init

init

Turns a project with no shared AI config into one where a single set of specs drives every tool.

Steps

1. Confirm the CLI is present: agnostic-ai --version. If missing, run the install skill first.

2. Check whether the project already has AI config worth keeping: CLAUDE.md, AGENTS.md, GEMINI.md, .cursor/rules/, .github/copilot-instructions.md. If any exist, stop and use the import skill instead; it captures them as specs rather than overwriting them.

3. Scaffold:

   agnostic-ai init --all      # every target, no prompt
   agnostic-ai init --demo     # plus example specs to start from
   

   init alone opens a target picker when stdin is a TTY. In a non-interactive session, pass --all or pipe a list: echo "claude,codex" | agnostic-ai init.

4. Ask which AI tools the team actually uses, then narrow targets: in agnostic-ai.yaml to those. Fewer targets means fewer generated files to review.

5. Write the project's real conventions as rules under .agnostic-ai/rules/, one file per rule. Keep each rule to a single concern with a description that says when it applies.

6. Run agnostic-ai sync, then show the user what appeared with git status --short.

Notes

• init writes gitignore.enabled: true by default: generated files stay out of git and each contributor runs sync. Pass --gitignore=false to commit them instead, which suits a team where not everyone has the CLI.
• One spec per file. agnostic-ai new rule <name> (also agent, skill, hook, mcp) scaffolds the right frontmatter for the kind.
• Never hand-edit a generated file such as CLAUDE.md or .cursor/rules/*.mdc. The next sync overwrites it. Edit the spec under .agnostic-ai/ instead.