rpi
把改動依序走過研究、設計、實作三階段，文件存在任務目錄
中文翻譯
--- name: rpi description: 只有在使用者明確用名稱呼叫這個 skill 時才使用。 --- # Research, Design, and Implement 幫使用者研究一個改動、把設計想清楚，然後實作出來。依照使用者要求的步驟，使用下面的各階段指示。 `$SKILLBASE` 指的是包含這個 `SKILL.md` 的目錄。 範本裡的 `<rpi-invocation>` 指的是 host 的 skill 指令：Codex 是 `$rpi`，Claude Code 是 `/rpi`，host 有提供的話則是它加上 plugin 前綴的名稱。把後續的指令用這個原生的呼叫方式輸出。 ## 選擇步驟 依照使用者要求的步驟。要選擇或說明流程、或接下來是什麼的時候，讀 `$SKILLBASE/references/workflow.md`。 如果只有任務描述、沒有指定步驟，就從研究問題開始。問使用者想走哪種流程：研究後接設計討論、PRD 或 TDD，或只做研究。等他們選好再開始那個階段，然後把選擇帶進建議的下一個指令。 請使用者選擇流程時，讀 `$SKILLBASE/references/workflow-choice.md`，並把它的方框連線圖連同問題一起顯示。 Create 會寫一份新文件；iterate 則是依使用者的回饋就地更新既有文件，或是接續一份做到一半的文件的流程。 從 `$SKILLBASE/references/<step>/process.md` 載入對應的流程： | 請求 | 步驟 | | --- | --- | | 研究問題 | `create-research-questions` 或 `iterate-research-questions` | | 研究 | `create-research` 或 `iterate-research` | | 設計討論 | `create-design-discussion` 或 `iterate-design-discussion` | | 產品需求 / PRD | `create-prd` 或 `iterate-prd` | | 技術設計 / TDD | `create-tdd` 或 `iterate-tdd` | | 實作大綱 | `create-structure-outline` 或 `iterate-structure-outline` | | 實作 | `implement-outline` | | 實作的回饋或修正 | `iterate-implementation` | 預設更新既有文件；需要或被要求時才建新的。有多份符合時，用最新的那份，並說明選了哪個檔案。 ## 使用任務目錄 使用者有給目錄就用那個。 否則，在 `.agents/artifacts/` 裡找既有的任務目錄，名稱和 workflow 檔名要跟請求或目前任務的 branch 相符。明顯符合的就直接重用，不用問。如果目錄看起來接近、但可能屬於別的任務，或有好幾個看起來都合理，寫入之前先問使用者要用哪個。檢查是否符合時，只讀所選階段允許讀的文件。 沒有符合的現有目錄時，如果 branch 能辨識出任務，就用 `.agents/artifacts/<branch-name>`，把 `/` 換成 `-`。在 `main` 上，或沒有 branch 能辨識任務時，從請求挑一個簡短的 kebab-case slug，用 `.agents/artifacts/<task-slug>`，不用問是否同意。說明選了哪個路徑，並視需要重用或建立目錄。 存在選定的路徑。如果 sandbox 擋住那裡的寫入，就請求限定範圍的寫入權限，或要一個可寫的任務目錄，然後等使用者決定。 RPI 框架的作者建議不要 commit 這些 artifact。 新的 workflow 文件，用任務目錄裡下一個可用的 `NN-` 前綴編號。更新文件時維持原有路徑。把輔助的 mockup 和圖表存在文件旁邊，當作一般的本機檔案連結。 Workflow 文件的 frontmatter 包含 `repos`，是零個或多個 repo 的清單，每個項目有 `identifier` 和 `sha`。沒有適用的就用 `repos: []`。字串值視需要加引號，確保是合法的 YAML。 ## Review 文件 寫完或更新 workflow 文件之後，讀 `$SKILLBASE/references/prose.md` 和 `$SKILLBASE/references/show-me.md`，然後依這兩份來 review 文件。讓它更短、更視覺化、更容易讓人消化：砍掉沒有存在價值的字和段落，用能講清楚重點的最小視圖取代密集的文字，並讓每個視覺元素緊貼著它所支持的簡短文字。 ## 建議下一步 用該階段的交接指示和 `$SKILLBASE/references/workflow.md` 來決定接下來做什麼。使用者預設每個階段都在獨立的 context window 裡執行。建議下一個指令時，要附上任務目錄和選定的流程。並說明他們可以在這裡跑、在 compaction 之後跑，或開新的 session 跑。 被問到下一步是什麼時，用 artifact 的檔名找出目前的階段，然後讀它的文件，看有什麼未完成的工作和已定案的選擇。要能辨識 `prd` 或 `product-requirements`、`tdd` 或 `tech-design`，以及 `structure-outline` 或 `outline`。 ## 使用 Subagent 依各階段的指示，使用隨附的 `codebase-locator`、`codebase-analyzer`、`codebase-pattern-finder`、`web-search-researcher` 和 `outline-implementer-agent` 的 agent 指示。 <important if="you do not have an Agent() or Task() or other tool for running subagents"> read the subagent instructions from $SKILLBASE/references/subagents/SUBAGENT_NAME.md </important> <important if="you have a tool for running subagents or delegated sessions, but the named subagents are not available"> read the subagent instructions from $SKILLBASE/references/subagents/SUBAGENT_NAME.md and pass them into the subagent with your main prompt, with <instructions> REFERENCE_CONTENTS </instructions> <message> YOUR_MESSAGE ON WHAT TO LOOK FOR OR BUILD </message> </important>
英文原文
--- name: rpi description: Only use when the user explicitly invokes this skill by name. --- # Research, Design, and Implement Help the user research a change, work through its design, and implement it. Follow their requested step, using the phase instructions below. `$SKILLBASE` means the directory containing this `SKILL.md`. `<rpi-invocation>` in templates means the host's skill command: `$rpi` in Codex, `/rpi` in Claude Code, or its plugin-qualified name when the host exposes one. Render next commands with that native invocation. ## Choose the Step Follow the user's requested step. Read `$SKILLBASE/references/workflow.md` when choosing or explaining a flow or what comes next. For a task description with no requested step, start with research questions. Ask which flow the user wants: research followed by design discussion, PRD, or TDD, or research only. Wait for their choice before starting the phase, then carry it into suggested next commands. When asking the user to choose a flow, read `$SKILLBASE/references/workflow-choice.md` and show its box-and-line chart with the question. Create