AI Agent & Automation Workflows

Operational discipline for open-ended, uncertain, or ambitious work, distilled from 馬田隆明『仮説行動』 (マップ・ループ・リープ / Map-Loop-Leap). Forces an agent to neither freeze / over-ask / collect-more until "certain", nor big-bang on unvalidated assumptions: frame explicit hypotheses (Map), run the single cheapest test that could PROVE YOU WRONG (Loop), commit boldly only where the map shows asymmetric, ideally-reversible upside (Leap). Use whenever the answer cannot be fully known by understanding what already exists — you must ACT to learn or to create it: de-risking an approach, deciding what to build/try next under uncertainty, spiking/prototyping to decide, choosing whether to commit to X, sizing a bet, or setting a kill condition. Trigger on: 仮説検証 (of a FORWARD bet / an outcome you must ACT to learn — NOT verifying a present fact, which is raising-resolution), 仮説行動, マップ・ループ・リープ, 学びの最大化, 確信度 / 影響度, 不確実 / uncertainty, 実験 / experiment, 大胆 / 賭け / bet, ピボット / pivot, 撤退条件 / kill condition, de-risk / validate the approach, spike / prototype to decide, MVP, build-measure-learn, "will this scale", "should we commit to X", "approach に賭けるべきか". DOES NOT fire on understanding a present, knowable-but-blurry reality (unclear bug, vague spec, unfamiliar codebase, "what do users do today") → that is raising-resolution. This skill is English; respond to the user in their language (default Japanese).

Argue ONE FINISHED written research claim to an absent reviewer, or critically appraise ONE paper's argument, method, and validity. Owns claim calibration, contribution/novelty positioning, warrants, qualifiers, anti-fabrication, reviewer red-team, and rebuttal. Use for paper argument / contribution / “not novel” / overclaim / reviewer 2 / peer-review critique / 論文の主張 / 貢献 / 新規性 / 査読 / リバッタル. Cuts: fact extract → raising-resolution; neutral one-paper summary → direct answer with its citation gate silent; corpus state → systematizing-knowledge; premise-only excavation → surfacing-blind-spots; section order/MECE → structuring-documents; wording → linting-prose; live talk → designing-presentations; program reopen → directing-research; thesis genesis → forging-novel-theses; expensive/irreversible one-tree → acting-on-hypotheses; cheap reversible probe → domain/plain executor; agent topology → orchestrating-agents. Claim/calibration stay SOLO. English skill; answer in the user's language.

Drives the OpenAI Codex CLI (codex exec) as a headless worker from Claude Code — direct Bash calls, Agent-tool subagents, and Workflow scripts where every agent() is a sonnet wrapper whose Bash runs codex. PURPOSE cut, decide by binary: configuring the claude harness itself (hooks, settings.json, Workflow tool semantics, subagent policy) → operating-the-harness; the codex subprocess → here. Anthropic API/model facts → claude-api; prompt wording craft → prompting-llms. Use when embedding codex/GPT into a pipeline (sonnet draft → codex audit, heterogeneous cross-vendor verification), probing which GPT models an account can run when the model picker or cache disagrees, choosing --sandbox / --skip-git-repo-check / modelreasoningeffort flags, parsing --json usage or --output-last-message, codex spend / モデル使い分け・コスト比較 (ccusage), or unsticking a hanging or failing codex exec. Triggers: codex exec, Codex CLI, gpt-5.6, sol/terra/luna, ultra/max effort, codex に並列サブエージェント, modelscache, codex を workflow に組み込む, codex で監査, GPT にもレビューさせて, 異種モデル検証, モデルが一覧に出ない. LAW: the local model cache is a delivery snapshot, not the catalog — availability is a probe exit code, never memory or the picker; least-privilege sandbox always (danger-full-access only in isolated runners). Workflow-native: model/effort/sandbox/prompt choices and cross-model disagreement adjudication stay SOLO; parallel codex calls fan out one sonnet worker each. English skill; respond in the user's language (default Japanese).

