Skip to content
/

prd-creator-skill

huchi996-prd-creator · huchi996/prd-creator-skill · ★ 17 · last commit 2026-02-17

Primitive shape 1 total
Skills 1
00

Summary

prd-creator-skill — Summary

prd-creator-skill is a single SKILL.md file (v2.0.0) that implements a 6-phase AI-driven SOP for creating hierarchically structured Product Requirement Documents in Chinese-language enterprise workflows. Phases run from 需求孵化 (Requirement Incubation — multi-round Q&A to clarify vague ideas) through 方案预研 (Solution Research — AI web search for best practices), 项目构建 (Project Construction — directory + README), 模块构建 (Module Construction — module-level PRD), 功能细化 (Feature Refinement — function-level PRD), to 多角色评审 (Multi-Role Review — simulated test/dev/product persona review). The key structural innovation is a three-tier document hierarchy: Project (P001-xxx/) → Module (M001-xxx.md) → Function (F001-xxx.md), with each level carrying its own Mermaid architecture diagrams rendered via the beautiful-mermaid library. At 17 stars, it is a community skill targeting Chinese-speaking product managers and developers who want LLM-assisted PRD authoring within their IDE. Compared to seeds, it most resembles the spec-kit approach of funneling a natural language idea through a structured clarification sequence into a spec artifact, but goes deeper on the document hierarchy and adds a multi-role review simulation absent from any seed.

01

Overview

prd-creator-skill — Overview

Origin

Created by huchi996. Published on GitHub as huchi996/prd-creator-skill. Version 2.0.0, pushed 2026-02-17. MIT license. The skill follows agentskills.io format (SKILL.md frontmatter with name, version, description). The v2.0.0 release notes are in RELEASE-v2.0.0.md.

Philosophy

The skill targets the common enterprise PM workflow where product requirements start as vague verbal descriptions ("我想做一个用户中心平台") and must be refined into structured multi-level specifications before development can begin. The 6-phase SOP is modeled on standard Chinese enterprise SDLC methodology (需求孵化 → 开发). Every phase produces a tangible artifact in a structured file tree rather than a single monolithic document.

Key design decisions:

  1. Three-tier hierarchy — P (project) → M (module) → F (function) — matches enterprise-scale product breakdown
  2. Multi-role review simulation — Phase 6 simulates test engineer, developer, and product manager review personas, catching specification gaps before development
  3. Mermaid diagrams as first-class artifacts — every PRD level includes architecture diagrams rendered by beautiful-mermaid
  4. AI web search in Phase 2 — uses SearchWeb to gather competitive analysis and technical best practices before writing specs

Manifesto-style quotes

From SKILL.md description field:

"Create comprehensive Product Requirement Documents (PRD) using an AI-driven 6-phase SOP workflow (需求孵化 -> 方案预研 -> 项目构建 -> 模块构建 -> 功能细化 -> 多角色评审) with multi-role review (测试、研发、产品等角色评审)."

Phase 6 rationale from SKILL.md:

"Based on review findings, update the module and function specs. Fix specification ambiguities. Resolve technical feasibility concerns. Address test coverage gaps."

Target audience

Chinese-speaking product managers, architects, and developers working in enterprise software shops who want a structured, AI-assisted PRD authoring workflow without learning a complex framework.

02

Architecture

prd-creator-skill — Architecture

Distribution

Single SKILL.md file + assets/examples/references directories. No CLI, no npm package.

Install

Copy SKILL.md into the AI tool's skills directory (e.g., ~/.claude/skills/prd-creator/SKILL.md).

Required runtime

None (just the AI host tool)

Directory tree (as installed)

prd-creator-skill/
├── SKILL.md             # The skill itself (6-phase SOP)
├── README.md
├── CHANGELOG.md
├── RELEASE-v2.0.0.md
├── assets/              # Mermaid diagram assets
├── examples/            # Example PRD outputs
└── references/          # Reference materials

Output directory tree (produced by skill)

prd/
├── P001-用户中心平台/           # Project directory
│   ├── README.md               # Project overview (scope, goals, roadmap)
│   ├── M001-用户认证.md        # Module document with Mermaid architecture
│   ├── M001/                   # Function subdirectory
│   │   ├── F001-注册登录.md
│   │   ├── F002-密码找回.md
│   │   └── F003-第三方登录.md
│   ├── M002-账户管理.md
│   └── M002/
│       ├── F001-资料修改.md
│       └── F002-账号注销.md
├── P002-xxx/
└── ...

