Set up a project for agents

Give the target repository an AGENTS.md holding what only that repo knows, and install the skills its stack needs.

Every step below ends on a stated completion criterion. Reach it before moving on.

What earns a place in the file

AGENTS.md holds what only this repository knows. Everything else already has a home that this skill leaves alone: the user's own global instructions, and the installed skills. Skill selection stays in the skill system; AGENTS.md carries no skill catalogue or mandatory-reading list.

So a line earns its place only when it survives two questions:

• Does something else already carry it? A standard restated here becomes a second copy that drifts, and the copy is where qualifying clauses get dropped. Leave it out rather than naming or quoting its skill. Testing is the clearest case: each framework has its own runner and layering, so the stack skill owns that workflow.
• Is it recoverable once violated? A guardrail is the exception to that principle. It protects against a loss no later reading repairs, so its text goes in full rather than behind a pointer.

1. Explore

Read the repository before asking anything. Ask only what reading cannot settle.