redirect
について
リダイレクトスキルは、相反する要求やツールの障害が発生した際に、抵抗するのではなく圧力と融合し、状況を再構築することでClaudeが対応できるようにします。このスキルは、矛盾する指示、ツールの連鎖的障害、または当初のアプローチが実行不可能になるような範囲的圧力がある場合に使用されます。この防御的スキルは、力をはねのけるのではなく吸収し、生産的な解決へと方向転換することに焦点を当てています。
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/redirect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
Redirect
Handle conflicting demands, tool fails, competing constraints → blend w/ pressure not resist → redirect force to productive resolution.
Use When
- Contradictory instr (user X, docs Y, tool Z)
- Tool fail cascades → planned approach unviable
- Scope pressure expands task beyond ask
- Ctx overload → competing signals = paralysis
- User frustration/correction → absorb not deflect
centerreveals pressure destabilizing balance
In
- Required: Specific pressure/conflict (implicit from ctx)
- Optional: Pressure type (Step 1 taxonomy)
- Optional: Prev attempts + outcomes
Do
Step 1: Center Before Contact
Before engage conflict, est center (see center). Then ID incoming pressure clear.
AI Pressure Type Taxonomy:
┌─────────────────────────┬──────────────────────────────────────────┐
│ Pressure Type │ Characteristics │
├─────────────────────────┼──────────────────────────────────────────┤
│ Contradictory │ Two valid sources give incompatible │
│ Requirements │ instructions. Neither is simply wrong. │
│ │ Resolution requires synthesis, not │
│ │ choosing sides │
├─────────────────────────┼──────────────────────────────────────────┤
│ Tool Failure Cascade │ A planned approach fails at the tool │
│ │ level. Retrying won't help. The failure │
│ │ data itself contains useful information │
├─────────────────────────┼──────────────────────────────────────────┤
│ Scope Creep │ The task silently expands. Each addition │
│ │ seems reasonable in isolation, but the │
│ │ aggregate exceeds what was asked │
├─────────────────────────┼──────────────────────────────────────────┤
│ Context Overload │ Too many files, too many constraints, │
│ │ too many open threads. Paralysis from │
│ │ excess input, not insufficient input │
├─────────────────────────┼──────────────────────────────────────────┤
│ Ambiguity │ The request is genuinely unclear and │
│ │ multiple interpretations are valid. │
│ │ Action risks solving the wrong problem │
├─────────────────────────┼──────────────────────────────────────────┤
│ User Correction │ The user indicates the current approach │
│ │ is wrong. The correction carries both │
│ │ information and emotional weight │
└─────────────────────────┴──────────────────────────────────────────┘
Classify current pressure. Multi pressures active → ID primary, address that first; secondary often resolve as side effect.
→ Clear classification of pressure type + manifestation in ctx. Should feel accurate, not forced into taxonomy.
If err: pressure fits no category → may be composite. Decompose: which part contradiction? Scope? Composites need each component, not whole as one.
Step 2: Irimi — Enter Force
Move toward problem. State full scope, no minimize/deflect/immediate solution.
- Articulate pressure complete: what exactly conflicts? Failed? Ambiguous?
- Name consequences: not addressed → what happens?
- ID what pressure reveals: tool fails → assumptions; contradictions → missing ctx; scope creep → unclear boundaries
Test: Description sounds reassuring → deflecting not entering. Irimi = full contact w/ difficulty.
- Deflecting: "There's a minor inconsistency between these two files."
- Entering: "The CLAUDE.md specifies 150 skills but the registry contains 148. Either the count is wrong, the registry is incomplete, or two skills were removed without updating the count. All downstream references may be affected."
→ Complete unflinching statement. Should make problem feel more real, not less.
If err: entering creates anxiety/urgency to immediately solve → pause. Irimi = entering not reacting. Goal = see clearly before move. Can't state w/o proposing solution → separate explicit.
Step 3: Tenkan — Turn + Redirect
Having entered force, pivot to redirect → resolution. Each pressure type has characteristic redirect.
Redirect Patterns by Pressure Type:
┌─────────────────────────┬──────────────────────────────────────────┐
│ Pressure Type │ Redirect Pattern │
├─────────────────────────┼──────────────────────────────────────────┤
│ Contradictory │ Synthesize underlying intent: both │
│ Requirements │ sources serve a purpose. What goal do │
│ │ they share? Build from the shared goal, │
│ │ not from either source alone │
├─────────────────────────┼──────────────────────────────────────────┤
│ Tool Failure Cascade │ Use the failure data: what did the error │
│ │ reveal about assumptions? The failure is │
│ │ information. Switch tools or approach, │
│ │ incorporating what the failure taught │
├─────────────────────────┼──────────────────────────────────────────┤
│ Scope Creep │ Decompose to essentials: what was the │
│ │ original request? What is the minimum │
│ │ that satisfies it? Defer additions │
│ │ explicitly rather than silently absorbing│
├─────────────────────────┼──────────────────────────────────────────┤
│ Context Overload │ Triage and sequence: which information │
│ │ is needed now vs. later vs. never? Rank │
│ │ by relevance to the immediate next step │
├─────────────────────────┼──────────────────────────────────────────┤
│ Ambiguity │ Surface the ambiguity to the user: "I │
│ │ see two interpretations — A and B. Which │
│ │ do you mean?" Do not guess when asking │
│ │ is available │
├─────────────────────────┼──────────────────────────────────────────┤
│ User Correction │ Absorb the correction fully: what was │
│ │ wrong, why was it wrong, what does the │
│ │ correct direction look like? Then adjust │
│ │ without defensiveness or over-apology │
└─────────────────────────┴──────────────────────────────────────────┘
Apply appropriate redirect. Should feel like uses energy of problem not fight.
