bdd

BDD spec authoring — create, update, or delete .feature files following best practices

SKILL.md

GitHub

BDD Skill

Single authority for Behavior-Driven Development spec authoring. Creates, updates, or deletes .feature files in the spec/ pipeline following BDD best practices. This skill is invoked before implementation skills (/fix, /refactor, /reel) because behavior specs drive development.


Determine Scope

Check if scope was provided (accessible as $ARGUMENTS):

If scope is provided ($ARGUMENTS is not empty):

  • Use the provided description to focus spec work
  • Examples: spec/active/installer/, add login feature, remove deprecated API

If no scope provided ($ARGUMENTS is empty):

  • Ask user: "What behavior needs to be specified? (e.g., new feature, changed behavior, removed capability)"
  • Wait for user response

Phase 1: Assess Spec Impact

Determine what kind of spec change is needed:

1. Identify Affected Feature Files

Search spec/active/ and spec/todo/ for existing feature files related to the scope:

  • Read relevant .feature files in the affected domain subdirectory
  • Check if behavior is already specified (scenario exists)
  • Check if existing scenarios need updating (behavior changed)
  • Check if scenarios should be deleted (behavior removed)

2. Classify the Change

New behavior → Create new scenarios (or new feature file) in spec/todo/ Changed behavior → Update existing scenarios in spec/active/ Removed behavior → Delete scenarios from spec/active/ Mixed → Combination of the above

3. Present Assessment

Summarize what spec changes are needed. Present a picker:

Use AskUserQuestion with header "BDD scope", options:

  1. "Proceed with spec changes" — Write the feature file changes as assessed
  2. "Adjust scope" — Modify which specs to create/update/delete

Phase 2: Personas (Database/State-Store Projects Only)

This phase applies only when the project has a database or state storage backend. Skip entirely for projects without persistent state. Personas define the actors and objects that appear in Given steps — they must be defined before writing feature files that reference them.

Detect State Storage Presence

Check for indicators that the project uses a database or state store:

Python projects: Scan pyproject.toml or requirements.txt for ORM/driver dependencies (sqlalchemy, django, peewee, tortoise-orm, mongoengine, psycopg2, pymysql, pymongo, redis)

General indicators:

  • Migration directories (alembic/, migrations/, prisma/migrations/)
  • Config files (alembic.ini, database.yml, ormconfig.json, knexfile.js)
  • Environment variables in .env or config: DATABASE_URL, REDIS_URL, MONGO_URI
  • Docker services in docker-compose.yml: postgres, mysql, redis, mongo, elasticsearch

If no indicators found: Personas are narrative aids only — use them in the "As a [role]" Feature narrative and scenario names, but do not create persona files or seed data. Skip the rest of this phase.

Persona Types

Actor personas — people who interact with the system:

  • Users, customers, admins, operators, guests
  • Define: name, credentials, role, permissions, profile attributes
  • Map to the "As a [role]" narrative in Feature descriptions

Object personas — domain entities with specific states:

  • Products, orders, subscriptions, contracts, configurations
  • Define: name, key attributes, state, relationships
  • Named archetypes that appear frequently across scenarios (e.g., "expired subscription", "high-value order")

Persona File Location and Format

Persona files live at project root personas/, organized by type:

personas/
  actors/          — People who interact with the system
  objects/         — Domain entities with specific states

Format-agnostic — use whichever format matches the project's conventions:

Gherkin (.feature) — BDD-native approach. Personas as scenarios with Given steps. Step definitions handle database seeding. Best for teams already comfortable with Gherkin.

Feature: Customer personas
  Scenario: Alice the premium customer
    Given a user "Alice" with email "alice@example.com"
    And "Alice" has role "premium_customer"
    And "Alice" has subscription plan "annual"

  Scenario: Bob the free-tier user
    Given a user "Bob" with email "bob@example.com"
    And "Bob" has role "free_user"

YAML (.yml) — Data-oriented approach. A fixture loader reads persona definitions and seeds the database. Best for large persona sets or data-heavy projects.

alice:
  type: user
  name: Alice
  email: alice@example.com
  role: premium_customer
  plan: annual

bob:
  type: user
  name: Bob
  email: bob@example.com
  role: free_user

JSON (.json) — Machine-friendly variant of YAML. Same structure, different syntax. Best when tooling expects JSON.

