manifest-generator（清单生成器）

ManifestGenerator

Given a project description, produce two files in the current directory:

1. PROJECT_MANIFEST.md — structured project context every factory agent reads
2. SOFTWARE_FACTORY_MANIFEST.md — the 6-agent SDLC pipeline blueprint

Execute each phase in order. Do not skip phases. Do not proceed past a STOP until you have the required input.

FILE_WRITE_GUARD

Call this procedure before writing either output file. Substitute the actual filename.

FILE_WRITE_GUARD(filename):
  Check if filename exists in the current directory (use Read tool).
  If it exists:
    Ask the user: "filename already exists. Overwrite, rename to filename.bak, or abort?"
    Overwrite → proceed.
    Rename → rename the existing file to filename.bak, then proceed.
    Abort → stop execution entirely.
  If it does not exist → proceed to write.

Phase 0: Resume Detection

Before anything else, check for existing output files (use Read tool — expect a "file not found" error if they don't exist, which is the normal case):

1. Does PROJECT_MANIFEST.md exist in the current directory?
2. Does SOFTWARE_FACTORY_MANIFEST.md exist in the current directory?

Both missing: Proceed to Phase 1.

PROJECT_MANIFEST.md exists, SOFTWARE_FACTORY_MANIFEST.md does NOT: Read PROJECT_MANIFEST.md into context. Skip Phases 1–5. Tell the user: "Found existing PROJECT_MANIFEST.md — skipping to factory manifest generation." Jump to Phase 6.

Both exist: Ask the user: "Both output files already exist. Regenerate from scratch, or abort?" Regenerate → proceed to Phase 1 (FILE_WRITE_GUARD will handle overwrites). Abort → stop.

Phase 1: Accept Input

Runtime detection probe: Attempt to call AskUserQuestion with the question below. If it succeeds → interactive mode: use AskUserQuestion one-at-a-time for all remaining phases. If it fails (tool not found) → text mode: for all remaining phases, print pending questions as a numbered list, wait for the user to answer, then continue.

Input check (first question / probe):

Check the user's current message or invocation arguments:

• Contains a file path (e.g., PROJECT_OVERVIEW.md): use the Read tool to read it.

- File not found → tell the user, ask them to paste the description directly. STOP. - File found but appears non-text (JSON, binary, package.json, etc.) → tell the user "That file doesn't look like a project description." Ask for inline text. STOP. - File found and text → use it as the description.

• Contains a project description (multiple sentences about what's being built): use it.
• Neither → ask: "Paste your project description, or describe what you're building. Include what it does, who uses it, the tech stack if chosen, and any constraints."

STOP. Do not proceed until a description is in context.

Phase 2: Extraction

Map the description against the 11 PROJECT_MANIFEST sections. For each, assign:

• EXTRACTED — specific, project-named content present (real technology names, actual user roles, concrete constraints, named entities)
• PARTIAL — some content present but incomplete or vague
• MISSING — no relevant content found

Use the WORKED EXAMPLE (at the bottom of this file, between the HTML comment markers) to calibrate what "EXTRACTED" looks like per section. Do not use example content as output.

Always-ask sections — ask if PARTIAL or MISSING, never default:

Section 1: Overview
Section 2: Tech Stack
Section 4: Domain Model  ← most important; never generate a default
Section 6: Constraints
Section 9: Success Criteria

Defaultable sections — generate from context if MISSING, with label:

Section 3: Project Structure
  → infer from tech stack
  → label: (proposed — update when scaffolded)

Section 5: Conventions
  → infer from tech stack (e.g. Node.js → kebab-case files, *.test.ts,
    conventional commits; Python → snake_case, test_*.py, etc.)
  → label: (default — update when scaffolded)

Section 7: Task Inputs
  → derive from Domain Model + fixed pipeline sequence
  → label: (pipeline-critical — verify before running factory)
  → note: weak Domain Model = weak Task Inputs; this breaks pipeline sequencing

Section 8: Services to Connect
  → empty table if not mentioned

Section 10: Review Standards
  → generic rules appropriate to the detected tech stack
  → label: (default — customize for this project)

Section 11: Release Criteria
  → generic checklist appropriate to the project type
  → label: (default — customize for this project)

After mapping, identify which always-ask sections are PARTIAL or MISSING. Proceed to Phase 3.

Phase 3: Gap-Fill

For each always-ask section that is PARTIAL or MISSING, ask one targeted question. Ask in this order: Overview → Tech Stack → Domain Model → Constraints → Success Criteria.

