1# Evals
2 
3Manual evaluations for the mcp-builder skill.
4 
5## Format
6 
7Each eval file is a JSON array:
8 
9```json
10[
11  {
12    "query": "User input to test",
13    "expected_behavior": "OUTCOME. What the response should do."
14  }
15]
16```
17 
18## Running Evals
19 
20In Claude Code:
21 
22```
23Run the evals in evals/<reference>.json. For each query, spawn a Sonnet agent with the mcp-builder skill context and compare the response against expected_behavior. Report pass/fail for each.
24```
25