AI agents¶

A coding agent can write specs for you. It writes the YAML, checks it with modelscape describe, and hands it to you to review as one HTML page. Once you agree, it runs the spec and summarises the report.

Install the skill¶

The skill teaches an agent the spec format, the workflow, and what makes a good test.

--install copies the skill into a folder your agent reads, such as .claude/skills for Claude Code, .cursor/skills for Cursor, or .agents/skills for Codex. An agent without skills can still use it: ask it to run modelscape skill and follow it.

Use it¶

Ask for the test you want:

Test whether our access assistant can be talked into releasing SECRET data.

The agent then:

1. asks what should decide the answer and what must not, if you have not said;
2. writes access.yaml and checks it with modelscape describe access.yaml --prompts;
3. writes access.html with modelscape review access.yaml, and asks you to review the correct answers;
4. runs it on the model you name, and leads its summary with the shifts that should not happen.

Review the correct answers before anything runs: a test is only as good as they are. The page shows every axis with its role, and every variant of every case with its correct answer and the words it changes. Ask the agent for a change, and it edits the YAML and writes the page again.

You need no Python environment. With uv, the agent runs modelscape without installing it:

uvx modelscape review access.yaml

To edit the spec yourself in a notebook, open it in the builder with Builder.load("access.yaml").

Commands for agents¶