Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Create a product requirement document (PRD) for a new feature. Use when the user wants to create a new feature, plan a feature, or write a PRD.
.claude/skills/platformplatform-create-prd/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-04 | ✗→✓ | ▲ Improved | 282% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 72% | 0% |
| case-09 | ✗→✓ | ▲ Improved | 95% | 0% |
| case-10 | ✗→✓ | ▲ Improved | 110% | 0% |
| case-15 | ✗→✓ | ▲ Improved | 1095% | 0% |
Your job is to work with the user through an interactive wizard to create a high-level PRD using language that is easy to understand for non-technical people. The PRD defines a feature] with all tasks] to be created in [PRODUCT_MANAGEMENT_TOOL].
Team leads: execute this workflow directly. Do not delegate it.
/.claude/reference/product-management/[PRODUCT_MANAGEMENT_TOOL].md to understand terminology, status mapping, ID format, and MCP configuration.Follow the steps below to create the PRD.
[PRODUCT_MANAGEMENT_TOOL]Follow initialization steps in /.claude/reference/product-management/[PRODUCT_MANAGEMENT_TOOL].md.
Use the AskUserQuestion tool to ask the user what feature] they want to build:
AskUserQuestion with:
- question: "What feature would you like to build?"
- header: "Feature"
- multiSelect: false
- options:
- label: "New feature", description: "Create a new feature"
- label: "Enhancement", description: "Enhance existing functionality"Users will typically use the custom text option to describe their feature].
If the user's answer comes back empty:
If you receive a valid answer:
Conduct deep research for a feasible solution that takes the existing codebase and features] into consideration:
main, account) the feature] belongs to. Back-office features live under account/Core/Features/BackOffice/ and are served on the back-office host.Now that you've done research, ask the user ALL required questions in ONE single AskUserQuestion call:
AskUserQuestion with 3 questions:
Question 1 - Feature name:
- question: "What is the name of this feature? (Use sentence case, e.g., 'User management' not 'User Management')"
- header: "Feature name"
- multiSelect: false
- options:
- label: "Custom name", description: "Enter your feature name"
Question 2 - Self-contained system (put the most likely SCS first based on research):
- question: "Which self-contained system (SCS) should this feature belong to?"
- header: "SCS"
- multiSelect: false
- options:
- label: "account", description: "Tenant and user management system (also hosts the back-office surface for support and system admin tools)"
- label: "main", description: "Primary shell application where you build your product"
- label: "[Suggested SCS based on research]", description: "Based on my analysis"
Question 3 - E2E tests:
- question: "Should this PRD include Playwright end-to-end tests?"
- header: "E2E Tests"
- multiSelect: false
- options:
- label: "Yes", description: "Include E2E tests as a separate [task]"
- label: "No", description: "Skip E2E tests for now"Ask additional questions:
After the first 3 questions, ask additional relevant questions to gather comprehensive requirements. Use multiple AskUserQuestion calls (max 4 questions per call, max 4 options per question).
Ask as many questions as needed to understand:
The more questions you ask, the better the PRD.
Implementation approach:
AskUserQuestion with:
- question: "Should we create frontend mockups first for UI/UX exploration?"
- header: "Approach"
- multiSelect: false
- options:
- label: "Yes", description: "Frontend mockups first to validate UI/UX before backend"
- label: "No", description: "Backend-first approach (default)"Based on all the research and user answers, draft the complete PRD.
Create the PRD content following the example PRD structure:
Examples based on common patterns:
Example 1 - Backend-first approach (default):
Example 2 - Frontend-first approach:
Example 3 - Backend-only feature]:
Example 4 - Large complex feature]:
Note: These are examples only. Adapt the task] structure to match the actual feature] requirements, scope, and user answers. All work is sequential -- one task] fully completed before the next starts.
[PRODUCT_MANAGEMENT_TOOL], create tasks] in the same order they appear in the PRD—this defines the implementation sequenceExample of WRONG task description (missing business rules):
### 1. Backend for team management
This task implements team CRUD operations with API endpoints and tests.
- Create Team aggregate
- Create CreateTeam command
- Create API endpoints
- Create testsExample of CORRECT task description (includes business rules):
### 1. Backend for team management
This task implements team CRUD operations with API endpoints and tests. Teams are managed by Tenant Owners and Admins only. Team names must be unique within a tenant.
- Create Team aggregate with name uniqueness validation
- Create CreateTeam command with Owner/Admin permission guard
- Create UpdateTeam command with Owner/Admin permission guard
- Create DeleteTeam command with Owner/Admin permission guard
- Create API endpoints for all operations
- Create tests covering permissions (403 for non-owners/admins), name uniqueness, tenant isolationFor frontend tasks, include ASCII art fat marker sketches showing UI layout and components:
### 2. Frontend for user management
This task implements the Users page UI. Users can only be managed by Tenant Owners or Admins.