Interactive mode: AskUserQuestion, one question per call. Wait for answer before asking the next.

Text mode: Print all pending questions as a numbered list. Wait for the user to answer all before continuing.

Never ask about a section that is already EXTRACTED.

Domain Model special rule: If the Domain Model answer contains no entity names (a single vague sentence like "it manages data" or "it has users"), ask one follow-up: "Can you name the core data entities? For example: Order, Customer, Staff — the main 'things' your system tracks and stores." Accept the second answer regardless. If still no entity names, mark FILLED but add label (weak — strengthen Section 4 before running factory).

STOP. Do not proceed until all always-ask sections are FILLED.

Phase 4: Validate

Check all 11 sections. A section is still a placeholder if it:

• Is empty
• Contains unedited template text (e.g., [Core entities, their fields...])
• Contains only generic content applicable to any project

For any defaultable section still a placeholder: generate a reasonable default and label it (generated — verify before use).

For any always-ask section still a placeholder: this should not happen (Phase 3 should have caught it). If it somehow did, ask one targeted question now.

Proceed to Phase 5.

Phase 5: Write PROJECT_MANIFEST.md

Call FILE_WRITE_GUARD("PROJECT_MANIFEST.md").

Write PROJECT_MANIFEST.md to the current directory using the Write tool. Use the PROJECT_MANIFEST template defined in the Schema section below. Fill every section. No section may remain empty or contain template placeholder text.

Report: "PROJECT_MANIFEST.md written. N/11 sections extracted from description, M filled via Q&A, K generated as defaults."

Phase 6: Generate SOFTWARE_FACTORY_MANIFEST

Use the PROJECT_MANIFEST content already in context (generated in Phase 5, or read in Phase 0 resume). Do not re-read from disk.

Fill the 7 SOFTWARE_FACTORY_MANIFEST sections:

Section 1 — Factory Overview One paragraph: project name + "This factory runs a 6-agent sequential pipeline..." + tech stack summary drawn from the project manifest.

Section 2 — Pipeline Sequence Fixed 6-agent sequence. Use the project name (lowercased, hyphenated) as:

1. Planner
   Reads:  feature request + PROJECT_MANIFEST.md
   Writes: work-packages/<slug>.md

2. Architect
   Reads:  Planner work package + Tech Stack section of PROJECT_MANIFEST.md
   Writes: docs/adr/NNNN-<slug>.md

3. Designer
   Reads:  Architect ADR + Domain Model section of PROJECT_MANIFEST.md
   Writes: design/<slug>-spec.md

4. Coder
   Reads:  Designer spec + Conventions section of PROJECT_MANIFEST.md
   Writes: src/ on feature branch <slug>-<feature>

5. Reviewer
   Reads:  code diff + Review Standards section of PROJECT_MANIFEST.md
   Writes: review-reports/<slug>-review.md

6. Deployer
   Reads:  Reviewer report + Release Criteria section of PROJECT_MANIFEST.md
   Writes: release-gates/<slug>-gate.md

Section 3 — Human Gates

Gate 1: After Architect — human approves ADR before Designer runs.
Gate 2: After Reviewer — human approves review report before Deployer runs.

Section 4 — Per-Agent System Prompt Seeds Use this template for each of the 6 agents:

"You are the {Agent} for {ProjectName}. You {role description} using
{relevant manifest sections} in PROJECT_MANIFEST.md."

Fill in:

• {ProjectName}: from Section 1 (Overview) of the project manifest
• {Agent}: Planner / Architect / Designer / Coder / Reviewer / Deployer
• {role description}: appropriate to the agent's function
• {relevant manifest sections}: the sections that agent reads

Each seed MUST contain the project name AND at least one domain entity name from Section 4 (Domain Model) of the project manifest. If the domain model is weak (labeled "weak"), add to each seed: (update with domain entities once Section 4 is strengthened).

Example seeds for "Fired Up Pizza":

Planner:   "You are the Planner for Fired Up Pizza. You decompose feature
           requests into work packages using the Domain Model and Tech Stack
           in PROJECT_MANIFEST.md."

Architect: "You are the Architect for Fired Up Pizza. You write architectural
           decision records using the Tech Stack and Constraints in
           PROJECT_MANIFEST.md."

Designer:  "You are the Designer for Fired Up Pizza. You write UX specs and
           interaction designs using the Domain Model and Conventions in
           PROJECT_MANIFEST.md."

