| name | prd-author |
| description | Create and edit Strategic PRDs from human-provided context. Use this skill whenever you need to start a new Strategic PRD from scratch, revise an existing one due to scope changes, incorporate amendments proposed by the prd-learner skill, or make any structural changes to a PRD document. Trigger when the user says things like "write a PRD", "create a product spec", "update the PRD", "the scope changed", "incorporate these learnings", or "draft requirements for X". This is the only skill authorised to write or modify Strategic PRD documents — always use it when a PRD is the output. |
Skill: prd-author
Purpose: Create and edit Strategic PRDs from human-provided context. This is the only skill authorized to write or modify Strategic PRD documents.
When to Use
- Starting a new product engagement (create mode).
- A client's scope changes significantly and the PRD needs revision (edit mode).
- The
prd-learner skill has proposed amendments that require incorporation (edit mode).
- A human requests structural changes to an existing PRD (edit mode).
Two Modes
Create Mode
Takes messy human input and produces a complete Strategic PRD (draft status). The skill works conversationally, in phases, to ensure quality and alignment before writing.
Edit Mode
Takes an existing Strategic PRD and applies targeted changes. Preserves unchanged sections. Never downgrades severity levels without explicit human instruction.
Create Mode Workflow
Mode detection: Before Phase 1, check whether playbook.md exists at the project root and has at least one entry.
- No
playbook.md or file is empty: teach mode -- show rationale inline on every question, announce the companion doc when it generates.
playbook.md has entries: standard mode -- tight questions, rationale available on request, companion doc generated without announcement.
- User declares "expert" in conversation: expert mode -- no inline teaching, companion doc generated silently, playbook appended without narration.
Phase 1: Gather and Assess Input
Collect all available source material:
- Call transcripts and recordings
- Client-provided documents (specs, decks, briefs)
- Prototypes, wireframes, screenshots
- Existing PRDs (if extending a product)
- URLs to external resources (fetch and incorporate)
- Notes from human conversations
Assess input completeness. Before proceeding, determine:
- Is the problem space clear?
- Are the target users identified?
- Is there enough to define a solution approach?
- What domains are likely relevant? (product, tech, data, design, AI, GTM)
If critical gaps exist, surface them in Phase 2 rather than guessing.
Phase 2: Interview
Use AskUserQuestion to fill gaps and confirm direction. Batch questions into 2-4 per call. Adapt based on what's missing from Phase 1.
Teach-through-choice: Every question follows the three-things format: the question, a one-line rationale ("why we ask"), and optionally 2-3 representative answers so the user learns what good sounds like by contrast. In standard mode, skip representative examples unless the user seems unsure. In expert mode, skip rationale unless asked.
Standard questions (adapt, don't ask verbatim):
-
What is the single most important outcome for this product/feature?
Why we ask: Forces prioritization before scope locks. Without a primary outcome, every goal becomes equally important and the PRD cannot guide trade-off decisions.
-
Who are the primary users, and what's their technical sophistication?
Why we ask: Sophistication shapes UX requirements, onboarding complexity, and how much guidance the product needs to build in.
-
What's the timeline and what must ship in the first release vs. later?
Why we ask: Phase boundaries prevent everything from being R1. Without a hard line, the PRD becomes a wish list.
-
Are there hard technical, regulatory, or commercial constraints we should know about?
Why we ask: Constraints surface domains. A product handling financial data has compliance requirements; one replacing an internal tool has migration requirements. These rarely surface unless asked directly.
Domain forcing function: Based on the input material, identify which domains are likely relevant. If the human hasn't mentioned a domain that seems important (e.g. no data/privacy discussion for a product handling PII), proactively raise it:
"The product will handle [type of data]. Should we include data privacy requirements, or is that handled elsewhere?"
This is one of the system's most important mechanisms — it ensures senior product thinking regardless of the author's experience.
Commercial context probe: In addition to domain forcing, determine the commercial context early. Ask a gating question: "Is this an external product sold to customers, or an internal tool for your organization?"
- If external: Probe for commercial dimensions: "Who buys this — not just who uses it? How do they discover it? What's the pricing model? What sales motion do you envision — self-serve, product-led growth, sales-led? What does the buyer journey look like before someone becomes a user?" These answers shape GTM requirements in Section 5 and the buyer journey in Section 3.
- If internal: Probe for stakeholder dimensions: "Who sponsors this internally? What does it replace? What's the maintenance plan after launch? Who supports it day-to-day? What's the build-vs-buy case — why build this rather than buy off the shelf?" These answers shape GTM requirements in Section 5 and the stakeholder journey in Section 3.
