PRD Engine

Agent-executable Product Requirements Documents, generated by a 3-phase pipeline of parallel research agents, synthesis, quality scoring, and adversarial review.

PRD Engine ships as a Claude Code plugin and as a drop-in skill bundle. Same content, two install paths. Use whichever fits your workflow.

---

What it does

You describe a feature in plain English. The engine:

1. Spawns 3-8 parallel research agents (codebase analysis, domain research, requirements, risks, implementation strategy).
2. Hands their reports to a synthesizer that fills a structured template with before/after code blocks for every requirement, verification tables, acceptance criteria, and explicit "what NOT to change" lists.
3. Runs a quality reviewer (0-100 rubric) plus 5 adversarial experts in parallel (blast radius, data integrity, UX regression, code patterns, scope) and decides: deliver, revise, rework, or escalate.

Output: a single PRD.md file at .prd-engine/{slug}/PRD.md that another agent can implement directly without follow-up questions.

---

Why this exists

Most PRDs read fine to humans but fail when you hand them to an autonomous coding agent. They reference files that no longer exist, claim BEFORE code that doesn't match the actual source, leave acceptance criteria vague ("works correctly"), and skip the negative space ("don't change these 12 things").

PRD Engine enforces the structural rules that make a PRD agent-executable, then runs adversarial review to catch the gaps a single drafting pass misses.

---

Install

As a Claude Code plugin

# In any Claude Code session
/plugin marketplace add addiplus/prd-engine
/plugin install prd-engine@addiplus

This auto-discovers the skill at skills/prd-engine/SKILL.md and the 10 agents at agents/prd/.

As a drop-in skill (no plugin system required)

# Clone or download this repo, then:
cp -r skills/prd-engine ~/.claude/skills/
cp -r agents/prd ~/.claude/agents/

Both install paths use the same content. The plugin path adds versioning and marketplace integration; the drop-in path is just files.

---

Quick start

You: "Write a PRD for adding dark mode to the dashboard"
You: "PRD for the payment integration API --deep"
You: "Quick PRD for fixing the header alignment bug"

The engine confirms the plan (agents, intensity, time estimate), waits for "go", then runs to completion.

---

See it without running it

demo/ holds one complete worked example, committed as files: a feature request going in, the five research reports, the validator and quality-gate reports, all five adversarial expert reports, and the delivered PRD coming out. Start at demo/README.md, which maps every file to the pipeline stage that produces it.

---

Intensity tiers

Tier	Total agents	Research	Quality + Adversarial	Time	Use case
quick	4	3 agents	none	5-8 min	Small bug fix, internal tool
standard (default)	13	5 agents	1 reviewer + 5 adversarial	12-18 min	Most features
deep	15	6 agents	1 reviewer + 1 contrarian + 5 adversarial	20-28 min	Cross-cutting, external-facing
exhaustive	19	8 agents	2 synthesizers + scorer + 1 reviewer + 1 contrarian + 5 adversarial	30-45 min	Mission-critical, zero ambiguity

Natural-language detection: "quick PRD" → quick. "thorough" / "deep" → deep. "exhaustive" / "production" → exhaustive. No signal → standard.

---

The 10 agents

Agent	Role
prd-engine-synthesizer	Composes PRD from research reports
prd-reviewer	Scores draft on 5-dimension rubric (0-100)
prd-blast-radius-analyst	Traces downstream consumer breakage
prd-data-integrity-adversary	Corruption, race, parsing-overflow attacks
prd-ux-regression-hunter	Layout, dark mode, discoverability gaps
prd-code-pattern-enforcer	BEFORE/AFTER block accuracy, style drift
prd-scope-critic	Overengineering and cross-PRD conflicts
adversarial-calibrator	Rates adversarial report quality
adversarial-heatmap-generator	Agent × PRD severity matrix
finding-deduplicator	Groups findings by semantic similarity

All agents run on Opus with extended context.

---

Score-to-action mapping

Score	Action
80-100	DELIVER as-is. Copy draft to final.
60-79	REVISE. Apply reviewer's top 3 suggestions, deliver.
40-59	REWORK. Prompt user with top issues; offer re-synthesis.
0-39	ESCALATE. Fundamental gaps; review raw research.

---

What gets enforced

• Every requirement has BEFORE/AFTER code blocks (no exceptions)
• Every acceptance criterion has a specific expected value (no "works correctly")
• The "What NOT to Change" section has ≥10 items
• The PRD is self-contained (no external file references)
• File count in metadata matches files in body
• Length budget: 200 + (requirements × 60), capped at 600 lines

---

Output layout

.prd-engine/{slug}/
  config.json
  research/
    01_codebase_analysis.md
    02_domain_research.md
    03_requirements.md
    04_risks_and_edges.md
    05_implementation_strategy.md
    09_validation.md
    10a_quality_review.md
    10b_contrarian.md           # deep+ only
    expert_1_blast_radius.md
    expert_2_data_integrity.md
    expert_3_ux_regression.md
    expert_4_code_patterns.md
    expert_5_scope_critic.md
  PRD_DRAFT.md
  PRD.md

---

Configuration

Flag	Values	Default
--intensity / -i	quick, standard, deep, exhaustive	standard
--number / -n	Integer	Auto-increment from highest existing PRD
--output / -o	Directory path	.prd-engine/{slug}/
--clarify	Flag	Enabled for deep+, disabled for quick/standard
--no-questions	Flag	Skip Socratic clarification
--target	Project path	Repository root
--skip-quality	Flag	Skip quality gate (not recommended)

---

Validation

Every push and pull request runs a validation job (.github/workflows/validate.yml) that checks this repo is still well formed. It is a structural check only. Nothing in it calls a model, spawns an agent, or generates a PRD.

The job runs .github/scripts/validate.py, which confirms:

• .claude-plugin/plugin.json and .claude-plugin/marketplace.json parse as JSON and carry the fields Claude Code needs, including a plugins[0].name that matches the plugin name so the documented install resolves.
• Every agent definition under agents/ and every skills/**/SKILL.md opens with YAML frontmatter that parses and declares a non-empty name and description.
• No Markdown or JSON file contains an em dash, en dash, or related long dash, which keeps the prose in plain ASCII punctuation.

It needs the Python standard library plus PyYAML. Run it locally the same way CI does:

pip install pyyaml
python .github/scripts/validate.py

---

License

MIT. See LICENSE.

Author

Built by Joe Smith. Companion repos:

• addiplus/mcp-cookie-auth-reference: MCP server reference for cookie-authenticated systems
• addiplus/claude-skill-pack: Bun-native Claude Code dev loop (4 skills)
• addiplus/mcp-builder: skill for building MCP servers with dual transport + OAuth 2.1