SkillMedev/skills

REST API Design

APIs are products: the consumer's mental model, not your database schema, decides the shape. The costly mistake this skill prevents is shipping an inconsistent surface - mixed naming, wrong status codes, offset pagination on a growing table - that every consumer then hardcodes around, making it effectively unfixable once integrated.

Operating procedure

Work in this order. Resources come before methods because method semantics depend on what the resource is; error and pagination contracts come before the spec is written because retrofitting them breaks consumers.

Step 1: gather inputs

Collect before designing. Label any assumption as a guess.

1. The consumers: internal services, first-party clients, or external developers. External consumers raise the consistency bar - default to external if unknown.
2. The core nouns of the domain and which ones own which (drives nesting).
3. Expected collection sizes and growth. A list that can exceed ~10,000 rows rules out offset pagination.
4. Read/write ratio and latency expectations per endpoint.
5. Auth model (affects 401 vs 403 correctness).

Step 2: name resources

• Plural nouns: /users, /orders, /products.
• Nest one level for ownership: /users/{id}/orders. Do not nest deeper than two levels - flatten with filters instead (/orders?user_id=... beats /users/{id}/orders/{oid}/items/{iid}).
• No verbs in paths; the HTTP method is the verb. For genuine actions that fit no CRUD verb (e.g. sending an invoice), model a subresource POST: POST /invoices/{id}/send-attempts.
• Consistent case: kebab-case for multi-word path segments, snake_case for JSON fields - pick once, apply everywhere.

Step 3: assign methods and status codes

• GET: read, idempotent, no request body. 200 with body.
• POST: create. Return 201 plus a Location header pointing at the new resource. Return the created representation in the body so the client skips a follow-up GET.
• PUT: full replace, idempotent. Missing fields mean "clear", not "keep".
• PATCH: partial update (JSON Merge Patch or JSON Patch - state which in the spec).
• DELETE: remove, idempotent, returns 204. Deleting an already-deleted resource returns 204 or 404 - pick one and document it.
• Errors: 400 for malformed input, 401 for missing/invalid credentials, 403 for valid credentials without permission, 404 for absent resources (also use 404 instead of 403 when revealing existence is a leak), 409 for state conflicts, 422 for well-formed but semantically invalid input, 429 for rate limits (with Retry-After), 500 for server faults - never expose stack traces or internal identifiers.

Step 4: fix the error contract

One shape for every error, machine-readable code first:

{
  "error": "validation_error",
  "message": "Email is required",
  "field": "email",
  "request_id": "req_abc123"
}

The request_id must be logged server-side so support can correlate. Human message is for developers reading logs, never for end-user display.

Step 5: fix pagination and filtering

• Cursor-based pagination for any collection that grows: return an opaque cursor and has_more flag. Offset pagination is acceptable only for small, bounded, admin-facing lists - it degrades linearly and skips/duplicates rows under concurrent writes.
• Default page size 25, maximum 100. Reject page sizes above the max with 400, do not silently clamp.
• Filtering and sorting via query params: ?status=active&sort=created_at:desc. Document which fields are filterable; reject unknown filter params with 400 rather than ignoring them (silent ignoring hides client bugs).

Step 6: apply performance and payload budgets

• p95 latency budget: 300 ms for interactive endpoints, 1 s for reports/exports. If an operation cannot meet 1 s, make it asynchronous: 202 Accepted plus a status resource to poll.
• Response payload budget: keep typical responses under 100 KB; anything routinely above 1 MB needs pagination, field selection (?fields=), or a dedicated export flow.
• Version breaking changes in the URL path (/v1/). Additive changes (new optional fields) do not require a version bump. Hand the deprecation timeline itself to api-versioning-strategist.

Worked contrast

Bad:

POST /getUserOrders?userId=42&page=3
→ 200 { "data": [...], "error": null }

Verb in the path, GET semantics on a POST, offset paging, 200 wrapping an error field so clients must parse the body to know if the call failed.

Good:

GET /users/42/orders?status=shipped&limit=25&cursor=eyJpZCI6OTh9
→ 200 { "orders": [...], "next_cursor": "eyJpZCI6MTIzfQ", "has_more": true }
→ 404 { "error": "not_found", "message": "User 42 does not exist", "request_id": "req_9f2" }

Noun path, ownership nesting, cursor paging, status codes carry the outcome.

Deliverable

Produce an endpoint spec containing: the resource table (path, methods, status codes per method), the shared error shape, the pagination contract (cursor format, defaults, max), the filtering whitelist, and the latency/payload budget per endpoint.

Do NOT

• Do not mirror the database schema in the API - consumers get coupled to internals you will want to change.
• Do not return 200 for failures with an error body; clients and monitoring both key off status codes.
• Do not use offset pagination on unbounded collections - it duplicates and skips rows under writes and degrades linearly.
• Do not nest paths deeper than two levels; deep nesting encodes one access path and forbids all others.
• Do not silently ignore unknown query parameters - a typoed filter that returns unfiltered data is a data-leak class bug.

Quality bar

• A consumer could implement a client from the spec without asking a single question.
• Every endpoint states its status codes, including error cases.
• All collection endpoints share one pagination contract; all errors share one shape.
• Identical concepts have identical names across every path and payload.
• Each endpoint has an explicit latency budget and the >1 s ones are async.