derive-theoretical-result
について
このスキルは、Claudeが公式や定理を第一原理から厳密に段階的に導出し、各ステップを明示的に正当化することを可能にします。数学的命題の証明、教科書の結果の検証、学術作業のための自己完結的な導出の作成などのユースケースを想定して設計されています。主な特徴には、特殊なケースの検証や、確立された公理や定理に論理的推論を積み重ねることが含まれます。
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/derive-theoretical-result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
Derive Theoretical Result
Rigorous step-by-step derivation from axioms/first principles/theorems. Every step justified. Limiting cases checked. Final result + notation glossary.
Use When
- Formula/theorem from first principles (e.g., Euler-Lagrange from action)
- Math proof by logic from axioms
- Re-derive textbook → verify/adapt
- Extend known → more general (flat → curved spacetime)
- Self-contained → paper/thesis/report
In
- Required: Target result (equation, inequality, theorem, relation)
- Required: Starting point (axioms, postulates, prior results, Lagrangian/Hamiltonian)
- Optional: Proof technique (direct, contradiction, induction, variational, constructive)
- Optional: Notation conventions
- Optional: Known intermediate results citable w/o re-deriving
Do
Step 1: State assumptions + target
Contract before calc:
- Axioms + postulates: Every assumption listed. Physics: symmetry group, action principle, QM postulates. Math: axiom sys + prior lemmas.
- Target: Precise notation. Equation → both sides. Inequality → direction + equality conds.
- Scope: Domain of validity (e.g., "non-relativistic, spinless, 3D"). State what not covered.
- Notation: Define every symbol. Self-contained.
## Derivation Contract
- **Starting from**: [axioms, postulates, or established results]
- **Target**: [precise mathematical statement]
- **Domain of validity**: [restrictions and assumptions]
- **Notation**:
- [symbol]: [meaning and units]
- ...
→ Complete unambiguous statement. Notation up front.
If err: Target ambiguous/assumptions incomplete → clarify before proceed. Hidden assumptions → unreliable.
Step 2: Math toolkit
Tools + applicability:
- Algebra: Tensor, commutator, matrix, series. Verify prereqs (convergence, invertibility).
- Calc/analysis: ODE/PDE, integration + domain, functional derivs, contour, distributions. Verify regularity (differentiability, integrability, analyticity).
- Symmetry/group theory: Irreps, Clebsch-Gordan, character orthogonality, Wigner-Eckart.
- Topology/geometry (if applicable): Manifolds, bundles, connections + topo constraints (boundary terms, winding, index).
- Identities/lemmas: Specific ones invoked (Jacobi, Bianchi, integration by parts, Stokes). State explicitly, cite by name.
## Mathematical Toolkit
- **Algebra**: [techniques and prerequisites]
- **Analysis**: [calculus tools and regularity conditions]
- **Symmetry**: [group theory tools]
- **Identities to invoke**: [list with precise statements]
→ Checklist w/ applicability verified.
If err: Unverified prereqs (e.g., term-by-term diff w/o uniform convergence) → flag gap. Prove or state as additional assumption.
Step 3: Execute w/ justification
Every step labeled + justified:
- One op per step: No combining.
- Justification labels:
[by assumption]— stated axiom/assumption[by definition]— prior definition[by {identity name}]— named identity (e.g., "by Jacobi identity")[by Step N]— prior step[by {theorem name}]— external theorem (Step 2)
- Checkpoints (every 5-10 steps):
- Units/dimensions consistent
- Symmetries preserved
- Correct transformation props
- Branches: Case analysis → each branch labeled sub-derivation, merge.
## Derivation
**Step 1.** [Starting expression]
*Justification*: [by assumption / definition]
**Step 2.** [Result of operation on Step 1]
*Justification*: [specific reason]
...
**Checkpoint (after Step N).** Verify:
- Dimensions: [check]
- Symmetry: [check]
...
**Step M.** [Final expression = Target result]
*Justification*: [final operation] QED
→ Linear sequence, no logic gaps. Every step verifiable.
If err: Step doesn't follow → gap. Insert intermediates or identify new assumption. No "it can be shown" unless well-known identity listed Step 2.
Step 4: Limiting cases + special values
Validate vs known:
-
Limits (≥3): Simpler prior formula (non-rel limit), trivial case (coupling=0), extreme regime (high/low T).
-
Special values: Known independent (n=1 hydrogen, d=3).
-
Symmetry: Correct under group. Scalar → invariant. Vector → transforms right.
-
Consistency: Ward identities, sum rules, reciprocity.
## Limiting Case Verification
| Case | Condition | Expected Result | Derived Result | Match |
|------|-----------|----------------|----------------|-------|
| [name] | [parameter limit] | [known result] | [substitution] | [Yes/No] |
| ... | ... | ... | ... | ... |
→ All limits + special values match. Internally consistent.
If err: Failed limit → err in derivation. Trace to first step producing fail. Common: sign, missing 2/π, wrong combinatorial coeff, wrong order of limits.
Step 5: Final w/ notation glossary
Polished:
- Narrative: Intro para → motivation, approach, main result.
- Body: Steps from Step 3 cleaned. Group → logical blocks w/ headings.
- Result box: Highlighted, separated.
- Glossary: Every symbol + meaning + units + first occurrence.
- Assumptions summary: All in one place, postulates vs technical (smoothness, convergence).
## Final Result
> **Theorem/Result**: [precise statement with equation number]
## Notation Glossary
| Symbol | Meaning | Units | First appears |
|--------|---------|-------|---------------|
| [sym] | [meaning] | [units or dimensionless] | [Step N] |
| ... | ... | ... | ... |
## Assumptions
1. [Fundamental postulate 1]
2. [Technical assumption 1]
3. ...
→ Self-contained doc, followable start to finish w/o external refs (except cited identities + theorems).
If err: Too long (>~50 steps) → break into lemmas. Derive each, assemble main result citing lemmas.
Check
- All starting assumptions stated before first calc
- Every step labeled justification (no unjustified leaps)
- Units/dimensions consistent at every checkpoint
- ≥3 limiting cases checked + match
- Special values match known
- Result transforms correctly under stated symmetry
- Glossary defines every symbol
- No deferred "it can be shown"
- Domain of validity stated w/ result
Traps
- Hidden assumptions: Analyticity, convergence, integral existence w/o stating. Every regularity cond = assumption, declare.
- Sign errs: Most common mech err. Track at every step. Cross-check dim analysis (sign err → dim inconsistent).
- Dropped boundary terms: Integration by parts / Stokes → boundary terms vanish only under conds. State why (e.g., "field decays > 1/r at infinity").
- Order of limits: Wrong order → diff results (thermodynamic before zero-T). State order explicit + justify.
- Circular reasoning: Using result as intermediate. Subtle for "obvious" formulas. Every step from stated start, not answer familiarity.
- Notation collisions: Same symbol for diff quantities (E = energy + E-field). Glossary prevents — IF written before derivation.
→
formulate-quantum-problem— formulate QM framework before derivingsurvey-theoretical-literature— find prior derivations for comparison
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。
