---
name: product-requirements-writing
description: Write a feature spec or PRD from a problem statement or feature idea. Use when turning a vague idea or user request into a structured document, scoping a feature with goals and non-goals, defining success metrics and acceptance criteria, or breaking a big ask into a phased spec.
argument-hint: <feature or problem statement>
license: Apache-2.0
metadata:
author: Anthropic
source: anthropics/knowledge-work-plugins/product-management/skills/write-spec
source_url: https://github.com/anthropics/knowledge-work-plugins/blob/1c7187c4fc17feefa6cde39517f12dae1249e6c4/product-management/skills/write-spec/SKILL.md
upstream-commit: 1c7187c4fc17feefa6cde39517f12dae1249e6c4
original-name: write-spec
packaged-by: AutoGPT
packaged-on: '2026-09-25'
---
> Packaging adaptation by AutoGPT, 2026-09-25. Original authorship and licence are retained. Changes are limited to the recorded name, metadata and local references; see ATTRIBUTION.md in this package.
# Write Spec
> If you see unfamiliar placeholders or need to check which tools are connected, see [CONNECTORS.md](references/CONNECTORS.md).
Write a feature specification or product requirements document (PRD).
## Usage
```
/write-spec $ARGUMENTS
```
## Workflow
### 1. Understand the Feature
Ask the user what they want to spec. Accept any of:
- A feature name ("SSO support")
- A problem statement ("Enterprise customers keep asking for centralized auth")
- A user request ("Users want to export their data as CSV")
- A vague idea ("We should do something about onboarding drop-off")
### 2. Gather Context
Ask the user for the following. Be conversational — do not dump all questions at once. Ask the most important ones first and fill in gaps as you go:
- **User problem**: What problem does this solve? Who experiences it?
- **Target users**: Which user segment(s) does this serve?
- **Success metrics**: How will we know this worked?
- **Constraints**: Technical constraints, timeline, regulatory requirements, dependencies
- **Prior art**: Has this been attempted before? Are there existing solutions?
### 3. Pull Context from Connected Tools
If **~~project tracker** is connected:
- Search for related tickets, epics, or features
- Pull in any existing requirements or acceptance criteria
- Identify dependencies on other work items
If **~~knowledge base** is connected:
- Search for related research documents, prior specs, or design docs
- Pull in relevant user research findings
- Find related meeting notes or decision records
If **~~design** is connected:
- Pull related mockups, wireframes, or design explorations
- Search for design system components relevant to the feature
If these tools are not connected, work entirely from what the user provides. Do not ask the user to connect tools — just proceed with available information.
### 4. Generate the PRD
Produce a structured PRD with these sections. See **PRD Structure** below for detailed guidance on what each section should contain.
- **Problem Statement**: The user problem, who is affected, and impact of not solving it (2-3 sentences)
- **Goals**: 3-5 specific, measurable outcomes tied to user or business metrics
- **Non-Goals**: 3-5 things explicitly out of scope, with brief rationale for each
- **User Stories**: Standard format ("As a [user type], I want [capability] so that [benefit]"), grouped by persona
- **Requirements**: Categorized as Must-Have (P0), Nice-to-Have (P1), and Future Considerations (P2), each with acceptance criteria
- **Success Metrics**: Leading indicators (change quickly) and lagging indicators (change over time), with specific targets
- **Open Questions**: Unresolved questions tagged with who needs to answer (engineering, design, legal, data)
- **Timeline Considerations**: Hard deadlines, dependencies, and phasing
### 5. Review and Iterate
After generating the PRD:
- Ask the user if any sections need adjustment
- Offer to expand on specific sections
- Offer to create follow-up artifacts (design brief, engineering ticket breakdown, stakeholder pitch)
## PRD Structure
### Problem Statement
- Describe the user problem in 2-3 sentences
- Who experiences this problem and how often
- What is the cost of not solving it (user pain, business impact, competitive risk)
- Ground this in evidence: user research, support data, metrics, or customer feedback
### Goals
- 3-5 specific, measurable outcomes this feature should achieve
- Each goal should answer: "How will we know this succeeded?"
- Distinguish between user goals (what users get) and business goals (what the company gets)
- Goals should be outcomes, not outputs ("reduce time to first value by 50%" not "build onboarding wizard")
### Non-Goals
- 3-5 things this feature explicitly will NOT do
- Adjacent capabilities that are out of scope for this version
- For each non-goal, briefly explain why it is out of scope (not enough impact, too complex, separate initiative, premature)
- Non-goals prevent scope creep during implementation and set expectations with stakeholders
### User Stories
Write user stories in standard format: "As a [user type], I want [capability] so that [benefit]"
Guidelines:
- The user type should be specific enough to be meaningful ("enterprise admin" not just "user")
- The capability should describe what they want to accomplish, not how
- The benefit should explain the "why" — what value does this deliver
- Include edge cases: error states, empty states, boundary conditions
- Include different user types if the feature serves multiple personas
- Order by priority — most important stories first
Example:
- "As a team admin, I want to configure SSO for my organization so that my team members can log in with their corporate credentials"
- "As a team member, I want to be automatically redirected to my company's SSO login so that I do not need to remember a separate password"
- "As a team admin, I want to see which members have logged in via SSO so that I can verify the rollout is working"
### Requirements
**Must-Have (P0)**: The feature cannot ship without these. These represent the minimum viable version of the feature. Ask: "If we cut this, does the feature still solve the core problem?" If no, it is P0.
**Nice-to-Have (P1)**: Significantly improves the experience but the core use case works without them. These often become fast follow-ups after launch.
