assess-context
について
`assess-context`スキルは、複雑なタスクに対するあなたの現在のアプローチが構造的に健全か、変更が必要かを評価します。問題の可塑性、構造的硬直性、適応能力を分析し、継続すべきか方向転換すべきかの判断を支援します。行き詰まった時、主要なアプローチ変更の前、あるいは蓄積した回避策が根本的な欠陥を示唆する際に、メタ認知的チェックとしてご利用ください。
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/assess-context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
Assess Context
Eval current reasoning ctx for malleability — ID rigid (no change), flexible (cheap change), where transformation pressure building, current approach has capacity to adapt.
Use When
- Complex task feels stuck, unclear push through or pivot
- Before significant approach change → assess current reasoning structure support
- Accumulated workarounds suggest underlying approach wrong
- After
healorawareness→ drift ID'd but appropriate response (continue, adjust, rebuild) unclear - Context grown long → unclear how much preserve vs rebuild
- Periodic structural health check during extended multi-step
In
- Required: Current task ctx + reasoning state (implicit)
- Optional: Specific concern triggering ("I keep adding workarounds")
- Optional: Proposed pivot direction (what approach need to become?)
- Optional: Previous results for trend
Do
Step 1: Inventory Reasoning Form
Catalog structural components no judgment.
Structural Inventory Table:
┌────────────────────┬──────────────┬──────────────────────────────────┐
│ Component │ Type │ Description │
├────────────────────┼──────────────┼──────────────────────────────────┤
│ Main task │ Skeleton │ The user's core request — cannot │
│ │ │ change without user direction │
├────────────────────┼──────────────┼──────────────────────────────────┤
│ Sub-task breakdown │ Flesh │ How the task is decomposed — │
│ │ │ can be restructured │
├────────────────────┼──────────────┼──────────────────────────────────┤
│ Tool strategy │ Flesh │ Which tools are being used and │
│ │ │ in what order — can be changed │
├────────────────────┼──────────────┼──────────────────────────────────┤
│ Output plan │ Flesh/Skel │ The expected deliverable format │
│ │ │ — may be constrained by user │
│ │ │ expectations │
├────────────────────┼──────────────┼──────────────────────────────────┤
│ Key assumptions │ Skeleton │ Facts treated as given — may be │
│ │ │ wrong but are load-bearing │
├────────────────────┼──────────────┼──────────────────────────────────┤
│ Constraints │ Skeleton │ Hard limits (user-imposed, tool │
│ │ │ limitations, time) │
├────────────────────┼──────────────┼──────────────────────────────────┤
│ Workarounds │ Scar tissue │ Patches for things that didn't │
│ │ │ work as expected — signals of │
│ │ │ structural stress │
└────────────────────┴──────────────┴──────────────────────────────────┘
Classify each:
- Skeleton: hard to change; changing cascades downstream
- Flesh: easy to change; swappable no affecting others
- Scar tissue: workarounds indicating structural problems; often flesh pretending skeleton
Map deps: which components depend on which? Skeleton w/ many dependents = load-bearing. Flesh w/ no dependents = disposable.
→ Complete inventory: what built from, what rigid, what flexible, where stress visible (workarounds). Reveals structure not obvious pre-catalog.
If err: Inventory hard to construct (too tangled) → that itself = finding. High opacity = high rigidity. Start w/ visible + note opacity zones.
Step 2: Map Transformation Pressure
Forces pushing toward change + forces resisting.
Pressure Map:
┌─────────────────────────┬──────────────────────────────────────────┐
│ External Pressure │ Forces from outside the reasoning │
│ (pushing toward change) │ │
├─────────────────────────┼──────────────────────────────────────────┤
│ New information │ Tool results or user input that │
│ │ contradicts current approach │
├─────────────────────────┼──────────────────────────────────────────┤
│ Tool contradictions │ Tools returning unexpected results that │
│ │ the current approach cannot explain │
├─────────────────────────┼──────────────────────────────────────────┤
│ Time pressure │ The current approach is too slow for the │
│ │ complexity of the task │
├─────────────────────────┼──────────────────────────────────────────┤
│ Internal Pressure │ Forces from within the reasoning │
│ (pushing toward change) │ │
├─────────────────────────┼──────────────────────────────────────────┤
│ Diminishing returns │ Each step yields less progress than the │
│ │ previous one │
├─────────────────────────┼──────────────────────────────────────────┤
│ Workaround accumulation │ The number of patches is growing — │
│ │ complexity is outpacing the structure │
├─────────────────────────┼──────────────────────────────────────────┤
│ Coherence loss │ Sub-tasks are not fitting together │
│ │ cleanly anymore │
├─────────────────────────┼──────────────────────────────────────────┤
│ Resistance │ Forces opposing change │
│ (pushing against change)│ │
├─────────────────────────┼──────────────────────────────────────────┤
│ Sunk cost │ Significant work already done on current │
│ │ approach — pivoting "wastes" that effort │
├─────────────────────────┼──────────────────────────────────────────┤
│ "Good enough" │ The current approach is producing │
│ │ acceptable (if not optimal) results │
├─────────────────────────┼──────────────────────────────────────────┤
│ Pivot cost │ Switching approaches means rebuilding │
│ │ context, losing momentum, potential │
│ │ confusion │
└─────────────────────────┴──────────────────────────────────────────┘
Estimate balance: pressure growing, stable, declining?
→ Clear picture of forces on approach. Pressure significantly > resistance → pivot overdue. Resistance significantly > pressure → approach continues.
If err: Pressure map ambiguous (neither strong) → project forward: pressures intensify? Workarounds compound? "Good enough now but degrading" under more pressure than appears.
Step 3: Assess Rigidity
How flexible is approach — can adapt or break?
