Auto-Coder docs
Work with the agent like a development system, not a magic button.
A good autonomous workflow starts with explicit project context, a bounded goal and a review step. These guidelines describe the intended operating model without assuming a specific private build or download package.
Mental model
Auto-Coder combines project context with agent execution. The useful loop is: inspect → plan → change → run → report → review.
Agent output can be wrong, incomplete or unsafe. Treat generated changes as proposed engineering work until tests and review say otherwise.
1. Start from a real project
Give the system a bounded codebase and make the project’s conventions visible. Useful context includes:
- README / architecture notes
- package and lock files
- tests and quality gates
- environment-variable documentation without secrets
- deployment expectations
2. Write a task the agent can verify
Prefer a concrete goal with acceptance criteria over a vague instruction.
Goal: Add password-reset email flow.
Constraints:
- keep existing auth provider
- no service-role key in browser
- mobile UI must work
Verify:
- unit tests
- reset-link happy path
- expired-token error state3. Review the plan and the diff
Before considering work finished, inspect changed files, commands run, test output, unresolved assumptions and any dependency changes.
4. Pick the right interface
Desktop
Use when you want the editor, terminal and agent workflow in one interactive workspace.
Web
Use when browser access and streaming project assistance are more important than local editor depth.
CLI
Use when you want headless or automation-friendly agent execution against a project path.
5. Production checklist
- Run the project’s test and lint gates.
- Review auth, secrets, payments and destructive actions manually.
- Test the main user flow in a real browser.
- Check mobile layout and keyboard navigation.
- Verify deployment health and logs after release.