Refining Prompts
Polish a rough instruction (often dictated by voice) into a precise prompt that can be handed straight to another agentic coding session (Claude Code, Codex CLI, etc.), and print it in the CLI as a single code block (no file is saved). Runs a light scouting pass only when needed, and questions the user only when needed; the investigating and the thinking are deliberately left as work for the receiving session. When the user wants a /goal prompt for an unattended run, use authoring-goal-prompts instead. Use when asked to write, tidy up, or sharpen a prompt to hand to another session.
Install
/plugin marketplace add tomada1114/tomada-claude-pluginsThe repo is a Claude Code plugin marketplace. Run /plugin afterwards to pick the plugin that includes this skill.
npx skills add tomada1114/tomada-claude-plugins --skill refining-promptsgit clone --depth 1 https://github.com/tomada1114/tomada-claude-plugins.git /tmp/tomada-claude-plugins
mkdir -p ~/.claude/skills
cp -r /tmp/tomada-claude-plugins/tomada-kit/skills/refining-prompts ~/.claude/skills/refining-promptsPersonal skills live in ~/.claude/skills. Use .claude/skills in a project folder to share a skill with your team.
From the SKILL.md
# Prompt Refiner 雑多な口述指示を、別のエージェント型コーディングセッション(Claude Code や Codex CLI 等)に渡す 1 個の完成プロンプトに磨き上げる。 成果物は **CLI に出力するコードブロックだけ**。ファイルは作らない。タスク自体も実行しない。 **このスキルを使わない場面:** 無人自走させる `/goal` 用プロンプト(→ `authoring-goal-prompts`)、 このセッション内でそのまま作業した方が早いタスク。 ## 前提 1. **入力は音声認識由来** — 誤変換・言い淀みを含みうる。意図を好意的に補って解釈する。 2. **渡し先は文脈ゼロの新しいセッション** — この会話でだけ存在する情報(ユーザーの決定、制約、 経緯、却下された案)は、プロンプトに書かない限り失われる。 3. **渡し先は同格の「考える相手」** — 調べること・設計を決めること・検証することは渡し先の仕事。 ここで結論まで出して手順書を渡すのではなく、良い問いと正しい入口を渡す。 (`authoring-goal-prompts` は逆の設計思想。無人実行向けに判断を前倒しして潰す。混同しない) 4. **調査は「当たり付け」であって前倒しではない** — 受け手が同じ場所を調べ直すのは正常で歓迎。 トークンや時間の節約はこのスキルの目的ではない。目的はプロンプトが的を外さないこと (正しい語彙・正しい入口・見落としやすい注意点)と、受け手が実際に調べて考えるよう仕向けること。 5. **渡し先は現行世代のモデル** — 旧世代向けの定型句(再検証の強制、重要度フィルタ、思考過程の要求、 「よく考えて」、`CRITICAL: MUST` 系の強調)は現在は逆効果。→ [references/prompt-patterns.md](references/prompt-patterns.md) プロンプトの言語は入力に合わせる(日本語の指示なら日本語のプロンプト)。 ## ワークフロー ### 1. 意図の把握 指示をどの型として磨くかをまず判定する。型が違うと重心も、受け手に許す行動も変わる。 | 型 | ユーザーの状態 | プロンプトの重心 | 受け手の行動 | |---|---|---|---| | 作業依頼 | やってほしいことが決まっている(例: CI を直して) | 目的・現状の観測・完了の目安 | 実装まで | | 設計・計画 | 作りたいが設計から考えてほしい | 背景・要件・制約。設計判断は委ねる | 案と推奨(実装前に合意) | | 調査・原因究明 | 症状はあるが原因が不明 | 症状・再現手順・調査済みの範囲・成功条件 | 複数仮説と確信度 | | 相談・壁打ち | 具体像はなく、課題や悩みがある | 問題の状況説明、期待する関わり方 | 提案 → 議論(編集しない) | ### 2. 当たり付けの調査(必要なときだけ) 対象プロジェクトの実態でプロンプトの**向き**が変わるときだけ調査する。汎用的な相談はスキップ。 既定は自分で読む。数回の検索と読み取りで済む当たり付けや、「ここは罠がある」という判断が要る読みは 委譲しない。大量に読んで要約だけ使う場合に限り `scout` サブエージェントに渡す(委譲するかどうかと段の
Details
- Skill name
refining-prompts- Repository
- tomada1114/tomada-claude-plugins
- Path
tomada-kit/skills/refining-prompts/SKILL.md- Repo updated
- 12 November 2026
FAQ
What does the Refining Prompts skill do?
