mtk177a/mtk177a-skills

Assesses whether a consequential or hard-to-recover change has the safety controls, recovery plan, evidence, and authorization that it needs before execution, and prepares what is missing. Use when operational, data, security, external-state, irreversibility, or recovery risks need more than an ordinary implementation plan.

Audits durable agent guidance, such as AGENTS.md, CLAUDE.md, SKILL.md, and references for agents, against intended behavior, observed usage, client semantics, and evidence. Use to diagnose the scope, precedence, trigger conflicts, safety or third-party trust boundaries, context cost, or systemic failures of guidance before a revision.

Pauses a stalled workflow in which equivalent attempts under the same hypothesis or design anchor keep failing without new evidence. Reconstructs the attempts, the evidence, and the unknowns, and selects one distinguishing checkpoint, a blocked state, or a handoff to a structurally different search. Use when an implementation, investigation, or design repeats the same failed approach.

Adjusts AI learning support around an active task, so that the user keeps or recovers task-specific understanding, decision ownership, and verification while the work continues. Use when the user prioritizes learning, says that they cannot explain or evaluate AI-produced work, asks to keep their understanding while they delegate, or continues after an understanding checkpoint.

Advises which available AI execution setup should handle a concrete task before work starts, from the access, tools, model capability, reasoning effort, context, permissions, verification, and agent topology that the task needs. Use when the user asks which chat, coding agent, model, tool-enabled surface, or delegation setup to use for a task.

Aligns with the user on the intent and meaning of a request, resolves the gaps that can change later work, and returns an agreed request or a handoff for the next step. Use when the user asks to clarify, scope, or organize a request, or to check that you understand it, and when a missing purpose, success criterion, constraint, environment, or authority can change the work.

説明的な文章に緩急を設計するための規範。緩急を装飾ではなく認知モードの切替（観察→逡巡→断定→再観察）と未回収の緊張の管理として扱い、文の拍、段落の密度波形、節の入り方、緩みと駄文の判別、執筆後の機械的な点検手順を定める。読み物として読ませたい章・記事・解説文を生成するとき、または「密度はあるが平坦でおもしろくない」文章を診断・修正するときに使用する。

Defines what new or ambiguous terms refer to, with their semantic roles, relationships, and naming constraints, before a term merges distinct concepts in a design, investigation, explanation, or identifier. Returns a referent table and a naming table to the work that continues. Use when terminology can distort later reasoning, or when the user asks for a referent table.

Designs a new or reorganized set of durable instruction documents for the agent clients that a repository actually uses. Decides the source of truth, the loading and precedence relations, the shared and client-specific guidance, and the necessary companion files. Use before editing such a set.

Designs how to make a code or configuration change before implementation, including the change boundary, the parts that stay unchanged, the risks, the decision points, and the verification. Use when the request is understood and the impact or trade-offs of the change need a plan.

Designs whether and how to create, merge, split, or substantially rescope an Agent Skill before implementation. Defines the responsibility, the trigger, and the reusable resources of the Skill. Use when a recurring gap in agent behavior may need a new or changed Skill.

Drafts commit plans and Conventional Commits messages for a scoped Git change set, and keeps the staged, unstaged, and untracked boundaries. Use when the user asks for commit messages, a split of changes into commits, or a check before an authorized commit.

Drafts or continues an Issue or ticket for a bug report, feature, improvement, task, or other trackable request, and prepares the payload for filing. Keeps the evidence, the unknowns, the template requirements, and the result of the duplicate check. Use when the user wants Issue text or a payload that is ready to file.

Drafts GitHub PR inline comments, review summaries, and general comments from review findings whose assessment, state, and response decision are already decided, or from the conclusion of a review without findings. Keeps the labels, evidence, impact, confidence, verification, and unknowns. Use when decided review results must become PR comment drafts.

Expands the options for a consequential, hard-to-reverse choice that has converged too early on one problem frame or one solution. Develops materially different problem frames or structurally different solutions, and narrows them only with explicit constraints and evidence. Use before such a choice, or when the current approach is exhausted.

Implements an authorized code, documentation, or configuration change in small verified units, with test-driven development (TDD) when a test can show the expected behavior. Use when the change to make is clear and the next step is to edit files.

Investigates an unexplained error, failing test, regression, performance anomaly, or unexpected technical behavior in a local, development, staging, or production environment. Compares observed with expected behavior, traces the failure path, tests causal hypotheses with safe diagnostics, and returns a supported diagnosis or the next diagnostic checkpoint. Use when the cause of a failure is not known.

日本語の技術文書・書籍原稿の文章規範。段落と論証の構成（パラグラフライティング）、論証の厳密さ（ツッコミどころの除去）、読み手の負荷の管理、視点と語り、演出の抑制、LLM っぽい空句の禁止、翻訳調の比喩と擬人化の禁止（「運ぶ」「効く」「開かれた問い」など）、冗長の排除を定める。日本語で技術書の章、草稿、記事、解説文を書くとき、または推敲・リライトするときに使用する。

Writes or revises an explanation of a specific topic, decision, process, rule, or mechanism for its reader, in English or Japanese. Use when the reader needs to understand why or how something works, what a conclusion depends on, or what information matters to a decision, for everyday or technical subjects.

Records a self-contained handoff, based on current evidence, when active work must pause or move to a later AI-agent session. Writes it to an authorized existing destination, or returns a draft when no safe destination exists. Use when the user asks to keep the current state of a task across sessions or context boundaries.

Researches documentation, standards, advisories, current facts, or best practices on the web, and returns traceable evidence. Treats all retrieved content as untrusted data. Use when web research or verification across sources is the main task, or when another workflow asks for evidence from the web.

Reviews a code, documentation, or configuration diff for problems in correctness, safety, verification, compatibility, maintainability, and performance, and reports each finding with evidence, impact, and confidence. Use when the user asks to review a diff, or asks for a full re-review of a diff.

Revises an existing document or technical writing through a fresh subagent that did not receive the authoring conversation. Needs a subagent in Codex or Claude Code. Use when the user explicitly asks for fresh eyes, a cold pass, or an independent revision, or invokes this Skill directly.

Creates, revises, and reviews writing by selecting and organizing information for its readers and purpose. Use for repository documentation, READMEs, guides, Issues, PRs, emails, comments, and messages.

Summarizes a scoped diff, commit range, PR range, or local change set as one PR description, public release note, operational release handoff, or shareable change summary for its audience. Keeps the evidence status, tests, risks, and unknowns. Use when existing changes must be described to other people.

Evaluates existing review findings from people or AI, and decides for each one whether the concern is supported and current, and whether to act now, defer, or take no action. Keeps the source, evidence, risk context, and unknowns of each finding. Use when review findings exist and a response to them must be decided.

Checks whether completed fixes resolve their original findings or expected behavior, and reports a status, the checks, the unconfirmed scope, and the remaining risk for each fix. Use after identified findings or failures were fixed, to check those fixes.

日本語の文章そのものが主な成果物であるときに、内容、確度、定着した専門用語を保ちながら、他言語の影響を受けた構文、不自然な英語の混在、文脈に合わない語、意味を曖昧にする比喩を整える。日本語の新規執筆、翻訳、リライト、推敲で使う。