writes a new document; iterate updates an existing document in place using the user's feedback or resumes the workflow for a partially finished document. Load the matching process from `$SKILLBASE/references/<step>/process.md`: | Request | Step | | --- | --- | | Research questions | `create-research-questions` or `iterate-research-questions` | | Research | `create-research` or `iterate-research` | | Design discussion | `create-design-discussion` or `iterate-design-discussion` | | Product requirements / PRD | `create-prd` or `iterate-prd` | | Technical design / TDD | `create-tdd` or `iterate-tdd` | | Implementation outline | `create-structure-outline` or `iterate-structure-outline` | | Implement | `implement-outline` | | Implementation feedback or fixes | `iterate-implementation` | Update an existing document by default; create one when needed or requested. When several match, use the newest and name the chosen file. ## Use the Task Directory Use the directory supplied by the user. Otherwise, look for existing task directories in `.agents/artifacts/` whose names and workflow filenames match the request or current task branch. Reuse an obvious match without asking. If a directory looks close but could belong to a different task, or several directories look plausible, ask the user which to use before writing. Read only documents allowed by the selected phase when checking a match. When no existing directory fits, use `.agents/artifacts/<branch-name>` with `/` replaced by `-` if the branch identifies the task. On `main`, or when no branch identifies the task, choose a short kebab-case slug from the request and use `.agents/artifacts/<task-slug>` without asking for approval. State the selected path and reuse or create the directory as needed. Save at the selected path. If the sandbox blocks writes there, ask for scoped write access or a writable task directory and wait for the user's choice. The RPI framework authors recommend not committing the artifacts. Number new workflow documents using the next available `NN-` prefix in the task directory. Keep the existing path when updating a document. Save supporting mockups and diagrams beside the documents and link them as ordinary local files. Workflow document frontmatter includes `repos`, a list of zero or more repositories with `identifier` and `sha` for each entry. Use `repos: []` when none apply. Quote string values as needed for valid YAML. ## Review the Document After writing or updating a workflow document, read `$SKILLBASE/references/prose.md` and `$SKILLBASE/references/show-me.md`, then review the document against both. Make it shorter, more visual, and easier for a person to digest: cut words and sections that do not earn their place, replace dense prose with the smallest view that makes the point, and keep each visual beside the short text it supports. ## Suggest the Next Step Use the phase's handoff instructions and `$SKILLBASE/references/workflow.md` to choose what comes next. Users run each phase in a separate context window by default. When suggesting a next command, include the task directory and chosen flow. Say they can run it here, after compaction, or in a new session. Use artifact filenames to locate the current phase when asked what's next, then read its document for unfinished work and settled choices. Recognize `prd` or `product-requirements`, `tdd` or `tech-design`, and `structure-outline` or `outline`. ## Use Subagents Use the bundled agent instructions for `codebase-locator`, `codebase-analyzer`, `codebase-pattern-finder`, `web-search-researcher`, and `outline-implementer-agent` as directed by each phase. <important if="you do not have an Agent() or Task() or other tool for running subagents"> read the subagent instructions from $SKILLBASE/references/subagents/SUBAGENT_NAME.md </important> <important if="you have a tool for running subagents or delegated sessions, but the named subagents are not available"> read the subagent instructions from $SKILLBASE/references/subagents/SUBAGENT_NAME.md and pass them into the subagent with your main prompt, with <instructions> REFERENCE_CONTENTS </instructions> <message> YOUR_MESSAGE ON WHAT TO LOOK FOR OR BUILD </message> </important>