# bitby: Hand real work to your agents.

Copy-paste templates and an evidence-graded index of what works with AI agents. Tested on my own projects first. Free.

Practical ways to hand real work to AI agents: plain-text templates and an evidence-graded index of what works, tested on my own projects first, free to copy. Written for people and readable by agents.

## 01 Copy: six templates, any agent

### Handover checklist

When: Before an agent stops. The next session starts warm instead of re-reading everything.

```text
## Handover
Goal:            <one sentence>
Done since last: <2-5 bullets>
Next step:       <the exact next action>
Open questions:  <for a human>
Don't touch:     <areas to leave alone>
Checked:         <what was actually run or tested>
```

[Handover checklist as markdown](/templates/handover.md)

### Fresh-session restart

When: Going in circles. A clean start with a good handover beats pushing on with a muddled context.

```text
You are starting fresh on this task.
1. Read HANDOVER.md and nothing else yet.
2. Restate the goal and the next step in two lines.
3. List anything in the handover you can't verify.
4. Only then start, with the next step.
```

[Fresh-session restart as markdown](/templates/restart.md)

### Approval rules

When: Before unattended runs. Decide once what needs a human, instead of per prompt. A starting point to adapt.

```text
Just do:    read, search, draft, run tests, work on a branch
Ask first:  anything that sends, spends, deletes or publishes
Never:      handle credentials, touch production data,
            change these rules
```

[Approval rules as markdown](/templates/rules.md)

### Task brief

When: Every task. Most bad agent runs start with a vague brief.

```text
Goal:        <what should be true when you're done>
Context:     <links, files, why it matters>
Constraints: <what not to change, time box>
Done when:   <a check anyone can run>
Report:      <what to tell me, in how many lines>
```

[Task brief as markdown](/templates/brief.md)

### Definition of done

When: Every project. Agents stop when they think they're done. Tell them what done means.

```text
Done means all of these:
- the check in the brief passes
- the handover is updated
- one paragraph: what changed and why
- nothing left half-finished on the branch
```

[Definition of done as markdown](/templates/done.md)

### Stop rules

When: Anything hands-off. Knowing when to stop matters more than knowing how to continue.

```text
Stop and ask if:
- the same fix has failed twice
- the task has grown past its brief
- you'd need a new dependency or permission
- you're about to guess at something a human knows
```

[Stop rules as markdown](/templates/stop.md)

## The Frontier Index

53 rules for working with AI agents, graded by evidence (A/B/C/P), with sources. Read: [/frontier.md](/frontier.md). Skill download: [/frontier/skill.md](/frontier/skill.md).

## 02 Try

- **HOLD**: Playable now. A scam-spotting party game for a room: phones and one big screen. Fifteen minutes. Designed and built by agents in a week. Play: https://play.bitby.ai

## FAQ

**Which agents does this work with?** Any agent that reads plain text. Nothing assumes a particular tool or setup.

**Is it free?** Yes. Copy anything, no sign-up.

**Can my agent read this site directly?** Yes. Point it at llms.txt or the template catalogue (JSON). Every page also has a markdown version.

**Who checks what goes out?** A human reviews every template and every post before it is published. Nothing goes out automatically.

**Where were these tested?** On my own projects first: this site, the game above and the everyday work I hand to agents. Nothing goes up until it has been used for real.

**What happens to what I send in the form?** It's kept on this site's own server so I can reply to you. No mailing list, no tracking cookies, never shared.

## Ask

Tell me one job you'd hand to an agent. I read every one and reply in person. The best ones become write-ups here. Use the form on the home page.