Naming convention

  • Project: P{三位编号}-{项目名}/ — e.g., P001-用户中心平台/
  • Module: P{编号}/M{三位编号}-{模块名}.md — e.g., M001-用户认证.md
  • Function: P{编号}/M{编号}/F{三位编号}-{功能名}.md — e.g., F001-注册登录.md

Target AI tools

Any agentskills.io-compatible tool with web search capability (Phase 2 uses SearchWeb). Claude Code is the primary target.

External dependencies

  • beautiful-mermaid library for Mermaid SVG rendering (referenced but not bundled)
  • SearchWeb capability in the AI host (for Phase 2 competitive research)
03

Components

prd-creator-skill — Components

Single Skill

Component Type Purpose
SKILL.md skill-md 6-phase PRD creation SOP (the entire framework)

6 Phases within the Skill

Phase Name (Chinese) Name (English) Output
1 需求孵化 Requirement Incubation 《需求描述文档》 — structured requirement description
2 方案预研 Solution Research 《技术预研报告》 + module list
3 项目构建 Project Construction P001/ directory + project README
4 模块构建 Module Construction P001/M001-xxx.md + architecture diagram
5 功能细化 Feature Refinement P001/M001/F001-xxx.md + flow diagram
6 多角色评审 Multi-Role Review 《评审报告》 + change list

Templates

Template Content
Project README Background, goals, scope boundary, module list, dependency graph
Module document Business value, user stories, architecture diagram (Mermaid), API specs, data model, test strategy
Function document Feature description, business rules, UI flow (Mermaid), edge cases, acceptance criteria
Review report Role-specific findings (test, dev, product), severity classification, change requirements

Role Simulations (Phase 6)

The skill instructs the AI to simulate:

  • 测试工程师 (Test Engineer): Checks testability, edge cases, coverage gaps
  • 研发工程师 (Development Engineer): Checks technical feasibility, API completeness, performance concerns
  • 产品经理 (Product Manager): Checks business alignment, user experience, priority clarity

External Tool Call

Phase 2 explicitly calls SearchWeb to gather competitive analysis and technical best practices — the only skill in this batch that makes web search a first-class phase.

05

Prompts

prd-creator-skill — Prompts

Excerpt 1: Phase 1 — Requirement Incubation Clarifying Questions

Source: SKILL.md Phase 1 section

### Key Questions to Ask
- 目标用户是谁?有哪些使用场景?
- 核心功能点有哪些?优先级如何?
- 有什么必须排除在范围外的功能?
- 预期上线时间或里程碑?

Technique: Minimal structured Q&A — four fixed clarifying questions that cover user, scope, exclusions, and timeline. Unlike blueprint-imbue's iterative multi-round Q&A, this fires all four questions in a single batch to define the requirement boundary quickly.


Excerpt 2: Phase 6 — Multi-Role Review Workflow

Source: SKILL.md Phase 6 section

## Phase 6: 多角色评审 (Multi-Role Review)

### Goal
通过多角色视角评审,确保PRD质量和可实施性。

### Review Roles
1. **测试工程师视角** - 评审可测性、边界条件、缺少的测试场景
2. **研发工程师视角** - 评审技术可行性、API完整性、性能风险
3. **产品经理视角** - 评审业务价值、用户体验、优先级合理性

### Review Process
For each role:
1. Read ALL module and function documents
2. Identify issues from this role's perspective
3. Classify severity: Critical/Major/Minor
4. Document specific change requirements

### Output: 《评审报告》
| 角色 | 问题 | 严重程度 | 修改要求 |
|------|------|----------|----------|

### Decision Gate
- No Critical/Major issues → proceed to development
- Issues found → update specs in Phase 5 (or Phase 4 if structural)

Technique: Persona-switching review simulation. The AI adopts three distinct reviewer roles sequentially, evaluating the same documents from different professional perspectives. This is a structured adversarial review without spawning separate agents.


