User Alignment and Agent-Ready PRDs

User Alignment & Agent-Executable PRDs

Purpose: A practical guide for turning messy user requests into aligned, testable Product Requirements Documents (PRDs) that autonomous or semi-autonomous agents can execute without drifting.

Two supporting files ship with this skill:

• references/prd-template.md — the full copy-paste agent-executable PRD template.
• references/reading-list.md — the research and reference map this guide is built on.

1. Core thesis

A good agent PRD is not just a product document. It is a shared operating contract between the user, the product owner, and the implementation agent.

It must do three jobs at once:

1. Align on user intent — what the user actually wants, why it matters, and what outcome would make them say “yes, that’s it.”
2. Remove ambiguity before execution — especially around scope, constraints, priorities, edge cases, and trade-offs.
3. Translate intent into executable work — sequenced phases, explicit files/systems, acceptance criteria, tests, and “do not do” boundaries.

For human teams, ambiguity can be resolved in meetings. Agents often resolve ambiguity by guessing. The PRD’s job is to make guessing unnecessary.

2. Research-backed principles

2.1 Start with the user problem, not the implementation

AI/product PRDs should begin with the user pain and the cost of the status quo, not with “we will use AI/model/tool X.” The model, framework, or agent is an implementation detail unless the user explicitly constrained it.

A strong problem statement should include:

• Who is affected.
• What they are trying to accomplish.
• What blocks them today.
• Why existing/manual/deterministic solutions are insufficient.
• What measurable improvement would matter.

Bad: “Build an AI assistant for support.”
Better: “Support agents spend 8 minutes triaging each ticket, and 23% are misrouted. We need ticket classification under 2 seconds with at least 92% routing accuracy, while escalating low-confidence cases.”

2.2 Treat the PRD as an executable artifact

Spec-driven development treats the spec as a durable source of truth, not disposable planning scaffolding. The spec should be checked into the repo and referenced by agents across sessions.

The PRD should answer:

• What should be built?
• Why does it matter?
• What is explicitly out of scope?
• What order should work happen in?
• How will each phase be verified?
• What should the agent never touch?

2.3 Write for two audiences: humans first, agents second

The top of the PRD should explain the product and user context in human language. The lower sections should become increasingly operational and literal for the agent.

Recommended split:

2.4 Break work into bounded phases

Traditional PRDs often describe the whole product and leave sequencing to engineers. Agent PRDs should define implementation phases with dependencies and testable outputs.

Each phase needs:

• Dependency: what must already exist.
• Scope: what this phase covers.
• Out of scope: adjacent work the agent must not do yet.
• Tasks: concrete implementation actions.
• Verification: commands, tests, screenshots, or manual checks that prove completion.

2.5 Define evaluation before implementation

For AI or agentic features, “works” is rarely binary. Define launch thresholds, target thresholds, and aspirational thresholds before building.

Include three metric classes:

1. Outcome metrics — user/business result, e.g. completion rate, time saved, conversion lift.
2. Quality metrics — correctness, relevance, tone, completeness, usefulness.
3. Operational metrics — latency, cost, reliability, throughput, escalation rate.

For agent-executed software projects, also define:

• Unit/integration/e2e tests required.
• Manual QA checks.
• Review criteria.
• Regression risks.
• Expected screenshots or artifacts.

2.6 Make constraints and prohibitions explicit

Agents overbuild when boundaries are unclear. Every PRD should include an explicit “do not do” section.

Examples:

• Do not change authentication.
• Do not modify database schema outside the listed migration.
• Do not install new dependencies without approval.
• Do not touch deployment files.
• Do not redesign unrelated UI.
• Do not commit secrets.
• Do not treat P1/P2 items as required for MVP.

2.7 Use semi-formal requirement language when precision matters

For behavioral requirements and acceptance criteria, use a lightweight syntax such as EARS or Given/When/Then.

Useful EARS patterns:

• WHEN [trigger] THEN [system] SHALL [response]
• IF [condition] THEN [system] SHALL [response]
• WHILE [state] [system] SHALL [continuous behavior]
• WHERE [context] [system] SHALL [contextual behavior]

Useful BDD pattern:

Given [context]
When [user action or system event]
Then [observable result]
And [additional observable result]

Avoid vague terms like “fast,” “simple,” “intuitive,” “robust,” or “user-friendly” unless they are backed by measurable thresholds.

2.8 Ingest User Session Replays & Droplet Context

When debugging or implementing fixes from user feedback, agents must not rely solely on text descriptions. They should actively parse available session recordings, screenshots, console error logs, and network payloads captured during the user session.

An agent-executable PRD must instruct the agent on how to:

• Correlate screenshots to specific UI components and coordinates.
• Parse stack traces and console errors to identify the exact files and lines of code responsible.
• Use network request/response payloads to trace API mismatches or data corruptions.

2.9 Leverage Persistent Memory & Preferences

To prevent context drift and avoid re-teaching the agent style preferences (e.g. using arrow functions, specific hook patterns) or workspace constraints across different chat sessions, the project must maintain a persistent memory file (e.g. .agentguard/parcel-memory.json).

The PRD should define:

• How the agent reads this file before every execution step.
• How the agent updates the history/state log in the memory file after completing a phase.
• An explicit rule requiring the agent to align all generated code with these persistent preferences.

3. User alignment workflow

Use this workflow before writing the PRD.

Step 1: Capture the raw request

Write the request down as-is. Do not immediately translate it into features.

## Raw Request
> [Paste the user’s exact words]

Step 2: Extract the intent

Summarize what you believe the user wants in one paragraph.

## Interpreted Intent
The user wants [outcome] for [user/persona] because [problem]. The desired result is [observable success state].

Step 3: Separate facts, assumptions, and unknowns

## Facts
- [Directly stated by user]

## Assumptions
- [Reasonable inference, but not confirmed]

## Unknowns
- [Information needed before implementation]

Risks

• Technical risks that could affect implementation
• Data privacy or security concerns
• Timeline or resource constraints
• External dependencies that may block progress

Identify high-risk items before implementation begins and surface them during alignment. A good agent should act on obvious defaults, but should not hallucinate material requirements. If an unknown changes architecture, cost, privacy, or scope, ask before execution.

Step 4: Ask only high-leverage clarification questions

Do not interrogate the user with twenty questions. Ask the few questions that materially change what gets built.

Use this priority order:

1. Outcome: What does success look like?
2. User: Who is this for?
3. Scope: What is in/out for v1?
4. Constraints: What systems, stack, data, deadline, or policy constraints apply?
5. Failure tolerance: What happens if the agent/system is wrong?
6. Approval: Who needs to sign off?

Step 5: Read back the aligned understanding

Before writing an executable PRD, produce a short alignment read-back:

My understanding:
- We are solving: [problem]
- For: [users]
- Success means: [metrics / user-visible outcome]
- MVP includes: [scope]
- MVP excludes: [non-goals]
- Key constraints: [constraints]
- Open questions: [remaining unknowns]

This catches mismatches early, before the agent turns them into code.

4. Agent-executable PRD template

The full copy-paste template lives in references/prd-template.md.

Copy it into docs/prds/[feature-name].md and fill it in. It covers: raw request, aligned understanding, goals/non-goals/success metrics, scope, user stories with acceptance criteria, UX requirements, technical context, agent instructions and prohibitions, implementation phases, testing/evaluation plan, rollout/monitoring/fallback, risks and open questions, and a readiness checklist.

5. PRD review rubric for agents

Before handing a PRD to an implementation agent, score it against this rubric.

A PRD is not ready for agent execution if two reviewers can reasonably disagree about what should be built.