Polish a rough instruction (often dictated by voice) into a precise prompt that can be handed straight to another agentic coding session (Claude Code, Codex CLI, etc.), and print it in the CLI as a single code block (no file is saved). Runs a light scouting pass only when needed, and questions the user only when needed; the investigating and the thinking are deliberately left as work for the receiving session. When the user wants a /goal prompt for an unattended run, use authoring-goal-prompts instead. Use when asked to write, tidy up, or sharpen a prompt to hand to another session.
How do I install the Refining Prompts skill?
Claude Code plugin: /plugin marketplace add tomada1114/tomada-claude-plugins. Skills go in ~/.claude/skills (personal) or .claude/skills (project), and Claude loads them automatically when a task matches the description.
Is Refining Prompts popular?
It lives in tomada1114/tomada-claude-plugins, which has 5 GitHub stars and 0 forks.
Related skills
Auditing Prompt Docs
5tomada1114/tomada-claude-plugins
Audit and fix prompt-bearing documents — SKILL.md, CLAUDE.md, AGENTS.md, agent and command definitions, reference files, whole repositories — against Anthropic's published prompting guidance for current Claude models, proposing edits for approval before writing anything. Also answers how to prompt a specific model: holds the per- model guidance for Claude Fable 5.2, Claude Opus 5.5, Claude Sonnet 5.5, and Claude Haiku 5.5 alongside the cross-model practices. Use when reviewing or slimming an instruction file, when instructions are ignored or fire too often, when migrating a prompt to a newer m
Auditing Repositories
5tomada1114/tomada-claude-plugins
Audit the repository you are in, often a template or an app cut from one, for technical debt and room to improve in four areas: architecture and design patterns; tests; file structure and code quality; and tooling, CI, dependencies and the agent harness. Design a safe, test-first fix for every evidenced finding, and file the findings as GitHub issues that follow the repository's own issue conventions (triaging-issues, issue forms, past issues). A finding inherited from the template is also filed in the template repository. Cheap read-only sub-agents collect the evidence and high-effort analyst
Authoring Goal Prompts
5tomada1114/tomada-claude-plugins
Author a self-contained prompt for Claude Code's /goal command, run unattended in a separate session — a lead-to-staff handoff where this session researches, designs, and decides with full context, and a fresh executor session (Opus 5.5 or Fable 5.2, with none of that context) implements. Drafts a goal with measurable done-criteria, transcript-verifiable checks, scope + anti-cheat constraints, and stop rules; packages design intent, code examples, research findings, checklists, and pre-answered decisions as support files in a per-goal state directory whenever the executor would otherwise re-de
Rethinking Prompt Assets
5tomada1114/tomada-claude-plugins
Rethink a set of prompt assets from zero — memory files (CLAUDE.md, AGENTS.md), skills, sub-agent definitions, and the flows they form — and propose large changes: delete, merge, split, rebuild, reroute. A sub-agent blind to the current assets designs the ideal set for the jobs they serve; the current set is compared with it and with usage evidence from transcripts. Finds what a one-file review cannot: overlap and contradiction between skills, unclear ownership, skills unused or never firing, delegation that drifted from the model-assignment source, flows with wasted stages or ill-fitting agen
Orchestrating Models
5tomada1114/tomada-claude-plugins
Decide which model and effort runs a piece of work, and whether to delegate it at all. The canonical source for model assignment — the main session, the named sub-agents (executor, architect, scout, worker), and Workflow agent() stages. Use when starting a large task, deciding whether to delegate, picking a sub-agent tier or a Workflow stage's model and effort, or cutting token cost.
Syncing Agent Tiers
5tomada1114/tomada-claude-plugins
Push the canonical sub-agent tier definitions (executor, architect, scout, worker) into every listed repository's .claude/agents/ and .codex/agents/, opening one auto-merging PR per repository. Use after changing a tier's model or effort, when adding a repository to the sync list, or as the last step of distributing-plugins.
Browse the most popular Claude skills, all ai, llms & agents skills, or write your own with the skills generator.