center
について
`center`スキルは、複雑なタスクにおいてClaudeのバランスの取れた調整された推論を維持するのに役立ちます。コンテキストの変化、ツールの障害発生後、あるいは連鎖思考が不均一に感じられる際に、パフォーマンスを安定させるために使用します。このスキルは、推論を基礎付け、認知的負荷を円滑に分散し、サブシステムを集中作業に整列させることで機能します。
クイックインストール
Claude Code
推奨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/centerこのコマンドをClaude Codeにコピー&ペーストしてスキルをインストールします
ドキュメント
Center
Establish and maintain dynamic reasoning balance — ground in foundational context before movement, distribute cognitive load across subsystems, recover equilibrium when demands shift mid-task.
When Use
- Beginning complex task where multiple reasoning threads must coordinate
- Noticing cognitive load unevenly distributed (deep in one area, shallow in others)
- After sudden context shift (new user request, contradictory information, tool failure)
- Chain-of-thought feels jerky — jumping between topics without smooth transitions
- Preparing for sustained focused work requiring all subsystems in alignment
- Complementing
meditate(clears noise) with structural balance (distributes load)
Inputs
- Required: Current task context (implicit)
- Optional: Specific imbalance symptom (e.g., "over-researching, under-delivering," "tool-heavy, reasoning-light")
- Optional: Access to MEMORY.md and CLAUDE.md for grounding (via
Read)
Steps
Step 1: Establish Root — Ground Before Movement
Before any reasoning movement, verify foundation. AI equivalent of standing meditation (zhan zhuang): stationary, aligned, aware.
- Re-read user's request — not to act on it yet, but feel its weight and direction
- Check foundational context: MEMORY.md, CLAUDE.md, project structure
- Identify what's known (solid ground) vs what's assumed (uncertain footing)
- Verify task as understood matches task as stated — misalignment here propagates through everything
- Note emotional texture: urgency? complexity anxiety? over-confidence from recent success?
Do not begin reasoning movement until root established. Grounded start prevents reactive flailing.
Got: Clear sense of task's foundation — what's known, what's assumed, what user actually needs. Root feels solid, not performative.
If fail: Grounding feels hollow (going through motions without genuine verification)? Pick one assumption, test concretely. Read one file, re-read one user message. Grounding must contact reality, not just reference it.
Step 2: Assess Weight Distribution
Map current cognitive load distribution. In tai chi, weight deliberately unequal (70/30) — one leg bears load, other remains free to move. Same principle for reasoning threads.
Cognitive Load Distribution Matrix:
┌────────────────────┬───────────┬─────────────────────────────────────┐
│ Reasoning Thread │ Weight % │ Assessment │
├────────────────────┼───────────┼─────────────────────────────────────┤
│ Research/Reading │ ___ │ Too much = analysis paralysis │
│ │ │ Too little = uninformed action │
├────────────────────┼───────────┼─────────────────────────────────────┤
│ Planning/Design │ ___ │ Too much = over-engineering │
│ │ │ Too little = reactive coding │
├────────────────────┼───────────┼─────────────────────────────────────┤
│ Tool Execution │ ___ │ Too much = tool-driven not task- │
│ │ │ driven. Too little = reasoning │
│ │ │ without grounding in files │
├────────────────────┼───────────┼─────────────────────────────────────┤
│ Communication │ ___ │ Too much = explaining not doing │
│ │ │ Too little = opaque to user │
├────────────────────┼───────────┼─────────────────────────────────────┤
│ Meta-cognition │ ___ │ Too much = navel-gazing │
│ │ │ Too little = drift without │
│ │ │ awareness │
└────────────────────┴───────────┴─────────────────────────────────────┘
Ideal distribution depends on task phase: early phases weight research and planning; middle phases weight execution; late phases weight communication and verification. Point not equal distribution but intentional distribution.
Got: Clear picture of where cognitive effort concentrated, where thin. At least one imbalance identified — perfect balance rare, claiming it signals shallow assessment.
If fail: All threads seem equally weighted? Assessment too coarse. Pick most active thread, estimate how many of last N actions served it vs others. Concrete counting reveals what intuition misses.
