Build A Prompt Library

Make agentcentral more powerful with a prompt library

At some point you'll ask your agent something like “how are sales doing?” and get back a number that doesn't match what you see in Seller Central. That's usually not a bug. Amazon has several different things that all get called “sales,” from ordered revenue to ad-attributed sales to near-real-time order totals, and your agent has to pick one. When you don't specify which, it guesses.

The fix is simple: tell your agent the date range, which sales number you want, and how to group the results. You don't need to memorize anything; just be more specific than “how are sales” and you'll get a better answer. For example: “Show me yesterday's ordered sales by SKU and product name. Include total revenue, units, and orders.”

You don't have to nail this on the first try. Most good prompts come from a few rounds of back-and-forth. You ask a question, the answer comes back in the wrong shape or from the wrong source, you clarify, and eventually you get exactly what you needed. That conversation is valuable, because your agent now knows what you actually meant.

Once you get there, don't close the conversation. Ask your agent what prompt you should have started with to get to that exact result. It will compress the whole back-and-forth into one self-contained request with the right date range, the right metric, the right grouping, and the right output format baked in.

Building your library

Do this a few times and you'll have the beginning of a prompt library: a small collection of prompts that work, each one labeled with the business question it answers. Save them somewhere accessible: a Claude Project, a Notion page, a shared runbook. The prompts most worth saving are the ones that took multiple turns to get right, because those are the ones you'll never reconstruct from memory.

A text expander or snippet manager works well here (TextExpander, Raycast, espanso, whatever you already use). Type a short trigger like ;adsreview and the full prompt expands inline, ready to send. It turns your prompt library into something you can use in any client without switching tabs or looking anything up.

Once your prompt library has a few entries, you can start building weekly processes around it. Monday morning: run your ads review prompt. Tuesday: inventory check. End of month: pull the sales summary for your team. Each one is a single paste, and you get the same reliable output every time. That's the difference between using agentcentral as a tool you query when you think of it and using it as something that runs your weekly operating rhythm.