Excerpt 3: Phase 2 — SearchWeb Integration

Source: SKILL.md Phase 2 section

### Steps
1. **联网搜索** - SearchWeb 搜索相关技术方案、竞品分析、开源解决方案
2. **方案对比** - 提供主流架构对比、推荐技术栈、潜在风险分析
3. **模块拆分** - 基于调研结果,生成模块清单(M001, M002...)

Technique: Web research as a first-class phase gate. Before writing a single line of spec, the AI gathers external evidence (competitive landscape, technical options) to ground the specification in real-world constraints.

09

Uniqueness

prd-creator-skill — Uniqueness

Differs from Seeds

Closest seed is spec-kit — both convert a natural language idea into a structured spec through a clarification loop. prd-creator diverges in three ways: (1) it adds a three-tier document hierarchy (Project → Module → Function) whereas spec-kit produces a single flat spec file; (2) it includes a Phase 2 competitive research step using SearchWeb before writing any spec, which no seed framework does; (3) it simulates multi-role review in Phase 6 (test/dev/product personas) as a specification quality gate, not a code review gate. Compared to openspec (which also targets full SDLC), prd-creator is narrower — it covers only the requirements authoring phase, not implementation or task management. It is the only Chinese-language skill in this batch and appears to target enterprise PM workflows in the Chinese tech industry.

Positioning

Occupies the "idea-to-structured-PRD" slot with unusual depth: most PRD tools stop at a single document, whereas prd-creator enforces hierarchical decomposition all the way to individual function files. The multi-role review simulation is the most distinctive feature — it operationalizes structured spec review without requiring separate review agents.

Observable Failure Modes

  1. SearchWeb dependency in Phase 2: If the AI host lacks web search capability, Phase 2 silently falls back to training-data knowledge. The spec may miss real competitive context.
  2. Persona simulation fidelity: The three-role review in Phase 6 is a single-model role-play, not genuine adversarial agents. A sufficiently aligned LLM may not challenge its own prior work effectively.
  3. Context window stress: Phase 6 reads ALL module and function docs simultaneously. On large projects (5+ modules, 10+ functions each), this can exceed practical context limits.
  4. Chinese-only terminology: The phase names and many output document titles are in Chinese, creating a language barrier for non-Chinese teams.
  5. No enforcement beyond prompting: Nothing prevents skipping Phase 6 review — the skill is guidance only.
04

Workflow

prd-creator-skill — Workflow

Phase Flow

Phase 1: 需求孵化 (Requirement Incubation)
  ↓ User confirms 《需求描述文档》
Phase 2: 方案预研 (Solution Research)
  → SearchWeb for competitive analysis + technical best practices
  ↓ User confirms module list
Phase 3: 项目构建 (Project Construction)
  → Create P001/ directory + README.md
  ↓ Proceed
Phase 4: 模块构建 (Module Construction)
  → Create M001-xxx.md per module with Mermaid architecture diagram
  ↓ Proceed (or iterate per module)
Phase 5: 功能细化 (Feature Refinement)
  → Create F001-xxx.md per function with Mermaid flow diagram
  ↓ Proceed
Phase 6: 多角色评审 (Multi-Role Review)
  → Simulate test/dev/product review personas
  ↓ Review passed? → enter development
  ↓ Review failed? → back to Phase 5 (or Phase 4 if structural issues)

Artifacts per Phase

Phase Artifact
1 《需求描述文档》 (requirement description doc)
2 《技术预研报告》 (technical research report) + module list
3 prd/P001-xxx/README.md
4 prd/P001/M001-xxx.md (one per module)
5 prd/P001/M001/F001-xxx.md (one per function)
6 《评审报告》 (review report) + change list

Approval Gates

  1. After Phase 1: User confirms the requirement description is accurate
  2. After Phase 2: User confirms module list and architecture direction
  3. After Phase 6: Review panel passes (or iterate)

Iteration Loop

Phase 6 → Phase 5 (or Phase 4) is an explicit iterative loop. The skill instructs that review findings trigger spec updates, not code changes — the PRD must be correct before development starts.

