dissolve-form
について
`dissolve-form`スキルは、中核機能を維持しながら、硬直化・石灰化したシステム構造を制御された形で解体します。このスキルは、技術的負債がすべての進歩を阻害する場合や、漸進的変更が不可能な状況(通常は`assess-form`評価の後)で使用されます。プロセスには、硬直性のマッピング、溶解順序の決定、安全な分解が含まれ、その後のアーキテクチャ再構築に向けてシステムを柔軟化させます。
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/dissolve-form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
Dissolve Form
Controlled dismantle of rigid systems → dissolve calcified arch + tech debt + org rigidity, preserve imaginal discs → seed new form.
Use When
assess-form→ PREPARE or CRITICAL (too rigid to transform direct)- So calcified incremental change impossible
- Tech debt blocks all fwd progress
- Org structure too rigid for new reqs
- Before
adapt-architecture→ soften before reshape - Legacy decommission → extract value before shutdown
In
- Required: Form assessment → high rigidity (
assess-form) - Required: ID essential capabilities → preserve (imaginal discs)
- Optional: Target form (post-dissolve; may be unknown)
- Optional: Timeline + constraints
- Optional: Stakeholder concerns on specific components
- Optional: Prior dissolve attempts + outcomes
Do
Step 1: ID Imaginal Discs
Bio metamorphosis → imaginal discs = cell clusters in caterpillar → survive dissolve → become butterfly organs. ID what must survive.
- Catalog every capability:
- User features
- Data processing
- External integrations
- Institutional knowledge in code/process
- Business rules (often implicit, undoc)
- Classify:
- Imaginal disc (must survive): core biz logic, critical integrations, irreplaceable data
- Replaceable tissue (rebuild): UI, infra, standard algos
- Dead tissue (discard): workarounds for fixed bugs, shims for dead systems, unused features
- Extract imaginal discs → portable:
- Doc biz rules explicit (may only exist as comments or tribal)
- Extract algos → standalone tested modules
- Export data → format-independent
- Record integration contracts + actual (not doc) behavior
→ Clear capability inventory: preserve/rebuild/discard. Essentials extracted portable before dissolve starts.
If err: imaginal disc ID uncertain (stakeholder disagree) → err on preserve side. Extract more than need → discard after dissolve easy, recover lost knowledge often impossible.
Step 2: Map Dissolve Sequence
Order → outer layers first, core last.
- Order by dep depth:
- L1 (outermost): no dependents → nothing breaks on remove
- L2: dependents only in L1 (already dissolved)
- L3: deeper deps → careful interface mgmt
- LN (core): load-bearing → dissolved last
- Per layer:
- What dissolved (removed, decommission, archive)
- What replaces (new comp, nothing, stub)
- Interfaces to maintain for remaining layers
- How verify post-dissolve
- Dissolve checkpoints:
- Post-layer → tested + verified operational
- Each checkpoint = stable state → pause poss
- If layer dissolve breaks → restore prior checkpoint
Dissolution Sequence (outside in):
┌─────────────────────────────────────────────────────────────────┐
│ Layer 1: Dead features, unused integrations, orphaned code │
│ → Remove. Nothing depends on these. │
│ │
│ Layer 2: Replaceable UI, standard infrastructure │
│ → Replace with modern equivalents or stubs │
│ │
│ Layer 3: Business logic wrappers, data access layers │
│ → Extract imaginal discs, then dissolve │
│ │
│ Layer 4 (core): Load-bearing structures, data stores │
│ → Dissolve last, with full replacement ready │
└─────────────────────────────────────────────────────────────────┘
→ Layer-ordered sequence, each step safe (checkpoint) + reversible (prior checkpoint restorable). Most critical dissolved last when team has most exp + confidence.
If err: dep mapping reveals circular (A→B→A) → break cycle first. Add interface A↔B, break cycle, then proceed.
Step 3: Interface Archaeology
Before dissolve → excavate + doc actual interfaces, not documented, actual in use.