Persona Authoring Conventions

Keep persona definitions minimal:

  • Only specify attributes that are semantically meaningful for the persona's role
  • Let factories or defaults fill in the rest — avoids brittleness when schema changes
  • A "premium customer" persona defines role and plan, not every database column

Use trait composition over persona explosion:

  • Define a base persona (e.g., "Alice") with core attributes
  • Apply traits in step definitions for variations: Given "Alice" with an expired payment method
  • Do NOT create separate personas for every slight variation (Alice-with-expired-card, Alice-with-two-addresses, etc.)

Name personas descriptively:

  • Actor names should be vivid and memorable: "Alice", "Bob", not "User1", "User2"
  • Object names should encode their key trait: "expired subscription", "high-value order"
  • The name alone should tell the reader what makes this persona interesting

Step Definition Patterns

Step definitions resolve persona names to seeded data. The implementation depends on the persona file format:

For Gherkin personas: Run the persona feature file as a setup step (or call shared Given steps directly). The step definitions create database records.

For YAML/JSON personas: A fixture or conftest function loads the persona file, iterates definitions, and seeds records via ORM or raw SQL. Example pattern for pytest-bdd:

# conftest.py
@given(parsers.parse('"{persona_name}" is logged in'), target_fixture='current_user')
def givenPersonaLoggedIn(persona_name, db_session):
    persona = PERSONAS[persona_name]  # loaded from YAML
    user = UserFactory(**persona)
    db_session.add(user)
    db_session.flush()
    return user

Test Isolation

  • Seed reference data (roles, plans, config) at session scope
  • Seed scenario-specific persona data at function scope with transaction rollback
  • Never share mutable persona state between scenarios
  • Each scenario gets fresh persona instances — never reuse database records across scenarios

Anti-Patterns to Avoid

  • Over-specifying: Defining every field when only 2-3 matter. Breaks when schema changes.
  • Shared mutable state: Scenarios reading/writing the same persona record without isolation.
  • Persona explosion: Dozens of personas for slight variations. Use trait composition instead.
  • Tight schema coupling: Persona files mirroring column names. Use a factory indirection layer.
  • Hidden setup: Burying critical preconditions in shared fixtures far from the scenario. Keep business-relevant setup visible.

Phase 3: Write Feature Files

Conventions

All feature files must follow these conventions:

File organization:

  • One feature per file, organized by domain in subdirectories
  • Filename: lowercase, hyphenated, .feature extension (e.g., user-authentication.feature)
  • New features go in spec/todo/<domain>/ (created during development)
  • Active specs live in spec/active/<domain>/
  • Domain subdirectories mirror the codebase's logical structure

Feature structure:

Feature: Short business-readable title
  As a [role]
  I want [capability]
  So that [benefit]

  Background:
    Given [shared preconditions — only if ALL scenarios share them]

  # Rule: Business rule grouping  (behave users: comment out Rule: lines — not supported)

    @id:abcd1234
    Scenario: Descriptive behavior outcome
      Given [context]
      When [action]
      Then [expected outcome]

Tagging (lines preceding each Scenario):

  • @id:<8-char-hex> — Mandatory. Generate via ~/.claude/scripts/bdd.sh --generate-id spec/
  • @depends-on:<id> — Optional. One per line. Establishes task ordering.
  • @acceptance:<shell-command> — Optional. Overrides BDD framework auto-detection for verification.

