Project Onboarding

project-onboarding

When starting work on a new codebase, the agent has zero context. This leads to hallucinated conventions, ignored project patterns, and wasted turns re-explaining the stack. This skill crawls a target directory, infers everything it can automatically, and generates a structured PROJECT.md that the agent loads for all future work on that project.

When to invoke

Invoke this skill when:

• Starting work on a codebase for the first time
• The agent begins making assumption errors about the project structure
• A new team member (agent) is added to an existing project

Onboarding protocol

Step 1 — Crawl the directory Scan the project root for:

• package.json, pyproject.toml, Cargo.toml, go.mod, pom.xml, etc. → infer tech stack
• Makefile, justfile, scripts/ → infer build/test commands
• .github/workflows/ → CI commands and quality gates
• Test file patterns (__tests__/, spec/, *_test.go) → testing conventions
• CONTRIBUTING.md, DEVELOPMENT.md, docs/ → explicit conventions
• src/, lib/, app/, internal/ → project structure

Step 2 — Infer key invariants Ask: what would break the project if violated? Examples:

• "All API responses must include a requestId field"
• "Never commit secrets — use .env.example"
• "All new routes must have a corresponding integration test"

Read existing test names, CI checks, and lint config to infer these.

Step 3 — Generate PROJECT.md Write PROJECT.md to the project root (or ~/.openclaw/workspace/<project-slug>.md):

markdownCopy code
# Project: <name>
## Stack
## Build & Test Commands
## Project Structure
## Key Conventions
## Things the Agent Must Never Do Here
## Known Gotchas

Step 4 — User validation Show the generated file and ask: "Does this look right? Anything missing or wrong?" Revise based on feedback.

Step 5 — Register in state Record the project path, PROJECT.md location, and onboarded date in state so the agent auto-loads context on future sessions.

Auto-load on session start

When working in a directory that matches a registered project, the agent should:

1. Read PROJECT.md into context before any work
2. Skip Step 1–4 (already onboarded)
3. Run python3 onboard.py --refresh monthly to catch drift

Companion script

python3 onboard.py --scan <path> performs Step 1 and outputs a structured JSON summary of detected stack/conventions — the agent uses this to populate Step 3.