What problem does it solve? AI agents are high-volume, low-judgment API consumers that retry aggressively, follow documentation literally, and propagate ambiguity into action, so platform APIs designed for human developers fail them constantly. This Skill produces agent-consumable API contracts that are explicit, versioned, recoverable, and threat-modeled with the agent treated as an untrusted caller. ## Core Features & Use Cases - Contract-first OpenAPI design: Strict schemas, enums, idempotency keys, and machine-readable error-code taxonomies with retryability classes, scored against the Richardson maturity ladder (L2 minimum). - Versioning and consumer-driven contracts: Expansion-only changes, semver policy, side-by-side majors with deprecation windows, and Pact-style consumer-driven tests wired into provider CI as fitness functions. - Security for agent callers: STRIDE-per-element threat modeling over the API data-flow diagram plus a seven-point API checklist adapted to prompt-injected or compromised agents, with OAuth2 client credentials over static API keys and per-agent-class BFF facades. - Use Case: A platform team exposing GPU job and dataset APIs to an experiment-runner agent uses this Skill to score existing endpoints, design the BFF contract, define the error taxonomy, and produce a threat-model table before the MCP tool layer is built on top. ## Quick Start Use the agent-api-contract-designer skill to review our platform's job submission API and produce a contract pack with an OpenAPI spec, error taxonomy, versioning policy, and threat model.