┌─────────────────────────────────────────┐
│ Users [+ Invite user] │
├─────────────────────────────────────────┤
│ ┌─────────────────────────────────────┐ │
│ │ Email Name Role │ │
│ ├─────────────────────────────────────┤ │
│ │ admin@... John Doe Owner │ │
│ │ member@... Jane Smith Member │ │
│ └─────────────────────────────────────┘ │
└─────────────────────────────────────────┘
- Add Users navigation menu item
- Create Users page with table
- Create CreateUserDialog (validates email uniqueness)
- Show/hide [+ Invite user] button based on role (Owner/Admin only)
- Create UserDetailsSidePane
- Integrate all API operationsASCII sketches help engineers visualize the UI before coding.
Show the complete PRD to the user - display the full content including all tasks] with their descriptions.
Ask for approval: "Does this PRD look good?" (Yes/No)
Follow your PRODUCT_MANAGEMENT_TOOL]-specific guide at /.claude/reference/product-management/[PRODUCT_MANAGEMENT_TOOL].md to understand how to create items based on the PRD.
Create:
Each task] description must include:
After creating all tasks], update each feature]'s description in PRODUCT_MANAGEMENT_TOOL] with the full PRD content for that feature] -- the intro paragraph, overview section, and core changes section (everything above the "Tasks overview" heading). This ensures the PRD context is available to anyone viewing the feature] in PRODUCT_MANAGEMENT_TOOL].
Inform user: The feature] and all tasks] have been created in PRODUCT_MANAGEMENT_TOOL]. Ask the user if they would like to start implementing the first task now.
✅ DO:
❌ DON'T:
[PRODUCT_MANAGEMENT_TOOL] before getting PRD approval in Step 5prd.mdDo the research. Read code and rule files. Ask comprehensive questions. Create excellent PRDs.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 8,117 | 7,653 | -6% | 1 | 1 | 0% | 1,395 | 3,454 | +148% | 0 | 0 | — |
case-02 | fail→fail | 9,722 | 4,935 | -49% | 1 | 1 | 0% | 1,511 | 3,084 | +104% | 0 | 0 | — |
case-03 | fail→fail | 7,160 | 8,509 | +19% | 1 | 1 | 0% | 1,160 | 3,629 | +213% | 0 | 0 | — |
case-04 | fail→pass | 4,975 | 1,575 | -68% | 1 | 1 | 0% | 788 | 3,014 | +282% | 0 | 0 | — |
case-05 | pass→pass | 9,360 | 3,930 | -58% | 1 | 1 | 0% | 1,509 | 3,402 | +125% | 0 | 0 | — |
case-06 | fail→fail | 8,120 | 1,462 | -82% | 1 | 1 | 0% | 1,315 | 3,004 | +128% | 0 | 0 | — |
case-07 | pass→pass | 12,147 | 2,770 | -77% | 1 | 1 | 0% | 1,889 | 3,210 | +70% | 0 | 0 | — |
case-08 | fail→pass | 11,167 | 2,501 | -78% | 1 | 1 | 0% | 1,850 | 3,180 | +72% | 0 | 0 | — |
case-09 | fail→pass | 11,652 | 3,203 | -73% | 1 | 1 | 0% | 1,676 | 3,273 | +95% | 0 | 0 | — |
case-10 | fail→pass | 11,470 | 3,189 | -72% | 1 | 1 | 0% | 1,643 | 3,444 | +110% | 0 | 0 | — |
case-11 | pass→pass | 9,410 | 1,735 | -82% | 1 | 1 | 0% | 1,415 | 3,015 | +113% | 0 | 0 | — |
case-12 | pass→pass | 12,644 | 6,063 | -52% | 1 | 1 | 0% | 1,929 | 3,702 | +92% | 0 | 0 | — |
case-13 | fail→fail | 12,044 | 1,975 | -84% | 1 | 1 | 0% | 1,867 | 3,099 | +66% | 0 | 0 | — |
case-14 | fail→fail | 11,951 | 2,282 | -81% | 1 | 1 | 0% | 1,834 | 3,095 | +69% | 0 | 0 | — |
case-15 | fail→pass | 2,147 | 2,356 | +10% | 1 | 1 | 0% | 263 | 3,144 | +1095% | 0 | 0 | — |
case-16 | fail→pass | 5,116 | 1,407 | -72% | 1 | 1 | 0% | 733 | 2,929 | +300% | 0 | 0 | — |
case-17 | fail→pass | 9,603 | 1,872 | -81% | 1 | 1 | 0% | 1,498 | 3,024 | +102% | 0 | 0 | — |
case-18 | fail→pass | 13,157 | 2,918 | -78% | 1 | 1 | 0% | 2,204 | 3,153 | +43% | 0 | 0 | — |
case-19 | fail→pass | 8,178 | 2,092 | -74% | 1 | 1 | 0% | 1,331 | 3,049 | +129% | 0 | 0 | — |
case-20 | pass→fail | 9,318 | 22,084 | +137% | 1 | 1 | 0% | 1,397 | 4,334 | +210% | 0 | 0 | — |
case-21 | pass→fail | 9,728 | 26,799 | +175% | 1 | 1 | 0% | 1,469 | 5,788 | +294% | 0 | 0 | — |
case-22 | fail→fail | 4,737 | 11,682 | +147% | 1 | 1 | 0% | 768 | 4,639 | +504% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 19 counted toward the lift figure. The other 3 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +32 percentage points is the difference between those two pass rates over the 19 comparable cases. 3 cases got worse with the skill loaded, and they are included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.