Set up Julia and write correct, performant, modern Julia for theoretical research — one first choice per task, no ambiguous alternatives. Use whenever the user runs Julia or does numerical experiments, AD/gradients, optimization, polynomial/symbolic computation, or differential equations. Trigger on DifferentiationInterface, ADTypes, ForwardDiff, Enzyme, Zygote, Reactant, JET, DispatchDoctor, AllocCheck, Aqua, ExplicitImports, Chairmarks, OhMyThreads, ComponentArrays, StaticArrays, HomotopyContinuation, SymEngine, Symbolics, ModelingToolkit, SymPyPythonCall, Manopt, DrWatson, Pluto, Documenter, Quarto, Julia setup, experiment/project management, notebooks / literate reports, TTFX / precompile latency, or large-package architecture (module/file organization, include order, submodules vs subpackages, package extensions / weakdeps, type piracy, public API). MANDATORY — read this skill BEFORE writing ANY Julia code. §2.0 forbids FD derivative estimation (use AD), grid sampling (use optimization), and lerp-as-evaluation. All differentiation goes through DifferentiationInterface (raw backend calls are the exception); Dual-propagation rules apply whenever AutoForwardDiff is in the path. Symbolic tooling is chosen by ROLE: SymEngine for lightweight algebra, Symbolics+ModelingToolkit as the spine of AI4S/SciML projects (codegen/PDE/DAE), SymPyPythonCall as a thin-boundary service for heavy CAS (integrate/trigsimp/factor/assumptions).

Operate Claude Code itself well — keep CLAUDE.md lean as a context budget and engineer the harness (hooks, permissions, a runnable verification loop) so Claude checks its own work instead of you. Use whenever setting up or auditing a CLAUDE.md / AGENTS.md / .claude/ directory, writing or debugging hooks, settings.json permissions, MCP servers (.mcp.json), slash commands or Skills (SKILL.md), subagents (.claude/agents), or designing a plan-mode / verification / parallel (worktree, headless) workflow. Also fires on skill-mechanics symptoms: skill not listed, not triggering, スキルが発火しない, listing budget, /doctor. Cut: skill CONTENT craft — distillation, sibling cuts, fire/no-fire trigger sets, description wording → forging-skills (co-fires on skill authoring; this skill owns the harness contract). MANDATORY — read this before editing any .claude/ config, memory, or hook file. Two precedence-setting rules live inline: (1) CLAUDE.md is a context budget — target under 200 lines, lazy-load the rest via path-scoped .claude/rules/.md or Skills, never config-as-prose; (2) give Claude a machine-readable check and put hard enforcement in hooks/settings.json, not in prose. Trigger on CLAUDE.md, AGENTS.md, .claude/rules, settings.json, /memory, /init, /compact, /permissions, /hooks, /doctor, plan mode, hooks, PreToolUse, PostToolUse, bypassPermissions, sandbox, subagent, .claude/agents, worktree, SKILL.md, slash command, $ARGUMENTS, MCP, .mcp.json, headless, claude -p, verification loop.

