Configure a project
agentbriefer init is the only interactive project command. It turns a guided
conversation into a typed YAML configuration.
Before the prompts
Run it from the project root:
agentbriefer init
If agentbriefer.yaml already exists, the wizard asks before overwriting it and
defaults to leaving it untouched. Agentbriefer then detects a supported manifest
and loads any saved developer profiles.
Stack answers
Detected values prefill language, framework, database, package manager, testing tools, and key dependencies. They remain editable because repositories can be polyglot or use a tool the detector does not recognize.
Comma-separated testing tools and dependencies are normalized into YAML lists. Optional fields can be left blank.
Behavioral policy
The structured settings are deliberately concrete:
| Setting | What it changes |
|---|---|
| Developer style | Scope, tradeoffs, teaching, maintainability, or security posture |
| Explanation style | Short, beginner-friendly, detailed, or tradeoff-focused reporting |
| Project type | Backend, frontend, CLI, full-stack, library, or docs priorities |
| Security level | Caution around auth, secrets, validation, permissions, and data access |
| Testing level | When tests are optional, practical, or required |
| Dependency policy | Whether agents may add, must explain, or must ask first |
| Architecture style | Flat, lightly layered, feature-based, or advanced existing patterns |
Every value is rendered with an agent-facing description; generated files do
not rely on an unexplained label such as practical or strict.
Custom instructions
Custom instructions hold project-specific guidance that does not fit a dropdown:
Run `make fixtures` before integration tests. Never edit generated files under
src/protocol/. Keep public error codes backward compatible.
The wizard opens $VISUAL or $EDITOR when available, falling back to the
platform editor. Closing without saving keeps the previous value.
Stop rules
Stop rules describe boundaries where an agent must pause or avoid continuing:
stop_rules:
- Stop before changing CI/CD configuration.
- Ask before modifying a public API.
- Do not access production credentials.
When no stop rules are configured, generated output still tells the agent to stop after the requested change is complete and verified.
Review and save
The final screen shows every answer. Select a field to edit it using its current value as the default, or select Save and finish. The wizard writes YAML only after this review.
init starts with no installed skills. Add them after saving with
agentbriefer skill list and agentbriefer skill add <id>.