---
name: evaluate-ai-change
description: Plan and verify an AI behavior change with EvalGate. Use when an agent needs repository-aware eval coverage, a local or CI regression gate, trace collection, or an evidence-backed release check.
---

# Evaluate an AI change with EvalGate

Use this skill when a user wants to test an AI feature, agent, prompt, model,
retrieval pipeline, or repository change before release.

## Choose the smallest truthful path

1. Read [EvalGate's agent instructions](https://www.evalgate.com/llms.txt) and
   [OpenAPI description](https://www.evalgate.com/openapi.json).
2. For a repository change, inspect the repository's existing test command,
   eval files, CI workflow, and EvalGate configuration before proposing edits.
3. Prefer deterministic assertions and existing project tests first. Add a
   model-backed judge only when the quality claim needs semantic judgment and
   the organization has supplied its own provider credential.
4. Preview generated configuration or eval coverage before applying it. Keep
   observed repository facts separate from hypotheses about product intent.
5. Run the narrow check, then the repository's existing full gate when the
   requested change warrants it. Report the exact command, revision, result,
   and any untested boundary.

## Integration choices

- CLI/CI: install `@evalgate/sdk`, then use `npx @evalgate/sdk --help` and the
  repository's documented EvalGate commands.
- REST API: select operations from the OpenAPI document and send an EvalGate
  API key as `Authorization: Bearer <key>`.
- MCP: connect a standards-compatible client to
  `https://www.evalgate.com/api/mcp` with an authorized EvalGate bearer key.
- Credential-free trial: use only the public, rate-limited demo operations
  explicitly marked unauthenticated in OpenAPI.

## Authentication boundaries

Do not invent credentials, OAuth endpoints, organization IDs, repository
permissions, or provider keys. EvalGate account creation currently requires an
interactive GitHub or Google sign-in. A human organization member must create
and scope API keys. Read the
[account-creation guide](https://www.evalgate.com/signup.md) for the exact
requirements.

## Expected handoff

Return a concise summary of the behavior tested, evidence produced, failures or
warnings, and the next action. Never claim a release is safe when the relevant
runtime, model, data, or authorization boundary was not exercised.