Coder:     "You are the Coder for Fired Up Pizza. You implement features
           following the Conventions and Task Inputs in PROJECT_MANIFEST.md."

Reviewer:  "You are the Reviewer for Fired Up Pizza. You enforce the Review
           Standards in PROJECT_MANIFEST.md against every code diff."

Deployer:  "You are the Deployer for Fired Up Pizza. You gate releases against
           the Release Criteria in PROJECT_MANIFEST.md."

Section 5 — Quality Gates Copy the Review Standards and Release Criteria sections from the project manifest. Organize as: "Stage N passes when: [criteria]" for each of the 6 stages.

Section 6 — Orchestrator Configuration

Coordination pattern: sequential pipeline with handoffs
Failure handling: stop pipeline at failing agent, surface error to human
Retry policy: no automatic retries (human decides whether to re-run)
Branch strategy: feature branch per work item, merge after Deployer gate passes

Section 7 — Conventions Reference Copy verbatim from Section 5 (Conventions) of the project manifest.

Phase 7: Validate and Write

Validation checks before writing. Fix any failure by regenerating the section using in-context project manifest content.

1. All 6 agents in Pipeline Sequence have non-empty Reads and Writes — PASS or fix.
2. Gate 1 is after Architect, Gate 2 is after Reviewer (not Deployer) — PASS or fix.
3. All 6 seeds contain the project name AND at least one domain entity — PASS or fix.
4. Quality Gates section is non-empty and non-generic — PASS or fix.

Call FILE_WRITE_GUARD("SOFTWARE_FACTORY_MANIFEST.md").

Write SOFTWARE_FACTORY_MANIFEST.md to the current directory.

Report: "SOFTWARE_FACTORY_MANIFEST.md written and validated. Both manifests ready — your 6-agent factory has context."

PROJECT_MANIFEST Schema

Write the output file with exactly this structure. Fill every section.

# Project Manifest: [Project Name]

## Overview

[One paragraph: what the software does and who uses it.]

## Tech Stack

| Layer    | Technology | Notes |
|----------|-----------|-------|
| Frontend |           |       |
| Styling  |           |       |
| State    |           |       |
| Routing  |           |       |
| Backend  |           |       |
| Database |           |       |
| Testing  |           |       |
| Linting  |           |       |

## Project Structure

[Top-level directory tree, annotated. Label (proposed) if pre-code.]

## Domain Model

[Core entities, their fields, and their relationships.]

## Conventions

- File naming:
- Test files:
- API routes:
- Commits:
- Branches:

## Constraints

- [Explicit out-of-scope items or disallowed approaches.]

---

## Task Inputs

| Agent     | Receives                  | From                     |
|-----------|--------------------------|--------------------------|
| Planner   |                          |                          |
| Architect |                          |                          |
| Designer  |                          |                          |
| Coder     |                          |                          |
| Reviewer  |                          |                          |
| Deployer  |                          |                          |

## Services to Connect

| Service | Purpose | Config |
|---------|---------|--------|
|         |         |        |

## Success Criteria

### Per-Feature Success

- [ ]
- [ ]

### Factory-Level Success

- [ ]
- [ ]

---

## Review Standards

### Spec Compliance

- [Rules the Reviewer enforces against the Designer's spec.]

### Style

- [Project-specific style rules.]

### Security

- [Project-specific security rules.]

### Severity Scale

- **Low**: cosmetic issues, minor inconsistencies
- **Medium**: functional gaps, missing edge cases
- **High**: data loss, security vulnerability, spec violation

---

## Release Criteria

### Required (all must PASS)

1. [ ]
2. [ ]

### Informational (reported but non-blocking)

- [metric 1]

SOFTWARE_FACTORY_MANIFEST Schema

Write the output file with exactly this structure.

# Software Factory Manifest: [Project Name]

## Factory Overview

[Project name. This factory runs a 6-agent sequential pipeline (Planner →
Architect → Designer → Coder → Reviewer → Deployer) with two human gates.
Tech stack: [summary from PROJECT_MANIFEST.md].]

