| name | new-feature |
| description | Pre-flight checklist that must complete before writing the first line of code for any new feature. Load this skill when starting a new endpoint, service method, or integration. Prevents the most common AI-coding failure mode — writing code before defining layer boundaries, which leads to features that work today and break in 3 months. |
new-feature
Do not open any file for editing until all 5 steps below are completed.
Step 1 — Define the API contract
Write the contract in plain text before any code. Answer these 4 questions:
Endpoint: POST /generate/image
Request: { prompt: str, model_id: str, width: int, height: int }
Response: { task_id: str, status: "queued" }
Side effect: creates a task, holds credits
If you cannot answer all 4 in 2 sentences, the feature scope is unclear. Stop and clarify with the user.
Step 2 — Identify layers touched
Use this decision matrix to know which files you will create or modify:
| What the feature needs | Layers touched | Files |
|---|
| New HTTP endpoint only | Router | routers/<domain>.py |
| Endpoint + business logic | Router + Service | + services/<domain>/ |
| Logic + database | Router + Service + Repo | + repositories/protocols.py, repositories/sqlalchemy/<domain>.py |
| Logic + external API | Router + Service + Provider | + providers/<name>.py |
| Logic + credits charge | All of the above + Credits | + services/credits/user.py (read-only — single writer!) |
Write the list: "This feature touches: Router, Service, Repo."
Do not touch more layers than listed. If you discover you need more, stop and re-plan.
Step 3 — Define the test scenario
Write the test scenario before code. Two scenarios minimum:
These become your actual pytest tests in tests/unit/ or tests/integration/.
Step 4 — Build bottom-up
Always build in this order. Never start from the router.
1. Repository (if new data access needed)
└── Add method to CreditsRepoProtocol
└── Implement in SQLAlchemyCreditsRepo
└── Implement in FakeCreditsRepo (tests/conftest.py)
2. Service
└── Accepts repo via Protocol (never imports SQLAlchemy)
└── Contains all business logic
└── Raises domain exceptions (ValueError, not HTTPException)
3. Router
└── Thin: validate schema → call service → return response
└── Maps domain exceptions to HTTP codes
└── No logic beyond 10 lines per endpoint handler
4. Schema
└── Request and response Pydantic models
└── Goes in schemas/<domain>.py, never inside routers/
Code template for a new router endpoint:
@router.post("/<resource>", response_model=ResourceResponse, status_code=202)
async def create_resource(
body: ResourceRequest,
user_id: str = Depends(get_current_user_id),
service: ResourceService = Depends(get_resource_service),
) -> ResourceResponse:
try:
result = await service.create(user_id=user_id, **body.model_dump())
except ValueError as e:
raise HTTPException(status_code=400, detail=str(e))
return ResourceResponse.model_validate(result)
Step 5 — Verify principles before committing
Run through this checklist before calling the feature done:
bash scripts/lint-architecture.sh
grep -rn "AsyncSession\|from sqlalchemy" app/services/<new-service>/
grep -n "get_current_user_id" app/routers/<new-router>.py
grep -n "user_id" app/repositories/sqlalchemy/<new-repo>.py
| Principle | Check |
|---|
| A1 — folder if 3+ endpoints in domain | ls app/routers/<domain>/ |
| A7 — file under 400 LOC | wc -l app/services/<file>.py |
| B1 — no SQLAlchemy in service | grep above |
| B2 — service via Protocol, not concrete | constructor accepts Protocol type |
| B10 — credits only via CreditsUserService | no direct repo.hold() calls |
Verification
The skill was applied correctly when: