Guide for subject matter experts iterating on agent skills, prompts, and test scenarios. Load when the user wants to modify agent behavior, create or edit skills, run tests, or compare agent runs. Always load when working in skills/ or scenarios/ directories.
You are helping a subject matter expert iterate on an AI agent's behavior. The SME understands the domain but does not write Python. Help them modify skills, run tests, and understand results.
The Markdown body — instructions, guidelines, guardrails, examples.
The tools_allowed list (choosing from available tools below).
The output_schema. The version number. Reference files in references/.
Do not change: the dispatch field or the "Last message contract"
section. These are engineering concerns. If the SME thinks a skill should
change from inline to subagent or vice versa, note it and ask engineering.
Create and edit freely based on domain knowledge.
orchestrator/, tools/, scripts/pyproject.toml, .envskills/<name>/SKILL.md. Make a focused edit — one change at a time.python scripts/run.py scenarios/<scenario>.yamlruns/<timestamp>/trace.md.
python scripts/compare.py runs/<before> runs/<after>git add skills/<name>/SKILL.md && git commit -m "skill(<name>): <description>"