Writing style — DECLARATIVE, not imperative:

  • Describe WHAT behavior is expected, not HOW to interact with the UI
  • Use business language, not technical implementation details
  • A product owner should be able to read and validate the scenario
  • Use vivid persona names (e.g., "Alice") not abstract labels (e.g., "User A")

Scenario guidelines:

  • Each scenario must be independent and self-contained
  • Keep to 3-5 steps: one Given block, one When, one Then
  • Scenario names describe the behavior and expected outcome
  • No scenario should depend on another having run first
  • Do not test implementation details (database tables, CSS selectors, API internals)

When to use each construct:

  • Background: — Only Given steps shared by ALL scenarios in the feature. Keep short (1-4 lines).
  • Rule: — Group scenarios illustrating a single business rule (Gherkin 6+, pytest-bdd compatible; not supported by behave — comment out Rule: lines if automating with behave)
  • Scenario Outline: with Examples: — Data-driven scenarios with 3+ variations of the same behavior pattern
  • Separate feature files — When a file exceeds ~10 scenarios, split by sub-domain

Anti-patterns to avoid:

  • Imperative style (clicking buttons, filling fields) — use declarative behavior descriptions
  • Conjunction steps ("When I login and navigate to settings") — split into separate steps
  • Incidental details (specifying every field when only one matters)
  • Too many steps (more than 5 indicates testing multiple behaviors)
  • Testing implementation (referencing internal methods, database schema, etc.)

ID Generation

For each new scenario, generate a unique ID:

  1. Run ~/.claude/scripts/bdd.sh --generate-id spec/ to get a collision-free 8-char hex ID
  2. Place the @id:<hex> tag on the line immediately before the Scenario: line
  3. Never reuse or manually craft IDs — always generate them

Creating New Feature Files

  1. Determine the domain subdirectory (create if needed under spec/todo/)
  2. Generate @id: tags for each scenario
  3. Write the feature file following all conventions above
  4. Add @depends-on: tags where scenarios have ordering requirements
  5. Add @acceptance: tags only when the default BDD framework detection is insufficient

Updating Existing Feature Files

  1. Read the current feature file from spec/active/
  2. Modify scenarios to reflect the changed behavior
  3. Preserve existing @id: tags (IDs are stable across refactors)
  4. Add new @id: tags only for new scenarios
  5. Update scenario names and steps to match new behavior
  6. Remove scenarios for deleted behavior

Deleting Feature Files or Scenarios

  1. Remove the scenario (or entire file) from spec/active/
  2. Use git rm for tracked files
  3. Verify no other scenarios have @depends-on: references to deleted IDs

Phase 4: Verify and Summarize

After writing spec changes:

  1. Show summary of what was created/updated/deleted
  2. List the feature files and scenario counts affected
  3. Confirm the spec changes are ready for the implementation skill to proceed
  4. Stage spec changes — Run git add on any new or modified .feature files created/updated in this session. This ensures spec files are tracked even in new projects where spec/ didn't previously exist.

If this skill was invoked as a prerequisite to another skill (/fix, /refactor, /reel), indicate that spec authoring is complete and the calling workflow should continue.


Integration with Other Skills

This skill is the mandatory first step in behavior-driven workflows:

  • /design Phase 4 → invokes /bdd to write specs → then invokes /fix, /refactor, or /reel
  • /fix → invokes /bdd first (specify expected behavior) → then implements the fix
  • /refactor → invokes /bdd first (verify/update specs if behavior surface changes) → then refactors

Invocation is mandatory by default. Callers must always invoke /bdd rather than pre-deciding to skip. If no spec changes are needed, /bdd will assess that and proceed without modifications.

Opt-out: To disable mandatory BDD checks, add skip-bdd-checks: true to the project's CLAUDE.md or the developer's CLAUDE.local.md. When present, /fix, /refactor, and /design will skip the /bdd invocation.

The /bdd skill can also be invoked standalone to audit or update specs independently.


Example Usage

/bdd                                    # Prompts for scope
/bdd add user login feature             # Create new feature spec
/bdd spec/active/installer/             # Audit/update installer specs
/bdd remove deprecated v1 API           # Delete specs for removed behavior