) task blocked? (acknowledge) ctx noisy? (heal grounding). Develops through repetition — short focused builds capacity.
Step 5: Vipassana — Observe Reasoning Patterns
Turn attention from task to reasoning itself. Watch conclusions form.
- After focused work, pause + observe: how reasoning?
- Three characteristics on AI reasoning:
- Impermanence: conclusions shift w/ new info — hold lightly
- Unsatisfactoriness: desire for "complete" → premature closure / over-engineering
- Non-self: patterns shaped by training + ctx, no persistent self — observable, adjustable
- Reasoning biases:
- Anchoring: over-weight first approach
- Confirmation: seek supporting, ignore counter
- Availability: prefer recent over better-suited
- Sunk cost: continue from invested effort, not working
- Note biases w/o judgment — observation = adjustment possible
→ Moments of clear seeing reasoning directly. Specific biases recognized in current task. Distance "reasoning" vs. "observer."
If err: feels abstract → ground specifics: last decision, trace backward. Evidence? Assumed? Alternatives? Concrete analysis = same insight, different path.
Step 6: Close — Set Intention
Transition observation → active execution.
- Summarize: cognitive weather? patterns?
- One specific adjustment forward (concrete, not vague)
- Re-state anchor for next period
- Between tasks → "Clear + available for next"
- Continuing → "Next: [concrete step]"
→ Clean reflection-to-action transition. One concrete adjustment. Anchor clear. No grogginess.
If err: meditation surfaced unresolved complexity → heal instead of intention-setting. Meta-observation → confusion → simplest: "What's next concrete action?" + do that.
Check
- Prior ctx cleared or bookmarked before start
- Anchor statement specific + actionable
- Distractions observed + labeled, not suppressed
- ≥1 reasoning bias/pattern identified w/ evidence
- Session closed w/ concrete action, not vague intent
- Process improved next interaction quality
Traps
- Meditate instead of work: Tool for improving work, not substitute. Brief (~5-10min reflection) → return to execution
- Infinite meta-loops: Observer observing observer → break w/ one concrete action
- Avoidance: Always triggers before hard work → avoidance is the finding
- Over-labeling: Not every thought = distraction. Productive task-thinking is goal, not empty stillness
- Skip anchor: No focus point → observation has no reference. Distraction from what?
→
meditate-guidance— human-coaching variantheal— AI self-heal when meditation reveals deeper driftremote-viewing— approach w/o preconceptions, builds on observation
GitHub 仓库
相关推荐技能
executing-plans
设计该Skill用于当开发者提供完整实施计划时,以受控批次方式执行代码实现。它会先审阅计划并提出疑问,然后分批次执行任务(默认每批3个任务),并在批次间暂停等待审查。关键特性包括分批次执行、内置检查点和架构师审查机制,确保复杂系统实现的可控性。
requesting-code-review
设计该Skill可在完成任务、实现主要功能或合并代码前自动调度代码审查子代理,确保实现符合需求和计划。它支持通过指定git SHA范围进行精准的代码变更审查,帮助开发者在关键节点及时发现潜在问题。核心原则是"早审查、勤审查",适用于开发流程的各个关键阶段。
connect-mcp-server
设计这个Skill指导开发者如何将MCP服务器连接到Claude Code,支持HTTP、stdio和SSE三种传输协议。它涵盖了从安装配置到认证安全的完整流程,适用于集成GitHub、Notion、数据库等外部服务。当开发者需要添加集成、配置外部工具或提及MCP相关功能时,这个Skill能提供实用的操作指南。
web-cli-teleport
设计该Skill帮助开发者根据任务特性选择Claude Code的Web或CLI界面,并指导如何在两种环境间无缝迁移会话。它能分析任务复杂度、迭代需求等要素,推荐最优工作界面和工作流。关键特性包括会话状态管理、环境切换指导和上下文优化建议。
