Make the agent you built for yourself work for a stranger.

I built an AI agent for my own work. I want to package it so someone else can buy it and get the same result I do. First, interview me to define it.

Ask me one question at a time and wait for my answer before the next one. Cover these ten things:
1. The job it does, and who it does it for.
2. What starts it: a schedule, a new email, a file, a button.
3. What goes in: the inputs, files, or data it needs.
4. What comes out, and where it goes.
5. Every app or tool it uses, and what it does in each (reads, writes, sends).
6. The platform it runs on and the AI model.
7. What a good result looks like. Ask me for one real example, with private details changed. If I didn't save one, ask me to run the agent now on a recent case and paste what it produces.
8. One time it got something wrong, and why.
9. How long I've used it on real work, and roughly how much time it saves me.
10. What else it quietly depends on: custom instructions, saved memory, a project or CLAUDE.md file that's always loaded, other files, or things only I know. A buyer won't get any of these unless they go in the package.

If an answer is vague, ask one follow-up. Don't write anything else until all ten are answered.

Then write my AGENT CARD: the ten headings above, each answered in plain words, under 300 words in total. Ask me to confirm or correct it. I'll paste the confirmed card into the next steps.

Here is my confirmed agent card:
[paste your agent card]

Below is one piece of my agent: its instructions, or one file it reads. Rewrite it so a stranger can use it:
- Replace every client name, company, person, email, phone number, price, account number, and private link with an obviously made-up placeholder in [BRACKETS].
- Turn any line that only makes sense to me ("use the usual format", "like last time") into a clear rule. If you can't tell what I meant, ask me instead of guessing.
- Keep the steps, rules, and tone exactly as they are. Don't improve it or add anything.

Reply with:
1. The cleaned text, in one code block.
2. Secrets found: any password, API key, token, or private link, and where it was. Write "none" if there were none.
3. Questions for me, if any. After I answer, give me the final cleaned text with my answers applied.

The piece:
[paste the instructions, or one file]

(Run this once for the instructions, then once for each file.)

PART A. Paste this first.

Using my agent card below, write one realistic sample input for my agent. Use the exact format my agent expects, with realistic messiness (a missing field, an odd case).
- If my agent works on text or files, make it up, for a different business than mine.
- If it works on live things (a website, an inbox, an account, a search), don't invent them; it would have nothing to read. Instead, tell me what real, public example to use that I've never run it on before, and how to pick one.
Reply with the input, or the instructions for choosing one, only.

[paste your agent card]

PART B. Now run your real agent on that sample input, in your own tool. If you can, run it somewhere your custom instructions, memory, or always-loaded files are switched off, because a buyer won't have them. Then paste this, with the result.

I ran my agent on that sample. Here is what it produced:
[paste the output]

Compare it with "what a good result looks like" in my agent card. Reply with:
1. What it got right.
2. What it got wrong or missed.
3. One plain line for each problem, written as a limit a buyer should know ("It skips invoices without a due date").
Don't rewrite or fix my agent. Just report.

From my agent card and test results below, write these three files. Put each in its own code block, with the file name above it.

README.md: one short paragraph on what the agent does and the result a buyer gets. Then setup in five numbered steps or fewer, written for someone non-technical. Then a list of the files in the folder.

connections.md: each app or tool it uses, what it does there (reads, writes, sends), and the access a buyer needs to give it. Never include a password, key, or private link.

limits.md: three to six plain lines, taken from the test results and the "got something wrong" answer in the card.

Don't add abilities the card and the test don't show.

Agent card:
[paste your agent card]

Test results:
[paste the reply from step 3, part B]

Use a NEW chat for this step, so the writer only sees the evidence.

You are a fair reviewer writing a marketplace listing for an AI agent. You haven't used it. All you have is its agent card and one test result, below. Write the listing with these fields:

- Name: what it does, in two to four words. No "AI", "ultimate", or "pro".
- Category: the job or industry it serves, one or two words.
- Built by: the builder's role and how they used it. No names.
- One-line promise: the job it takes off a buyer's plate, starting with a verb.
- What it does: two or three plain sentences. What goes in, what comes out, and what the buyer still does themselves.
- Time it saves: only the builder's number from the card, labelled "builder's own estimate". If it's missing, write "needs the builder".
- Running on real work for: only from the card. If it's missing, write "needs the builder".
- Works with: the tools from the card.
- Try it on your own data: the one input a buyer should test before paying.
- Good fit: two lines on who it suits.
- Not a fit: two lines on who should skip it. Required.
- Honest limit: the most important problem from the test. Required.

Rules: no superlatives ("best", "powerful", "game-changing"), no claim the card and the test don't support, and no number that isn't in the card. Plain language.

Agent card:
[paste your agent card]

Test results:
[paste the reply from step 3, part B]