- Instrument interfaces:
- Log every call/msg/exchange at each interface
- Run ≥1 full biz cycle (daily/weekly/monthly)
- Capture actual payload shapes, not just doc schemas
- Compare actual vs documented:
- Doc interfaces never called? (L1 candidates)
- Undoc interfaces actively used? (hidden deps → preserve or explicit replace)
- Edge cases in traffic doc doesn't mention
- Build contract from actual behavior:
- Contract → spec for replacement
- Real input/output examples
- Doc actual error handling (not should-be)
→ Empirical contract: actual communication + undoc behaviors + hidden deps.
If err: instrumentation too invasive (perf or code changes) → sample traffic. Biz cycle too long → available data + stakeholder interviews on "what calls what when".
Step 4: Execute Dissolve
Systematic remove + maintain imaginal disc viability.
- Start L1 (outermost, no dependents):
- Remove dead features + unused code
- Archive (don't delete) for reference
- Verify: all tests pass, no runtime errs
- Per layer:
- Per component dissolved: a. Verify imaginal discs extracted (Step 1) b. Install replacement or stub (if dependents remain) c. Remove component d. Run valid. suite e. Monitor for side effects
- Per checkpoint: doc state + verify operational
- Handle resistance:
- Some resist (hidden deps surface)
- When remove breaks: a. Restore checkpoint b. Investigate hidden dep c. Add to interface archaeology (Step 3) d. Explicit stub for dep e. Re-attempt
- Track progress:
- Components remaining vs dissolved
- Imaginal discs extracted + verified portable
- Unexpected deps found + handled
→ Systematic verified dissolve of non-essential. Post-layer: smaller, simpler, operational. Imaginal discs preserved portable.
If err: cascading failure → layer order wrong, hidden deps deeper than expected. Stop, restore, remap deps, re-sequence. Imaginal disc more complex than expected → more extract time.
Step 5: Prep Foundation for Reconstruction
Post-dissolve → minimal viable core + extracted imaginal discs ready.
- Assess post-state:
- What remains? (minimal core + extracted capabilities)
- Maintainable? (team understands + modifies)
- All imaginal discs accessible + verified? (portable + tested + doc'd)
- Reconstruction manifest:
- Per disc: contract + data + test suite
- Target arch (or TBD)
- Gaps: partial extracts or quality concerns
- Handoff:
- Target known →
adapt-architecturew/ minimal core - Target unknown → operate on minimal core while designing
- Either way: system flexible → reshape-able
- Target known →
→ Minimal maintainable system + doc'd extracted capabilities. Foundation clean, ready for reconstruction in any form.
If err: post-state less maintainable than expected → some essential was dissolved. Check imaginal disc inventory → if critical capability missing, may still be in archive. Minimal core too minimal → "replaceable" was actually essential → restore from checkpoint.
Check
- Imaginal discs IDed + extracted + verified portable
- Sequence: outermost (no dependents) → core
- Interface archaeology → actual (not just doc) behavior
- Each layer verified checkpoint
- No essential lost
- Post-state minimal + maintainable + operational
- Reconstruction manifest: capabilities + gaps
Traps
- Dissolve w/o extract: Remove rigid component before extract → destroys irreplaceable knowledge. Extract imaginal discs first.
- Trust docs over observation: Docs often diverge from actual. Interface archaeology reveals truth, docs show intent.
- Core first: Load-bearing before dependents → cascading failure. Outside-in.
- Total dissolve: Everything gone "clean slate" → lose institutional knowledge, edge-case handling, operational continuity. Preserve imaginal discs.
- Dissolve as punishment: "Because it's bad" w/o reconstruction plan → vacuum. Dissolve is prep for rebuild, not end itself.
→
assess-form— prereq assessment → IDs rigidity, triggers dissolveadapt-architecture— reconstruction after dissolverepair-damage— targeted repair vs full dissolvebuild-consensus— consensus before dissolve → prevents team fragmentationdecommission-validated-system— formal decommission for regulatedconduct-post-mortem— shares investigative rigor w/ dissolve
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。
