ai-dev-tasks: PRD to Task List to One-Subtask-at-a-Time AI Workflow

## What it is / what pain point it solves ai-dev-tasks is a set of 'Markdown workflow templates for AI coding tools.' The specific pain point it solves: handing a large block of requirements directly to AI often results in it changing too many files at once, drifting off course, and being hard to interrupt mid-stream. This workflow breaks development into small steps with checkpoints — first write a PRD (product requirements document) to align on the goal, then break the PRD into a checkable task list, and finally complete only one subtask at a time, pausing after each one for your confirmation before proceeding. ## Why this source is worth using Apache-2.0 licensed, free to use and modify commercially; the approach is simple enough that it needs no CLI installation at all — it's just a few Markdown templates plus an execution discipline, usable with any AI coding tool that supports @file references (Cursor, Claude Code, Windsurf, Amp, etc.). It breaks 'controllable and verifiable' down to a fine grain, which is especially good for people just starting to use AI for feature development who haven't found their rhythm yet. ## How to use it (steps) 1. Get the two templates from the project: create-prd.md and generate-tasks.md. 2. Reference the first with @create-prd.md, describe the feature you want to build, and have AI produce a PRD (it will first ask clarifying questions). 3. Use @your-feature-PRD.md together with @generate-tasks.md to turn the PRD into a parent/child task list with `- [ ]` checkboxes. 4. Have AI start implementing from the first item, strictly following the rule of 'one subtask at a time, stop and wait for your yes before continuing'; mark completed items as `- [x]`, and once all subtasks under a parent are done, run tests and commit. ## When to use it Good for feature development where you want things 'step-by-step, checkable mid-process, and want AI to avoid writing too much on its own initiative'; too heavyweight for scenarios where you just want a one-off small snippet from AI. ## Honest note This project once had a third file, process-task-list.md, describing the task execution protocol, in its early stages; the repository root currently only contains the two workflow files create-prd.md and generate-tasks.md — but the execution discipline of 'one subtask at a time, pause for confirmation, mark with [x], test and commit once the parent task is done' remains its core spirit, and has been organized into the full_prompt above for you to use directly. Source: snarktank/ai-dev-tasks (Apache-2.0 License) — this content has been rewritten and reorganized; see the source link for the original.

填下面的欄位，上方 prompt 會即時替換 [方括號] 內容。填好後按「複製組好的 prompt」直接丟進工具。

# AI 開發三步流：PRD → 任務 → 逐項實作（改寫自 snarktank/ai-dev-tasks）

## 步驟 1：產生 PRD（產品需求文件）
你是一位產品工程助理。針對我描述的功能，先別寫程式，請產出一份 PRD，至少包含：目標與動機、目標使用者、功能性需求（編號條列）、明確的非目標（不做什麼）、驗收標準、開放問題。產出前若有不清楚之處，先列問題問我。
我要建的功能：{{FEATURE_DESCRIPTION}}
技術 / 環境背景：{{TECH_CONTEXT}}

## 步驟 2：把 PRD 拆成任務清單
根據上面確認好的 PRD，產出一份 Markdown 任務清單：
- 先列出「父任務」（高層步驟），等我說「Go」再展開子任務。
- 每個任務都用可勾選框 `- [ ]` 表示，採父任務／子任務兩層結構並編號（1.0 / 1.1 / 1.2 …）。
- 最後附一個「相關檔案（Relevant Files）」區塊，列出預計新增或修改的檔案與用途。

## 步驟 3：逐項實作（核心紀律）
依任務清單實作，但嚴格遵守：
1. **一次只做一個子任務**，做完就停下來，回報你做了什麼，然後等我回覆「yes / y / 繼續」才能動下一個子任務。
2. 完成一個子任務後，把它從 `- [ ]` 改成 `- [x]`。
3. 當某父任務底下所有子任務都完成時：先跑測試／驗證，確認沒壞，再更新「相關檔案」區塊，然後才把該父任務也標記為 `- [x]`，並做一次有意義的提交（commit message 講清楚這步做了什麼）。
4. 全程保持任務清單檔案與實際進度同步，禁止一口氣連做多個子任務。

---
📎 來源：snarktank/ai-dev-tasks（Apache-2.0 授權）— 本篇內容經改寫整理。原專案提供 create-prd.md 與 generate-tasks.md 兩份範本，搭配上述「逐子任務、做完等確認」的執行紀律；實際用法是在 Cursor / Claude Code 等工具中以 @檔名 引用這些範本。