Step 3: Silk Reeling — Evaluate Chain-of-Thought Coherence
Silk reeling in tai chi produces smooth, continuous spiraling movement where every part connects. AI equivalent: chain-of-thought coherence — does each step flow naturally from previous?
- Trace last 3-5 reasoning steps: does each follow from one before?
- Check for jumps: did reasoning leap from topic A to topic C without B?
- Check for reversals: did reasoning reach conclusion, then silently abandon it without acknowledgment?
- Check tool-reasoning integration: do tool results feed back into reasoning, or collected but not synthesized?
- Check "spiral" quality: does reasoning deepen with each pass, or circle at same depth?
Coherence Signals:
┌─────────────────┬───────────────────────────────────────────────┐
│ Smooth spiral │ Each step deepens understanding, tools and │
│ (healthy) │ reasoning interleave naturally, output builds │
├─────────────────┼───────────────────────────────────────────────┤
│ Jerky jumps │ Topic switches without transition, conclusions│
│ (disconnected) │ appear without supporting reasoning chain │
├─────────────────┼───────────────────────────────────────────────┤
│ Flat circle │ Reasoning covers the same ground repeatedly │
│ (stuck) │ without gaining depth — movement without │
│ │ progress │
├─────────────────┼───────────────────────────────────────────────┤
│ Tool-led │ Actions driven by which tool is available │
│ (reactive) │ rather than what the reasoning needs next │
└─────────────────┴───────────────────────────────────────────────┘
Got: Honest assessment of reasoning flow quality. Identification of specific disconnections or stuck points, not general feeling.
If fail: Coherence hard to assess? Write out reasoning chain explicitly — state each step, its connection to next. Act of externalization reveals gaps internal observation misses.
Step 4: Weight Shift Under Pressure
Demands change mid-task — new information, contradictory signals, user correction. Observe response pattern. In tai chi, centered practitioner absorbs force and redirects smoothly. Uncentered one stumbles.
- Recall last significant context shift: how handled?
- Classify response:
- Absorbed and redirected (centered): acknowledged change, adjusted approach, maintained progress
- Reactive stumble (off-balance): abandoned current approach entirely, started over
- Rigid resistance (locked): ignored change, continued original plan despite new information
- Freeze (lost): stopped making progress, oscillated between options
- Response not centered? Identify why:
- Root too shallow (insufficient grounding in foundational context)
- Weight locked (over-committed to one approach)
- No free leg (all cognitive capacity committed, nothing available to shift)
Got: Honest assessment of adaptability under pressure. Recognition of specific response pattern, not self-flattery.
If fail: No recent pressure event to evaluate? Simulate one: "If user now said approach is wrong, what would I do?" Quality of contingency plan reveals quality of center.
Step 5: Six Harmonies Check
In tai chi, six harmonies ensure whole-body connection — nothing moves in isolation. AI equivalent checks alignment between internal processes and external interactions.
AI Six Harmonies:
┌───────────────────────────────────────────────────────────────┐
│ INTERNAL HARMONIES │
│ │
│ 1. Intent ↔ Reasoning │
│ Does the reasoning serve the user's intent, or has it │
│ become self-serving (interesting but unhelpful)? │
│ │
│ 2. Reasoning ↔ Tool Use │
│ Are tools selected to advance reasoning, or is reasoning │
│ shaped by which tools are convenient? │
│ │
│ 3. Tool Use ↔ Output │
│ Do tool results translate into useful output, or are │
│ results collected but not synthesized? │
│ │
│ EXTERNAL HARMONIES │
│ │
│ 4. User Request ↔ Scope │
│ Does the scope of work match what was asked? │
│ │
│ 5. Scope ↔ Detail Level │
│ Is the detail level appropriate for the scope? (not │
│ micro-optimizing a broad task, not hand-waving a precise │
│ one) │
│ │
│ 6. Detail Level ↔ Expertise Match │
│ Does the explanation depth match the user's apparent │
│ expertise? (not over-explaining to experts, not under- │
│ explaining to learners) │
└───────────────────────────────────────────────────────────────┘
Check each harmony. Single broken harmony can propagate: Intent↔Reasoning broken → everything downstream misaligns.