## Pipeline Sequence

1. **Planner**
   - Reads: feature request + PROJECT_MANIFEST.md
   - Writes: work-packages/[slug].md

2. **Architect**
   - Reads: Planner work package + Tech Stack section
   - Writes: docs/adr/NNNN-[slug].md

3. **Designer**
   - Reads: Architect ADR + Domain Model section
   - Writes: design/[slug]-spec.md

4. **Coder**
   - Reads: Designer spec + Conventions section
   - Writes: src/ on feature branch [slug]-[feature]

5. **Reviewer**
   - Reads: code diff + Review Standards section
   - Writes: review-reports/[slug]-review.md

6. **Deployer**
   - Reads: Reviewer report + Release Criteria section
   - Writes: release-gates/[slug]-gate.md

## Human Gates

- **Gate 1 — After Architect:** Human approves ADR before Designer runs.
- **Gate 2 — After Reviewer:** Human approves review report before Deployer runs.

## Per-Agent System Prompt Seeds

**Planner:** "[seed]"

**Architect:** "[seed]"

**Designer:** "[seed]"

**Coder:** "[seed]"

**Reviewer:** "[seed]"

**Deployer:** "[seed]"

## Quality Gates

[Per-stage pass criteria drawn from Review Standards and Release Criteria in
PROJECT_MANIFEST.md. Format: "Stage N (AgentName) passes when: [criteria]".]

## Orchestrator Configuration

- Coordination pattern: sequential pipeline with handoffs
- Failure handling: stop pipeline at failing agent, surface error to human
- Retry policy: no automatic retries (human decides whether to re-run)
- Branch strategy: feature branch per work item, merge after Deployer gate passes

## Conventions Reference

[Verbatim copy of Conventions section from PROJECT_MANIFEST.md.]

Worked Example: Fired Up Pizza

Input description

Fired Up Pizza is a web app for a small neighborhood pizza restaurant. Customers browse the menu, customize pizzas (size, crust, toppings), place an order against their phone number, and track it through "placed → preparing → ready → delivered." Staff use the same app to manage menu items and advance order status as pizzas move through the kitchen. It replaces a phone-only workflow that loses orders and leaves customers in the dark.

Goals: customers can place and track an order without a phone call; staff can see every live order in one view and update status in one click; runs on a single machine with npm install && npm run dev — no cloud setup.

Stack: React 18 + TypeScript + Vite; Tailwind; Node.js + Express REST API; SQLite via better-sqlite3; Vitest + React Testing Library.

Out of scope: online payments (pay at counter), SMS/push notifications, multi-location support, loyalty accounts.

Constraints: no external auth (phone-number identity), no external DB server, no real-time infra — polling is acceptable.

Users: Customer (places and tracks orders, identified by phone number); Staff (views live order queue, advances order status, edits menu items).

What "EXTRACTED" looks like (annotated excerpts)

Section 1 — Overview (EXTRACTED): Fired Up Pizza is a web app for a small neighborhood restaurant that lets customers browse the menu, customize and place orders by phone number, and track status through placed → preparing → ready → delivered. Staff view the live order queue and advance status. Replaces a phone-only workflow. [Good: project-named, specific actors, specific workflow described.]

Section 2 — Tech Stack (EXTRACTED): Frontend: React 18 + TypeScript + Vite | Styling: Tailwind CSS | Backend: Node.js + Express (REST) | Database: SQLite via better-sqlite3 (single-machine, no external server) | Testing: Vitest + React Testing Library [Good: named technologies with versions where available, notes on constraints.]

Section 4 — Domain Model (EXTRACTED): MenuItem: id, name, description, base_price, available (boolean) Order: id, phone_number, status (placed|preparing|ready|delivered), created_at OrderItem: id, order_id, menu_item_id, size, crust, toppings[], unit_price Customer: identified by phone_number (no account entity) Staff: role only, no separate entity needed [Good: named entities, key fields, relationships implied. This is what "EXTRACTED" looks like.]

Section 4 — Domain Model (PARTIAL): "The app manages orders and menu items for customers and staff." [Bad: entity names present (orders, menu items, customers, staff) but no fields, no relationships, no statuses. Would trigger a follow-up question.]

Section 6 — Constraints (EXTRACTED):

• No online payments (pay at counter only)
• No SMS or push notifications
• No multi-location support
• No external auth — identity is phone number only
• No external DB server — SQLite, single machine
• Polling acceptable (no real-time infra needed) [Good: explicit, enumerated, actionable.]

Section 9 — Success Criteria (EXTRACTED): Per-Feature: customer places order without phone call; customer tracks through 4 status stages; staff sees all live orders in one view; staff advances status in one click; staff edits menu items. Factory-Level: app starts with npm install && npm run dev; all tests pass on clean checkout. [Good: specific, testable, includes factory-level criteria.]