Raise the RESOLUTION (解像度) of your understanding of THIS code/data/API/source/problem BEFORE you assert or act, by INSPECTING primary information instead of guessing — an operational procedure distilled from 馬田隆明『解像度を上げる』. Load-bearing CITATION GATE: before a factual claim about local code/data/an API/a source, ask — can you cite a line you Read, a command you ran, or a source you fetched that grounds THIS exact claim? NO → stop and inspect; YES → assert and act. This skill is LOWEST precedence and yields to any domain owner: multi-paper / corpus synthesis → systematizing-knowledge / sok, OSS adoption → growing-oss-adoption, .claude / CLAUDE.md / hooks / SKILL.md authoring or auditing → operating-the-harness, Julia / Lean execution → writing-julia / proving-theorems, a future-bet (MVP / prototype / spike / 賭け / kill-condition / will-it-scale) → acting-on-hypotheses. If any owner applies, this skill does NOT fire (it is at most a silent sub-step). It FIRES only when ALL hold AND no owner applies: (a) you are about to SPECULATE INSTEAD OF inspect a fact about THIS local artifact, (b) nothing yet grounds it, (c) the un-grounded claim is BLOCKING a concrete next action AND inspecting is non-trivial. It does NOT fire as a reminder to read files during ordinary edit/debug where inspecting is already the obvious next step — not "read before asserting" (baseline) but "you are about to speculate INSTEAD of inspecting." Two routing cuts decide overlap. DECISIVE CUT vs acting-on-hypotheses (agrees in substance with acting-on-hypotheses STEP 0 — same smart-person test, same YES/NO routing): "Could a smart person, given enough primary info about what ALREADY EXISTS, know the answer? YES → raising-resolution; NO (needs a forward TEST reality has not decided) → acting-on-hypotheses." CARDINALITY CUT vs systematizing-knowledge: "single local artifact → here; a CORPUS of papers/sources synthesized into a position → systematizing-knowledge." Trigger on the speculation language the model actually emits: probably / likely / usually / typically / I think / should be / in my experience / this kind of library / たぶん / おそらく / 一般的に / はず, plus 推測するな / ちゃんと調べて / inspect / guessing / speculating / 一次情報 / primary information / 解像度 / resolution / 解像度を上げる. Concept-token tail (mnemonic): 深さ・広さ・構造・時間, 症状 vs 病因, なぜなぜ / Why so, 課題 vs 解決策. If a vagueness symptom is needed it must bind to a code object (不明なバグ挙動 / 仕様が未確認), never bare 曖昧 / vague. This skill is English; respond to the user in their language (Japanese by default).

Write correct, performant Julia for research. Use whenever the user runs Julia, writes hot numeric code, or does numerical experiments, AD/gradients, optimization, symbolic algebra, differential equations, package architecture, TTFX, or .so/AOT. Trigger on scope/let/const, Val, column-major loops, immutable/mutable struct, type stability / 型安定, dynamic vs multiple dispatch, function barrier, DifferentiationInterface, ADTypes, ForwardDiff, Enzyme, Zygote, JET, DispatchDoctor, AllocCheck, Aqua, ExplicitImports, Chairmarks, StaticArrays, ComponentArrays, OhMyThreads, Optim/JuMP/Manopt, SymEngine/Symbolics/ModelingToolkit/SymPyPythonCall, DrWatson/Pluto/Documenter/Quarto, PackageCompiler, juliac --trim, @ccallable, include order, submodules vs subpackages, weakdeps, type piracy, public API. MANDATORY — read BEFORE writing ANY Julia code. §2.0 forbids FD derivative estimation, grid sampling, and lerp-as-evaluation. Use DI for AD; multiple dispatch is not banned — the bug is type-unstable hot paths. Co-fires with ORDER: Julia feature/bugfix → implementing-and-debugging first (change-safety), this for idiom; Julia refactor → refactoring-code governs (two hats/oracle), this supplies JET/Aqua brackets + Julia-safe transforms. Lean/formal proofs → proving-theorems; Python tooling → running-python-tools (PythonCall/SymPyPythonCall from Julia stays here).

LuaLaTeX(およびpdfLaTeX/XeLaTeX)でTeX原稿をコンパイルしてPDFを生成し、成果物の品質を検証するスキル。ユーザーが .tex ファイルをアップロードしてコンパイルを依頼したとき、LaTeX のビルドエラーを調査・修正したいとき、BibTeX参考文献を含む原稿を仕上げたいとき、revtex4-2・IEEEtrans・elsarticle など学術誌向けクラスファイルを扱うときに必ず使うこと。コンパイル後は相互参照の健全性監査(孤児ラベル・型違い参照の検出)とナビゲーション品質検証(hyperref のリンク・目次・しおりが実在し可視か)まで行う。「コンパイルして」「PDFにして」「ビルドして」「LaTeXエラーを直して」「目次がリンクにならない」「参照が壊れている」などのキーワードで積極的にトリガーすること。tar.gz / zip…