06

Memory Context

prd-creator-skill — Memory & Context

State Storage

File-based. All outputs are written to the prd/ directory under the project root:

  • prd/P001-xxx/README.md
  • prd/P001/M001-xxx.md
  • prd/P001/M001/F001-xxx.md

Persistence

Project-level. The three-tier file tree survives session boundaries and serves as the context source for subsequent AI interactions.

Cross-Session Handoff

Implicit — the prd/ directory tree is the handoff artifact. Any subsequent agent session can read the hierarchical PRD files to resume where the previous session left off.

Context Window Management

No explicit compaction handling. The skill instructs the AI to read "ALL module and function documents" during Phase 6 review, which could stress context windows on large projects with many modules.

Memory Type

File-based markdown. No database, no vector search.

Search Mechanism

None (file tree browsing only).

Phase 2 External Memory

Phase 2 uses SearchWeb to pull in external knowledge (competitor analysis, technical options) as ephemeral context that informs module design. This is not persisted beyond the 《技术预研报告》 artifact.

07

Orchestration

prd-creator-skill — Orchestration

Multi-Agent

No. Single agent executes all 6 phases. Phase 6 simulates multiple reviewer personas through prompt-based role-switching, not separate agent spawns.

Orchestration Pattern

Sequential. Phases execute in fixed order: 1 → 2 → 3 → 4 → 5 → 6, with an optional back-loop from 6 → 5 on review failure.

Isolation Mechanism

None. All writes go to the prd/ directory in the current working directory.

Multi-Model

No. Single model handles all phases including the role-simulation in Phase 6.

Execution Mode

Interactive-loop with approval gates at Phase 1 (requirement confirmation) and Phase 2 (module list confirmation).

Crash Recovery

Implicit — the file-based output tree means the AI can read existing prd/ artifacts and resume from the appropriate phase on restart. No explicit state machine.

Context Compaction

No explicit handling. Phase 6 explicitly requires reading ALL produced documents, which may exceed context on large specs.

Subagent Definition Format

None.

Prompt Chaining

Yes — the 《需求描述文档》 from Phase 1 is the input to Phase 2's SearchWeb query formation; the module list from Phase 2 drives Phase 3's directory creation; each phase's output is the next phase's primary input.

08

Ui Cli Surface

prd-creator-skill — UI & CLI Surface

CLI Binary

None. This is a single SKILL.md file.

Local UI

None.

IDE Integration

Via SKILL.md in the host AI tool's skills directory. No dedicated IDE plugin.

Invocation

  • Trigger: user says "create a PRD", "write product requirements", "draft a spec", "建立PRD目录结构", "需求评审", etc.
  • The SKILL.md description field contains the trigger phrases for automatic skill activation.

Observability

No audit log, no structured metrics. The prd/ file tree is the only persistent output.

Tool Requirements

Requires the host AI tool to have SearchWeb capability for Phase 2. Without web search, Phase 2 degrades to AI knowledge only (no competitive analysis).

Related frameworks

same archetype · same primary tool · same memory type

BMAD-METHOD ★ 48k

Provides a full agile delivery lifecycle with named expert-persona AI collaborators that elicit the human's best thinking rather…

Agent OS ★ 4.6k

Extracts implicit codebase conventions into token-efficient markdown standards files and injects them selectively into AI agent…

Claude Conductor ★ 367

Gives Claude Code a persistent, cross-linked, auto-analyzed documentation system so it retains codebase context across sessions.

Spec-Driver (Greenfield Spec-Driven Development) ★ 25

Prevents spec rot in AI-assisted development by making implementation changes flow back into evergreen, authoritative specs via…

Anthropic Knowledge Work Plugins ★ 16k

Role-specialized plugin bundles with live MCP connectors that turn Claude into a domain expert for enterprise knowledge workers.

Codex Integration for Claude Code (skill-codex) ★ 1.3k

Single Claude Code skill that handles Codex CLI invocation correctly (stdin blocking, thinking token suppression, session resume)…