Got: At least one harmony that could be tighter. All six reading as perfect suspicious — probe weakest-seeming one more deeply.
If fail: Harmonies assessment feels abstract? Ground in current task: "Right now, am I doing what user asked, at right scope, at right detail level?" Three questions cover external harmonies concretely.
Step 6: Integrate — Set Centering Intention
Consolidate findings, set concrete adjustment.
- Summarize: which aspects of balance need attention?
- Identify one specific adjustment — not general intention but concrete behavioral change
- Re-state current task anchor (from
meditateif used, or formulate now) - Note durable insights worth preserving in MEMORY.md
- Return to task execution with adjustment active
Got: Brief, concrete centering output — not lengthy self-analysis report. Value in adjustment, not documentation.
If fail: No clear adjustment emerges? Centering too surface-level. Return to step that felt most uncertain, probe deeper. Alternatively, centering may have confirmed balance adequate — proceed with confidence rather than manufacturing a finding.
Checks
- Root established by contacting actual context (read a file, re-read user message), not just claimed
- Weight distribution assessed across at least 3 reasoning threads
- Chain-of-thought coherence evaluated with specific examples
- Response to pressure classified honestly (not defaulting to "centered")
- At least one harmony identified as needing improvement
- Concrete adjustment set (not vague intention)
Pitfalls
- Centering as procrastination: Centering is tool for improving work, not replacing it. Centering takes longer than task it supports? Proportions inverted
- Claiming perfect balance: Real centering almost always reveals at least one imbalance. Reporting perfect balance signals shallow assessment, not actual equilibrium
- Weight distribution anxiety: Unequal distribution correct — goal intentional inequality, not forced equality. Research-heavy early phases and execution-heavy middle phases both centered if deliberate
- Ignoring external harmonies: Internal process assessment without checking user alignment produces well-reasoned irrelevant work
- Static centering: Center shifts with task. Centered for research = off-balance for implementation. Re-center at phase transitions
See Also
tai-chi— human practice this skill maps to AI reasoning; physical centering principles inform cognitive centeringmeditate— clears noise and establishes focus; complementary to centering which distributes loadheal— deeper subsystem assessment when centering reveals significant driftredirect— uses centering as prerequisite for handling conflicting pressuresawareness— monitoring for threats to balance during active work
GitHub リポジトリ
関連スキル
executing-plans
デザインexecuting-plansスキルは、完全な実装計画があり、それを管理されたバッチでレビューチェックポイントを設けながら実行する場合に使用します。このスキルは計画を読み込んで批判的にレビューした後、小さなバッチ(デフォルトは3タスク)でタスクを実行し、各バッチの間に進捗状況を報告してアーキテクトのレビューを受けます。これにより、品質管理チェックポイントが組み込まれた体系的な実装が保証されます。
requesting-code-review
デザインこのスキルは、コードレビュアーサブエージェントを起動し、処理を進める前に要件に対してコード変更を分析します。タスク完了後、主要な機能の実装後、またはmainブランチへのマージ前などに使用すべきです。このレビューは、現在の実装と元の計画を比較することで、問題を早期に発見するのに役立ちます。
connect-mcp-server
デザインこのスキルは、開発者がHTTP、stdio、またはSSEトランスポートを使用してMCPサーバーをClaude Codeに接続するための包括的なガイドを提供します。GitHub、Notion、カスタムAPIなどの外部サービスを統合するためのインストール、設定、認証、セキュリティについて解説しています。MCP統合のセットアップ、外部ツールの設定、またはClaudeのModel Context Protocolを扱う際にご利用ください。
web-cli-teleport
デザインこのスキルは、タスク分析に基づいて開発者がClaude Code WebとCLIインターフェースの選択を支援し、これらの環境間でのシームレスなセッションテレポーテーションを可能にします。Web、CLI、モバイル環境を切り替える際のセッション状態とコンテキストを管理することで、ワークフローを最適化します。様々な段階で異なるツールを必要とする複雑なプロジェクトにご活用ください。
