A lot of “AI workflows” are just the same long prompt pasted into ChatGPT every week.
Take a Monday invoice check: search Gmail for the last seven days, skip anything already in Notion, add the new ones, then summarize what showed up. None of that is hard. The annoying part is that the path lives in prose. One week the model keeps your filter. The next week it broadens the search or invents a Notion property you never had.
If you already automate for a small team — or you’re the person who wired up Claude for the company — you’ve seen this. The assistant is good at improvising. It is unreliable at repeating a fixed procedure.
Prompts are a bad place to store procedures
Generic “connect AI to Gmail” catalogs help when the action really is generic. A Monday invoice ritual usually isn’t. It has opinions: which query, which database, what counts as a duplicate.
You don’t need a thousand actions for that. You need a small tool that already encodes the path, then a short ask in chat — something like “run the weekly invoice check.”
What to put in the tool
Build one tool (or a tiny set) that does the boring steps in order:
- Search mail with the query you trust.
- Write only new invoice rows to Notion.
- Return a short result the model can talk about.
Keep secrets in settings, not in the prompt. Connect the tool to ChatGPT, Claude, or Cursor the way you already work. The model can still draft a follow-up email or widen the window to two weeks — it just shouldn’t reinvent how invoices get logged.
In Smia that means designing the tool as a graph, hosting it on a product server URL, and letting the assistant call it. The product is the tool surface, not another essay about invoices inside the chat.
This is not “replace the Zap”
A Zap is the right shape when a trigger and an action are the whole job, and you don’t need to stay in the conversation. Here the job still lives next to chat: you want to ask a follow-up, change the date range, or draft a reply after the check runs.
So the unit is a tool the assistant can call when you ask — not a background agent you market as “scheduled AI.” You can still run it yourself on Monday, or attach the same tool from a Cursor Automation / Claude Routine later. What matters is that the procedure isn’t re-litigated in English every week.
A simple check
After a few Mondays: did you paste the long prompt again?
If yes, the workflow is still a conversation. If you only asked for the invoice check and got a dull, correct summary, the tool is doing its job.
Sharing the server URL with a coworker can wait. The immediate win is quieter: the same path, on purpose, every time.