Agents

AI-powered development with Takeout

Takeout treats agent-readability as a first-class concern. The codebase is structured so AI coding agents can understand patterns, find examples, and verify their work without constant hand-holding.

The result: describe what you want, and agents can often implement features end-to-end—data layer, UI, permissions, tests—using the patterns already in place.

Skills

Skills are instruction sets that teach agents how to work in this codebase. They live in .claude/skills/ and map to docs in docs/ and packages/docs/.

When an agent starts a session, it loads the core guidelines automatically. When working on a feature, it can pull in relevant skills—say, takeout-zero when adding a synced data type, or takeout-triggers when setting up database automation.

Each skill includes:

• the pattern and why it exists
• working examples from the codebase
• common mistakes to avoid
• verification steps

This means agents don’t guess at conventions. They look up the skill, follow the pattern, and produce code that fits.

Available Skills

Over-Zero

The over-zero package structures your data layer in a way that’s easy for both humans and agents to understand. Models define schema, permissions, and mutations together. Queries are plain functions that become synced automatically.

When an agent needs to add a new data type, the pattern is clear:

1. create the model in src/data/models/
2. add queries in src/data/queries/
3. run bun zero/generate to get types
4. use the generated hooks in your UI

There’s no ambiguity about where things go or how permissions work. Agents can implement features involving offline sync, optimistic updates, and row-level security without needing to figure out the architecture.

See the Zero docs for the full picture.

Scripts

The tko CLI gives agents (and you) a discoverable set of operations. Run bun tko to see what’s available:

bun tko              # list all scripts
bun tko check types  # run type checking
bun tko web build    # build for web

Scripts are organized by category—check/, ci/, web/, db/—so agents can find the right tool without asking.

Verification

Session hooks ensure agents know to verify their work. The core guidelines load automatically and include explicit verification steps—run bun check:all for quick checks, bun run ci --dry-run for the full pipeline (see Releasing for details).

bun check:all        # types, lint, tests
bun run ci --dry-run # full CI pipeline locally

This catches issues early. When an agent says “done,” you can trust the tests passed and types check.

What This Enables

A typical agent workflow:

1. You describe a feature (“add a favorites system”)
2. Agent checks relevant skills (zero, triggers, routes)
3. Implements using established patterns
4. Runs verification
5. You review working code

The structure removes guesswork.