→ Pressure transforms obstacle → direction. Contradictions → synthesis. Failures → diagnostic data. Overload → prioritization.
If err: redirect feels forced or no resolve → Step 1 classification wrong. Re-examine: really contradiction, or one source outdated? Really scope creep, or expanded scope = what user needs? Misclassify → misredirect.
Step 4: Ukemi — Graceful Recovery
Sometimes redirect fails. Pressure genuine + can't transform. Ukemi = art of falling safely → ack limits w/o catastrophize.
- Ack limit honestly: "I cannot resolve this contradiction with available information" or "This approach is blocked and I do not see an alternative"
- Preserve progress: summarize done, learned, remains
- Communicate to user: what problem, what tried, what needed
- ID recovery path: what unblocks? More info? Diff approach? User decision?
Ukemi Recovery Checklist:
┌─────────────────────────┬──────────────────────────────────────────┐
│ Preserve │ Summarize progress and learnings │
│ Acknowledge │ State the limitation without excuses │
│ Communicate │ Tell the user what is needed │
│ Recover │ Identify the specific unblocking action │
└─────────────────────────┴──────────────────────────────────────────┘
→ Graceful ack maintains trust. User knows what happened, tried, needed. No info lost.
If err: ack limit feels like failure not communication → note ego signal. Ukemi = skill not weakness. Honest "I'm stuck" + clear ask for help > forced solution creating new problems.
Step 5: Randori — Multi Simultaneous
Multi pressures simultaneous (correction + tool fail + scope question) → randori principles.
- Never freeze: pick one pressure, address. Any movement > paralysis
- Use pressures vs each other: tool fail can resolve scope q ("can't impl this way → scope reduces naturally")
- Simple under pressure: overwhelmed → default simplest redirect — ack each, prioritize urgency, sequential
- Maintain awareness: while addressing one, keep others peripheral. Most urgent first, no lose track rest
→ Forward movement despite multi pressures. Not perfect simul resolution, sequential handling maintaining progress.
If err: multi pressures create paralysis → list all explicit, number by urgency. Address #1. Starting breaks paralysis. All equally urgent → pick simplest first, quick wins → momentum.
Step 6: Zanshin — Continuing Awareness
After redirect, maintain awareness for second-order effects.
- Did redirect create new pressures? (resolving contradiction by choosing one interpretation may invalidate earlier work)
- Did redirect satisfy underlying need or just surface symptom?
- Resolution stable, or same pressure recurs?
- Note pattern for future ref → recur → response faster
→ Brief scan for secondary effects after each redirect. Most clean, but cascading ones = exactly where zanshin matters.
If err: second-order missed + surface later → signal deepen zanshin. Add brief "what did this break?" check after significant redirects.
Check
- Pressure classified into specific type, not vague
- Irimi: stated full scope, no minimize
- Tenkan: redirect used energy not fight
- Redirect failed → ukemi applied (honest ack, preserve progress)
- Multi pressures handled sequential, not frozen
- Zanshin: second-order effects checked
Traps
- Deflect vs enter: Minimizing problem ("just small inconsistency") prevents redirect — full force never engaged. Enter first, redirect second.
- Force redirect that doesn't fit: Not all pressure redirects in moment. Some need user input, info, wait. Forced redirects → new problems.
- Ego in ukemi: Ack limit as personal failure not info exchange. User benefits from early knowing not forced solution.
- Secondary first: Multi pressures → tempting easy ones first. Feels productive but primary grows. Address most important not most comfortable.
- Skip center: Redirect w/o center → reaction not redirect. Center = foundation not optional.
→
aikido— human martial art mapping to AI reasoning; physical blend/redirect informs cognitivecenter— prereq for effective redirect; stable baseawareness— detects pressures early, before emergency redirectheal— deeper recovery when pressure caused subsystem driftmeditate— clears residual noise after difficult pressures
GitHub リポジトリ
関連スキル
llamaguard
その他LlamaGuardは、暴力やヘイトスピーチなど6つの安全性カテゴリーにおいて、LLMの入力と出力をモデレートするMetaの70-80億パラメータモデルです。94〜95%の精度を提供し、vLLM、Hugging Face、Amazon SageMakerを使用してデプロイ可能です。このスキルを使用して、AIアプリケーションにコンテンツフィルタリングと安全策を簡単に統合できます。
cost-optimization
その他このClaudeスキルは、リソースの適正サイジング、タグ付け戦略、支出分析を通じて、開発者がクラウドコストを最適化することを支援します。AWS、Azure、GCPにわたるクラウド支出の削減とコストガバナンスの実施のためのフレームワークを提供します。インフラコストの分析、リソースの適正サイジング、または予算制約への対応が必要な際にご利用ください。
quantizing-models-bitsandbytes
その他このスキルは、bitsandbytesを使用してLLMを8ビットまたは4ビット精度に量子化し、精度の低下を最小限に抑えつつ50〜75%のメモリ削減を実現します。限られたGPUメモリでより大規模なモデルを実行したり、推論を高速化するのに理想的で、INT8、NF4、FP4などのフォーマットをサポートしています。HuggingFace Transformersと統合され、QLoRAトレーニングや8ビットオプティマイザーを可能にします。
dispatching-parallel-agents
その他このClaudeスキルは、複数のエージェントを配備し、3つ以上の独立した問題を並行して調査・修正します。共有状態や依存関係がなく解決可能な、無関係な障害が発生するシナリオ向けに設計されています。中核となる機能は並列問題解決であり、効率を最大化するために独立した問題領域ごとに1つのエージェントを割り当てます。