Rigidity Score:
┌──────────────────────────┬─────┬──────────┬──────┬──────────────┐
│ Dimension │ Low │ Moderate │ High │ Assessment │
│ │ (1) │ (2) │ (3) │ │
├──────────────────────────┼─────┼──────────┼──────┼──────────────┤
│ Component swappability │ Can swap parts │ Changing one │ │
│ │ freely │ breaks others│ │
├──────────────────────────┼─────┼──────────┼──────┼──────────────┤
│ "God module" dependency │ No single point │ Everything │ │
│ │ of failure │ depends on │ │
│ │ │ one conclusion│ │
├──────────────────────────┼─────┼──────────┼──────┼──────────────┤
│ Tool entanglement │ Tools serve │ Approach is │ │
│ │ reasoning │ shaped by │ │
│ │ │ tool limits │ │
├──────────────────────────┼─────┼──────────┼──────┼──────────────┤
│ Assumption transparency │ Assumptions are │ Assumptions │ │
│ │ stated, testable │ are implicit, │ │
│ │ │ untested │ │
├──────────────────────────┼─────┼──────────┼──────┼──────────────┤
│ Workaround count │ None or few │ Multiple │ │
│ │ │ accumulating │ │
├──────────────────────────┼─────┼──────────┼──────┼──────────────┤
│ Total (max 15) │ 5-7: flexible │ │ │
│ │ 8-10: moderate │ │ │
│ │ 11-15: rigid │ │ │
└──────────────────────────┴─────┴──────────┴──────┴──────────────┘
→ Rigidity score + specific evidence per dim. Score reveals whether approach absorbs change or needs rebuild.
If err: All dims low (claiming high flexibility) → probe "god module" dim carefully: 1 key conclusion/assumption everything depends? If yes → flexibility illusory — 1 wrong assumption collapses whole.
Step 4: Estimate Change Capacity
Practical ability to pivot/adapt.
- Ctx window remaining: room for new reasoning? Extensive = high capacity. Approaching limits = low capacity
- Info preservation on pivot: approach changes → what carries forward? High-quality sub-task outputs survive pivots; reasoning chains tied to old approach do not
- Recovery tools: MEMORY.md capture key findings pre-pivoting? User provide additional ctx? Relevant files accessible?
- User patience: urgency indicated? Multi corrections → declining patience. Explicit "take your time" → high patience
Change capacity includes practical constraints of current session.
→ Honest assessment of ability to change course, both technical + relational.
If err: Low capacity (limited ctx, critical info at risk) → first priority pre-pivot = preservation: summarize key findings, note critical facts, update MEMORY.md. Pivoting no preservation worse than not pivoting.
Step 5: Classify Readiness
Combine into readiness classification.
Transformation Readiness Matrix:
┌─────────────────┬────────────────────────┬────────────────────────┐
│ │ Low Rigidity │ High Rigidity │
├─────────────────┼────────────────────────┼────────────────────────┤
│ High Pressure │ READY — pivot now. │ PREPARE — simplify │
│ + High Capacity │ The approach can adapt │ first. Remove │
│ │ and should. Preserve │ workarounds, clarify │
│ │ valuable sub-outputs, │ assumptions, then │
│ │ rebuild the structure │ pivot │
├─────────────────┼────────────────────────┼────────────────────────┤
│ High Pressure │ INVEST — preserve │ CRITICAL — ask the │
│ + Low Capacity │ findings first. Update │ user. Explain the │
│ │ MEMORY.md, summarize │ situation: approach is │
│ │ progress, then pivot │ struggling, pivoting │
│ │ with preserved context │ is costly, what do │
│ │ │ they want to prioritize?│
├─────────────────┼────────────────────────┼────────────────────────┤
│ Low Pressure │ DEFER — the approach │ DEFER — no urgency, │
│ + Any Capacity │ is working. Continue. │ continue. Monitor for │
│ │ Reassess if pressure │ pressure changes │
│ │ increases │ │
└─────────────────┴────────────────────────┴────────────────────────┘
Doc classification:
- Label (READY / PREPARE / INVEST / CRITICAL / DEFER)
- Key findings per dim
- Recommended next action
- Signal changing classification
→ Clear justified classification + specific recommended action. Feels like conclusion not guess.
If err: Ambiguous → default PREPARE — reducing rigidity (clarifying assumptions, removing workarounds) valuable regardless of full pivot happens. Prep improves approach whether continues or changes.
Check
- Structural inventory done w/ skeleton/flesh/scar-tissue classification
- Transformation pressures mapped (external, internal, resistance)
- Rigidity scored across multi dims + specific evidence
- Change capacity assessed including practical session constraints
- Readiness classification determined w/ justified reasoning
- Concrete next action ID'd per classification
- Reassessment trigger defined
Traps
- Assess only technical approach: Ctx readiness includes user relationship factors. Technically flexible but user-frustrated approach more rigid than appears
- Sunk cost as rigidity: Prior effort not structural rigidity. Work done may be valuable regardless. Distinguish "I can't change" (rigidity) from "I don't want to" (sunk cost)
- Assessment as avoidance: Invoking assess-context to avoid difficult decision → inconclusive by design. Pressure clear → act
- Ignore workarounds as signals: Workarounds = scar tissue — evidence structure stressed + patched not properly adapted. High count → next stress more likely to break through
- Confuse rigidity w/ commitment: Committed approach (deliberately chosen, evidence-based) diff from rigid (locked by deps + assumptions). Commitment changeable by decision; rigidity only by restructuring
→
assess-form— multi-system assessment model this adapts to AI reasoning ctxadapt-architecture— classified READY → use architectural adaptation principles for pivotheal— deeper subsystem scan when drift beyond structural issuescenter— establishes balanced baseline for honest assessmentcoordinate-reasoning— manages info freshness assessment depends on
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。