- If not applicable (purely technical module, infrastructure, or the human says "skip"): Omit GTM from the PRD.
If the human provides a URL: Use WebFetch to retrieve and incorporate the content.
If extending an existing product: Read existing PRDs (depends_on candidates) and reference them in the interview to check for conflicts or shared requirements.
Phase 3: Outline and Confirm
Before writing the full PRD, produce a structured outline for human approval:
PRD-XXX: [Proposed Title]
Context: [1-2 sentence summary of the problem]
Goals (R1): [Bullet list of 3-5 primary goals]
Non-goals: [Bullet list of what's explicitly excluded]
Users: [Primary and secondary user roles]
Commercial context: [external / internal / not applicable] — [1-line summary]
Solution approach: [1-2 sentence summary]
Domains covered: [List which domain subsections will be included]
Key risks: [1-2 top risks]
Present this to the human. Wait for approval or redirection before proceeding.
Why this phase matters: Writing a full PRD is a significant token investment. Getting alignment on the outline first catches misunderstandings early and avoids expensive rewrites.
Phase 3.5: Shape-Check the Outline
Before writing the full PRD, assess the outline the human has confirmed. Keep this to 2-4 sentences. Name what is strong. Name what is thin. Do not block -- shape-check is tutoring, not a gate.
What to assess:
- Problem statement: Grounded in specific evidence, or generic? ("Users want better reporting" is generic; "users export to CSV because the chart builder doesn't support multi-series comparisons" is specific and buildable.)
- Goals: Tied to a phase and measurable, or at minimum noted explicitly as qualitative?
- Non-goals: At least two named? An outline with no non-goals is a scope risk -- scope expands wherever no boundary is drawn.
- Users: A real segment with a named problem, or an assumed persona built from guesses?
In expert mode: skip this phase. In standard mode: surface only the one or two items that are genuinely thin. In teach mode: run all four checks and name what good would look like.
Phase 4: Author the Strategic PRD
Write the complete Strategic PRD following the 8-section structure defined in the PRD System Definition.
Section-by-section guidance:
Section 1 (Context & Problem): Ground this in specific evidence from the input material. Quote or reference transcripts, client docs, and existing pain points. Avoid generic problem statements.
Section 2 (Goals, Non-goals & Success Metrics): Every goal should be tied to a phase. Non-goals are mandatory — if the human didn't specify any, propose them based on scope signals. Success metrics should be measurable where possible; qualitative is acceptable when quantitative isn't realistic.
Section 3 (Users & Scenarios): User stories follow the format: US-01 (Role, Phase, Severity): As a [role], I want [action] so that [outcome]. Include 2-3 narrative scenarios that walk through end-to-end usage. If the commercial context probe identified this as an external product, add a Buyer Journey subsection mapping the stages from unaware through churning/graduating — who is involved at each stage, what triggers progression. If internal, add a Stakeholder Journey mapping the organizational lifecycle from need identification through ongoing maintenance.
Section 4 (Solution Overview): Architecture-level, not implementation-level. Name components, services, and integrations. A builder reading this should understand the shape of the system before seeing detailed requirements.
Section 5 (Requirements by Domain): This is the bulk of the PRD. For each relevant domain:
- Use the tabular format:
| ID | Severity | Requirement | Acceptance Criteria | Phase |
- Every
must requirement needs at least one testable acceptance criterion.
- Be prescriptive: "API must return paginated results with max 50 items per page" not "API should handle large datasets."
- Cross-reference user stories where applicable.
- Include only domains that are genuinely relevant. A simple CRUD app doesn't need a GTM section.
- GTM domain guidance: If external, GTM requirements should cover: pricing tiers and limits, free-tier boundaries, trial mechanics, onboarding conversion targets, positioning statements that the product must support. If internal, GTM requirements should cover: migration steps from the tool being replaced, training requirements, rollback plan, SLA expectations, ownership and support model. If the commercial context probe returned "not applicable," omit GTM.
Section 6 (Guardrails): These are the do's and don'ts. They must be:
- Specific — Name exact patterns, tools, APIs, conventions.
- Verifiable — An agent (or human) can objectively check compliance.
- Non-redundant — Don't restate requirements from Section 5 as guardrails. Guardrails are cross-cutting rules, not feature specs.
Section 7 (Risks & Open Questions): Every risk needs a mitigation. Every open question needs an owner and a target resolution date. If the human didn't specify owners, assign them and flag it for confirmation.
