- Home
- AI support
- Agent Workbench Setup
Agent Workbench Setup
Choose how you work, fill what you already know, download the files. Put them at the repo root. The coding agent asks once about anything still empty, writes the answer into docs/, and does not ask again.
What it is for?Hooks rebuild the local graph after commit, branch switch, merge, pull and rebase. They do not call a model.
Ask later leaves that row empty. The agent asks it once, writes the answer, and does not ask again. None means do not add that layer.
The zip is AGENTS.md plus docs/. Files are English. Drop them at the repository root.
What it is for?
This is not an application. It is a starter pack for a coding agent (Cursor, Codex, Claude, Copilot): AGENTS.md at the root, and a docs/ folder the agent is told to keep up to date. If you would not write those files yourself, this still gives you a usable base.
What you get
- AGENTS.md: standing rules (smallest change, no extra caution that hurts speed, related tests only, docs/ in English) plus how the agent should talk to you.
- docs/project.md and docs/stack.md: language, database, cache, Docker, queue, Graphify. Empty rows are asked once, then written, then not asked again.
- docs/TODO.md: checkboxes for later work, plain bullets for notes ("remember this").
- docs/testing.md, docs/adr/ with one example decision record (ADR) and a template. If Graphify is on, docs/graphify.md as well.
How the agent is supposed to use it
First chat: if a fact is still empty, one message of questions, then the agent updates the markdown. Later chats read the files. "We are adding Redis now" updates stack.md and an ADR; it does not restart the questionnaire. Using Docker is not the same as creating Compose files; scaffolding needs a separate yes.
It stays in the browser
Nothing is uploaded. The zip is built in this tab. Put the files in git so every agent session sees the same base.