Epilogue · Design your own AI button

You've seen five buttons. Here's the thinking behind all of them, so you can design the sixth for your own app.

Start from a chore, not from AI

Every button here replaces a small, boring task someone already does by hand:

Look through your app for the same kind of thing: copy-paste, retyping, reading to find one detail, writing a similar text again. That's where an AI button feels like a normal button.

Five questions for any AI button

1. What exactly goes in, and what comes out? Write the schema first. If you can't, the feature isn't clear yet.
2. What does "I don't know" look like? A null, a "not sure" flag, or "send to a person". Design it on purpose.
3. What does code check? Formats, ids, counts, "is this quote really in the source?" Anything you can check without AI, check without AI.
4. Who decides? The model proposes. Code or a person makes it happen. Nothing irreversible without a click.
5. What happens when it fails? The page should work without it.

Pick the smallest shape that works

From simplest to most complex:

1. One call → Tidy up, Fill from email
2. Route, then specialist → Route it
3. Fan out, then combine → This week's themes
4. Write, check, revise once, hand over → Draft and check
5. An agent loop → only when you can't write the steps down in advance (see The Agent Loop)

Most features stop at 1 or 2. Go one step up only when an eval shows it helps.

Where to go next

• The Agent Loop, the companion to this repo: build a mini coding agent, for when you really do need step 5.
• Anthropic, Building effective agents. The patterns here, explained by the people who make the model.
• Claude docs on structured outputs, for the schema details.

Build one button this week. Then tell me about it on X: @georgehanu. I'd love to see what chore it replaced.