Section 8 (Learnings): Leave empty for new PRDs. Add a placeholder note:
## 8. Learnings
No learnings recorded yet. This section will be populated by the `prd-learner` skill as implementation progresses.
Phase 5: Present and Iterate
Share the draft with the human. Explicitly flag:
- Any assumptions made due to incomplete input.
- Any domain sections that feel thin and might need more context.
- Any open questions that block implementation.
The PRD status is draft until the human approves it. Upon approval, update status to active.
Phase 5.5: Companion Doc and Playbook
After status moves to active, generate two artifacts. Do not wait for the human to ask.
Companion doc: Create PRD-XXX-companion.md in the same directory as the PRD, using templates/prd-companion.md as the base. Fill each section with content specific to this session: name the actual moves made, the actual decisions, the actual anti-patterns surfaced. A generic companion doc is useless. The value is in the specificity.
Playbook entry: Append one entry to playbook.md at the project root. Create the file from templates/playbook.md if it does not exist.
## [date] PRD-XXX: [short title]
Decomposition move: [the most reusable framing or structural decision from this session]
What to watch for: [one anti-pattern surfaced during authoring]
Mode behavior:
- Teach mode: surface the companion doc actively, mention where it lives, walk through one section to show what it contains.
- Standard mode: mention that the companion doc was generated and name its path.
- Expert mode: generate both silently.
Edit Mode Workflow
Step 1: Read the Current PRD
Load the full Strategic PRD. Parse frontmatter and all sections.
Step 2: Understand the Requested Change
Changes can come from:
- Human request (direct edit instruction).
prd-learner output (proposed amendments with learning context).
- Scope change (new client request that modifies existing requirements).
Step 3: Apply Changes
Principles for editing:
- Surgical edits. Use targeted changes. Don't rewrite sections that aren't affected.
- Additive by default. Prefer adding new requirements, guardrails, or learnings over removing existing ones. Removal requires explicit human instruction.
- Never downgrade severity. Changing a
must to a should requires explicit human approval with documented reasoning.
- Preserve IDs. Never change requirement IDs. If a requirement is superseded, deprecate it (strikethrough or note) and create a new one.
- Cross-reference learnings. If the edit is driven by a learning, reference the learning ID (e.g.
L-001) in the amended text or as a comment.
- Bump metadata. Update
updated_at in frontmatter after any edit.
Step 4: Summarize Changes
After editing, produce a brief change summary:
Changes to PRD-XXX:
- Section 5.2: Added TECH-05 (must) — require rate limiting on all public endpoints. [Driven by L-003]
- Section 6: Added DON'T — expose stack traces in API error responses. [Driven by L-003]
- Section 7: Resolved OQ-02 — confirmed OAuth2 as auth standard.
- Frontmatter: updated_at bumped to 2026-03-15.
This summary is for the human reviewer and for prd-taskmaster to know which task contexts need regeneration.
PRD Frontmatter Template
---
id: PRD-001
title: "Feature Name – Phase"
product: "Product Name"
area: core-feature
status: draft
phase: R1
priority: P0
owner: owner.name
contributors:
- name.one
- name.two
depends_on:
- PRD-002
labels:
- ai-native
links:
design: https://...
transcript: https://...
repo: https://...
created_at: 2026-03-01
updated_at: 2026-03-01
---
Quality Standards
Tone and Language
- Prescriptive, not advisory. Use "must", "will", "is required" — not "should consider", "might want to."
- Specific, not vague. Name exact tools, versions, patterns, thresholds.
- Evidence-grounded. Tie requirements to user needs, technical constraints, or business goals — not assumptions.
- Balanced. Acknowledge trade-offs honestly. Don't oversell.
Calibration by Scope
| PRD Type | Expected Length | Focus Areas |
|---|
| Single feature | 200-400 lines | Sections 3-6 heavy, minimal 7 |
| Product module | 400-700 lines | All sections substantial |
| Full product (R1) | 700-1200 lines | All sections comprehensive, multiple domain subsections |
What Makes a Good PRD
- Every
must requirement has testable acceptance criteria.
- Non-goals are explicit and specific, not generic ("we won't do everything").
- The domain forcing function was applied — relevant domains are covered even if the human didn't raise them.
- Risks have mitigations. Open questions have owners and dates.
- Guardrails are verifiable, not aspirational.
- A builder reading sections 4-6 can start working without needing to ask clarifying questions.