| name | write-user-story |
| description | Write user stories with Gherkin acceptance criteria, edge cases, anti-requirements, and ISC-validated criteria from a feature description or PRD section. |
| argument-hint | [feature description or PRD reference] |
| user-invocable | true |
| allowed-tools | Read, Write, Edit, Glob, Grep |
Write user stories for $ARGUMENTS.
Follow every step below. Each story must be complete enough to hand directly to an engineer without additional conversation.
Step 1: Identify the User Types
Before writing any stories, list every distinct user type involved. Do not use generic labels.
Anti-patterns (never use these):
- "As a user" — which user? This is always too vague.
- "As an admin" — what kind of admin? What are they trying to accomplish?
- "As the system" — systems do not have goals. Reframe as the user who benefits from the system behaviour.
Good user types are specific about role AND context:
- "As a team lead reviewing my direct reports' weekly submissions"
- "As a first-time visitor who arrived from a Google search"
- "As a billing admin who manages subscriptions for 50+ seats"
- "As a developer integrating our API for the first time"
For each user type, note:
- Their goal (what they are trying to accomplish overall, not just in this story)
- Their technical sophistication
- Their frequency of interaction (daily power user vs. monthly visitor)
Step 2: Write One Story Per Behaviour
Each story captures exactly one user behaviour. Apply this test: if the story has "and" in the action clause, it is two stories. Split it.
Story Format
### US-[N]: [Short descriptive title]
**As a** [specific user type with context],
**I want** [concrete action the user takes — not what the system does],
**so that** [the value to the user — why they care, not how it works].
Rules for Each Clause
User type clause:
- Must be a specific person from Step 1, not a generic role
- Include situational context when it matters: "a returning user who has saved items in their cart"
Action clause:
- Describes what the user does, not what the system does
- Bad: "I want the system to send me a notification"
- Good: "I want to receive a notification when my build fails"
- Must be a single, discrete action — no compound actions with "and"
Value clause:
- States the benefit to the user, not the implementation mechanism
- Bad: "so that the database is updated"
- Good: "so that I can track which items I have already reviewed"
- Must answer "why would the user care about this?"
Step 3: Write Gherkin Acceptance Criteria
Every story must have acceptance criteria written in Gherkin format. Gherkin makes criteria unambiguous and directly translatable to automated tests.
Scenario: [Descriptive name of the scenario]
Given [precondition — the state of the world before the action]
When [action — what the user does]
Then [outcome — what the user observes]
Rules for Gherkin Criteria
- One scenario per behaviour. If a scenario has multiple
When steps, split it.
- Given sets up context, not actions.
Given I am logged in is context. Given I click the login button is an action — put actions in When.
- Then describes observable outcomes.
Then the database is updated is not observable to the user. Then I see a confirmation message "Item saved" is observable.
- Use concrete values, not placeholders.
When I enter "john@example.com" is better than When I enter a valid email.
- Include the negative case. For every happy-path scenario, write the corresponding error scenario.
Example
Scenario: Successful password reset request
Given I am on the login page
And I have a registered account with email "user@example.com"
When I click "Forgot password"
And I enter "user@example.com" in the email field
And I click "Send reset link"
Then I see the message "Check your email for a reset link"
And an email is sent to "user@example.com" within 60 seconds
Scenario: Password reset with unregistered email
Given I am on the login page
When I click "Forgot password"
And I enter "unknown@example.com" in the email field
And I click "Send reset link"
Then I see the message "Check your email for a reset link"
And no email is sent
# Security: same message shown to prevent email enumeration
Step 4: Apply the ISC Splitting Test
Every acceptance criterion (Gherkin scenario) must pass all three checks:
Independent
Can this scenario be verified without first executing another scenario? If Scenario B requires Scenario A to have run first, they are not independent — either merge them or restructure so B sets up its own preconditions in the Given block.
Small
Does this scenario test exactly one behaviour? If the Then block asserts multiple unrelated things, split into separate scenarios. Multiple related assertions (e.g., "Then I see the item in the list And the count updates to 5") are fine if they are consequences of the same action.
Complete
Does this scenario cover the boundary? Check:
- What happens with empty input?
- What happens with maximum-length input?
- What happens with special characters?
- What happens when the user does not have permission?
- What happens with concurrent access?
If any boundary is missing, add a scenario for it.
Step 5: Document Edge Cases
For every story, enumerate edge cases explicitly. Do not leave them implicit. Common categories:
| Category | Questions to answer |
|---|
| Empty state | What does the user see when there is no data yet? First-time experience? |
| Error handling | What happens on network failure? Timeout? Invalid input? Server error? |
| Permissions | What happens if the user lacks permission? Expired session? |
| Concurrency | What if two users act on the same item simultaneously? |
| Scale | What happens with 0 items? 1 item? 1,000 items? 100,000 items? |
| Accessibility | Can the entire flow be completed with keyboard only? With a screen reader? |
| Undo/recovery | Can the user reverse the action? What is the recovery path for mistakes? |
Write a Gherkin scenario for each edge case that could affect user experience. Minor edge cases (purely cosmetic at extreme scale) can be noted without full Gherkin.
Step 6: Define Anti-Requirements
List what this story deliberately does NOT cover. Anti-requirements prevent scope creep and make implicit exclusions explicit.
Format:
**Anti-requirements:**
- This story does NOT cover bulk operations — that is a separate story (US-[N])
- This story does NOT handle the admin workflow — admin stories are in US-[N] through US-[N]
- This story does NOT include email notifications — notifications are tracked in [reference]
Every anti-requirement should reference where the excluded behaviour IS tracked (another story, a future initiative, or "out of scope entirely").
Step 7: Story Sizing Validation
Before finalising, validate each story against these size constraints:
- Too small: A story that takes less than half a day is probably a task, not a story. Combine it with related work.
- Right size: Completable by one engineer in 1-5 days, including tests.
- Too large: If an engineer would need more than one week, split the story further. Apply the INVEST criteria to find the split point.
Output Format
Group related stories under a feature heading. Number stories sequentially.
# User Stories: [Feature Name]
## User Types
1. **[Type A]** — [one-line description with context]
2. **[Type B]** — [one-line description with context]
---
## [Feature Area 1]
### US-1: [Title]
**As a** [specific user type],
**I want** [action],
**so that** [value].
**Acceptance Criteria:**
\```gherkin
Scenario: [Happy path]
Given ...
When ...
Then ...
Scenario: [Error case]
Given ...
When ...
Then ...
Scenario: [Edge case]
Given ...
When ...
Then ...
\```
**Edge Cases:**
- [Edge case not covered by Gherkin scenarios, noted for awareness]
**Anti-requirements:**
- Does NOT cover [excluded behaviour] — tracked in [reference]
**Size:** [S/M/L]
---
Write the output to a file if writing more than 3 stories: docs/stories-[feature-name].md.
Related Skills
/product-owner:groom-backlog — groom the backlog to identify which items need user stories written./product-owner:write-prd — user stories implement requirements from the PRD. Reference the PRD for context.
