| name | project-start |
| description | Orient an agent to a new or unfamiliar repository and establish a focused execution context before substantial work. Use when starting a project, entering an existing codebase, planning the first feature, creating repository instructions, or when output quality suggests missing project context. |
Project Start
Purpose
Build the smallest reliable project map needed to make correct early decisions without loading the entire repository or inventing conventions.
Inputs
- User outcome, repository path, current branch/status, and authorization boundary.
- Root instructions, README, manifests, configuration, CI, and architecture docs.
- Relevant source, test, schema, and deployment entry points.
- Existing issue/spec and known constraints.
Scope Boundary
Use this for orientation before substantial work. Skip it for a narrow task when the relevant files, command, and contract are already known.
Process
- Inspect repository status and locate instruction files without modifying anything.
- Read root documentation, manifests, commands, CI, and the directory structure.
- Identify languages, frameworks, package boundaries, entry points, tests, generated code, and public contracts.
- Trace one representative workflow related to the user's next task.
- Record a compact project map: purpose, architecture, commands, conventions, boundaries, risks, and unresolved questions.
- Separate evidence from inference; verify commands rather than assuming common defaults.
- Define the first scoped outcome and non-goals.
- Add or update project instructions only when requested or when they are a normal deliverable of project setup; keep them concise and repository-specific.
- Choose the next workflow skill based on the task, not the category list.
Constraints
- Do not read every file before useful work can begin.
- Do not create a new architecture when the repository already has one.
- Do not run destructive setup, migrations, deployments, or dependency upgrades during orientation.
- Do not paste secrets or local environment values into project context.
- Do not create generic rules files that repeat platform behavior.
Verification
- The project map cites real files and verified commands.
- The relevant workflow and ownership boundary are understood.
- Generated, vendor, secret, and user-owned files are identified before edits.
- The first outcome is testable and narrow enough to review.
- Remaining unknowns are genuinely material rather than discoverable routine facts.
Failure Modes
- Starting implementation from the README alone.
- Loading so much context that the current task becomes unclear.
- Assuming package-manager, test, or deployment commands.
- Writing a generic
AGENTS.md that conflicts with local conventions.
- Treating an inferred architecture diagram as authoritative.
Examples
Existing monorepo: identify package ownership, shared contracts, affected tests, and the command for the target workspace before editing.
New project: define audience, V1 outcome, non-goals, technical constraints, verification path, and a small initial structure before scaffolding.
Sources
See ../../sources/project-start.sources.md.