Starbase / 02 RecipeAgents

RecipeAgents / Guide

the pillar guide

How to build an AI agent (the small, useful kind)

Most "how to build an AI agent" guides start with frameworks and vector databases. This one starts with your inbox. An agent, for our purposes, is a trigger, a model that makes a decision, and an action, plus a person checking it until you trust it. You can build one this afternoon in Zapier, Make or n8n without writing code.

1. Pick one boring, repeated job

Good first agents have three things in common: the input arrives on its own (an email, a form, a file), the decision is small (which label, which category, which fields), and a mistake is cheap. Sorting invoices fits. Replying to angry customers on your own doesn't, at least not yet. Browse the collections and pick one from Personal admin if you're nervous, because those only touch your own stuff.

2. Write down the input and the output

Before you open any tool, write one real example of what comes in and what should come out. Every recipe here has a sample run, written for illustration, that shows the format. If you can't write the output by hand, the model can't produce it reliably either.

3. Write the prompt like a form, not a chat

Ask for JSON with named fields, give the allowed values ("one of: PRICE, TIMING, COMPETITOR…"), and say what to do when information is missing ("null, never guess"). That last rule prevents most of the bad outputs we describe in the failure-mode sections.

4. Build it: trigger → AI step → action

In n8n that's three nodes, and you can download ready-made files from n8n templates. In Zapier it's a trigger, a ChatGPT step and an action, and each recipe has a blueprint file you follow by hand (Zapier has no import format). Our inbox triage walkthrough shows every setting.

5. Test on 20 real examples

Put inputs and outputs side by side in a sheet and mark each one right or wrong. Fix the pattern, not the single case: one new line in the prompt per type of mistake. Keep drafts and approvals in place until the error rate is one you'd accept from a new hire.

6. Then let it run, and watch it

Tag every AI-made action so you can review a sample weekly. Watch costs too: each AI step is billed by the model provider and each run by the automation tool. Filtering inputs before the AI step is the cheapest optimisation there is.

SPONSOR SLOT · AVAILABLEFeature your product here

Every recipe, by job

Sales

Ops

Support

Content

Personal admin

Bigger builds

Questions

Do I need to code to build an AI agent?
No. Every recipe here uses Zapier, Make or n8n with a prompt. Code helps later, but isn't needed to start.
How long does it take?
Our setup estimates per recipe range from about 15 minutes to 2 hours. Those are estimates, not measurements.
Which tool should I use?
Zapier is quickest for straight-line flows. Make shows branching more clearly. n8n can be self-hosted, and we provide importable n8n files.
Newsletter

One small automation every Friday

Each Friday email has one recipe, the prompt for it, and the way it broke when we wrote it up. It's free, and every email has an unsubscribe link.