A tool sounds like a big thing. It usually isn't. Strip away the marketing and most "AI tools" are exactly one thing: a good prompt, wrapped so you can run it again on new input without retyping it. That's the whole trick. Once you see it, you'll build them constantly.
You already have the two hard parts from earlier in this path: a sharp prompt and, when quality matters, a loop that checks its own work. A tool is just those, saved and pointed at a stream of inputs instead of one.
Step 1 — find the repetitive task
Don't start with "what could AI do." Start with what do I retype every week? Classifying support emails. Turning meeting notes into actions. Drafting replies in your voice. The best first tool automates a thing you're already doing by hand, so you can tell instantly whether it's any good.
Step 2 — write the prompt once, properly
Make it general enough to handle any input of that kind, not just today's. Name the role, the constraints, the exact output format, and — critically — leave a slot where the input goes:
You are my support triage assistant. For the message below, return ONLY JSON:
{"category": "billing|bug|feature|other", "urgency": "low|med|high", "draft_reply": "..."}
Choose only from the listed values. Keep the reply under 60 words, plain and warm.
MESSAGE:
<<< input goes here >>>That <<< input >>> slot is what turns a one-off prompt into a tool.
Step 3 — wrap it. Two ways.
No-code. Save the prompt as a reusable step you can fire on new input — that's exactly what a loop tool like ailoop is for: you define the objective, the metric that judges each result, and the boundary, then run it on input after input without rebuilding it. (Full disclosure: ailoop is ours.) If you want quality control built in, this is the fastest path to a tool that checks itself.
Code. If you're comfortable with a few lines, the whole tool is: take input → drop it in the prompt → send to a model API → return the output.
// the entire "tool", in shape:
const prompt = TEMPLATE.replace("<<< input >>>", userInput);
const result = await callModel(prompt); // your API of choice
return JSON.parse(result);That's it. Everything else — a web form, a spreadsheet button, a Slack command — is just a different doorway into those three lines.
The point
A tool isn't a model and it isn't magic. It's the moment you stop having a conversation with AI and start running it. Build one for the most annoying repetitive thing you do, use it for a week, and you'll feel the shift the rest of this site is about: from typing prompts to designing the machine that types them for you.