**Future Considerations (P2)**: Explicitly out of scope for v1 but we want to design in a way that supports them later. Documenting these prevents accidental architectural decisions that make them hard later.
For each requirement:
- Write a clear, unambiguous description of the expected behavior
- Include acceptance criteria (see below)
- Note any technical considerations or constraints
- Flag dependencies on other teams or systems
### Open Questions
- Questions that need answers before or during implementation
- Tag each with who should answer (engineering, design, legal, data, stakeholder)
- Distinguish between blocking questions (must answer before starting) and non-blocking (can resolve during implementation)
### Timeline Considerations
- Hard deadlines (contractual commitments, events, compliance dates)
- Dependencies on other teams' work or releases
- Suggested phasing if the feature is too large for one release
## User Story Writing
Good user stories are:
- **Independent**: Can be developed and delivered on their own
- **Negotiable**: Details can be discussed, the story is not a contract
- **Valuable**: Delivers value to the user (not just the team)
- **Estimable**: The team can roughly estimate the effort
- **Small**: Can be completed in one sprint/iteration
- **Testable**: There is a clear way to verify it works
### Common Mistakes in User Stories
- Too vague: "As a user, I want the product to be faster" — what specifically should be faster?
- Solution-prescriptive: "As a user, I want a dropdown menu" — describe the need, not the UI widget
- No benefit: "As a user, I want to click a button" — why? What does it accomplish?
- Too large: "As a user, I want to manage my team" — break this into specific capabilities
- Internal focus: "As the engineering team, we want to refactor the database" — this is a task, not a user story
## Requirements Categorization
### MoSCoW Framework
- **Must have**: Without these, the feature is not viable. Non-negotiable.
- **Should have**: Important but not critical for launch. High-priority fast follows.
- **Could have**: Desirable if time permits. Will not delay delivery if cut.
- **Won't have (this time)**: Explicitly out of scope. May revisit in future versions.
### Tips for Categorization
- Be ruthless about P0s. The tighter the must-have list, the faster you ship and learn.
- If everything is P0, nothing is P0. Challenge every must-have: "Would we really not ship without this?"
- P1s should be things you are confident you will build soon, not a wish list.
- P2s are architectural insurance — they guide design decisions even though you are not building them now.
## Success Metrics Definition
### Leading Indicators
Metrics that change quickly after launch (days to weeks):
- **Adoption rate**: % of eligible users who try the feature
- **Activation rate**: % of users who complete the core action
- **Task completion rate**: % of users who successfully accomplish their goal
- **Time to complete**: How long the core workflow takes
- **Error rate**: How often users encounter errors or dead ends
- **Feature usage frequency**: How often users return to use the feature
### Lagging Indicators
Metrics that take time to develop (weeks to months):
- **Retention impact**: Does this feature improve user retention?
- **Revenue impact**: Does this drive upgrades, expansion, or new revenue?
- **NPS / satisfaction change**: Does this improve how users feel about the product?
- **Support ticket reduction**: Does this reduce support load?
- **Competitive win rate**: Does this help win more deals?
### Setting Targets
- Targets should be specific: "50% adoption within 30 days" not "high adoption"
- Base targets on comparable features, industry benchmarks, or explicit hypotheses
- Set a "success" threshold and a "stretch" target
- Define the measurement method: what tool, what query, what time window
- Specify when you will evaluate: 1 week, 1 month, 1 quarter post-launch
## Acceptance Criteria
Write acceptance criteria in Given/When/Then format or as a checklist:
**Given/When/Then**:
- Given [precondition or context]
- When [action the user takes]
- Then [expected outcome]
Example:
- Given the admin has configured SSO for their organization
- When a team member visits the login page
- Then they are automatically redirected to the organization's SSO provider
**Checklist format**:
- [ ] Admin can enter SSO provider URL in organization settings
- [ ] Team members see "Log in with SSO" button on login page
- [ ] SSO login creates a new account if one does not exist
- [ ] SSO login links to existing account if email matches
- [ ] Failed SSO attempts show a clear error message
### Tips for Acceptance Criteria
- Cover the happy path, error cases, and edge cases
- Be specific about the expected behavior, not the implementation
- Include what should NOT happen (negative test cases)
- Each criterion should be independently testable
- Avoid ambiguous words: "fast", "user-friendly", "intuitive" — define what these mean concretely
## Scope Management
### Recognizing Scope Creep
Scope creep happens when:
- Requirements keep getting added after the spec is approved
- "Small" additions accumulate into a significantly larger project
- The team is building features no user asked for ("while we're at it...")
- The launch date keeps moving without explicit re-scoping
- Stakeholders add requirements without removing anything
### Preventing Scope Creep
- Write explicit non-goals in every spec
- Require that any scope addition comes with a scope removal or timeline extension
- Separate "v1" from "v2" clearly in the spec
- Review the spec against the original problem statement — does everything serve it?
- Time-box investigations: "If we cannot figure out X in 2 days, we cut it"
- Create a "parking lot" for good ideas that are not in scope
## Output Format
Use markdown with clear headers. Keep the document scannable — busy stakeholders should be able to read just the headers and bold text to get the gist.
## Tips
- Be opinionated about scope. It is better to have a tight, well-defined spec than an expansive vague one.
- If the user's idea is too big for one spec, suggest breaking it into phases and spec the first phase.
- Success metrics should be specific and measurable, not vague ("improve user experience").
- Non-goals are as important as goals. They prevent scope creep during implementation.
- Open questions should be genuinely open — do not include questions you can answer from context.