đ Works best with your call recorder. Connect Fathom, Fireflies, or Zoom and Jeremy AI reads the walkthrough for you, no pasting.
Paste what you said while walking through a process. Get one file back: a process summary, the steps in the order you spoke them, open questions and a claims check. It writes down only what you said. Nothing is shared.
You'll see the SAMPLE transcript pasted in with the rules, the process guide written, the file saved with nothing shared, the summary with the owner marked not provided, the five steps in spoken order and the open questions with the claims check.
What this use case does and who it's for
Every team has a process that lives in one person's head. You do it the same way every time. You can explain it out loud in a minute. But nobody ever writes it down, so the next hire guesses.
Here's the shortcut. Record yourself walking through the process, get the transcript, and give it to Jeremy AI. It turns your words into a clean process guide. Four parts:
A process summary. Three columns: item, detail and source. It covers purpose, owner, tools and time.
A steps table. Four columns: step, action, tool and done when. One row per step, in the order you said them.
An open questions list. Anything your transcript leaves unclear.
A claims check. Three columns: claim, source and status. Each claim is marked "stated by speaker" or "speaker estimate."
It's for agency owners, ops leads and team leaders who want a written process guide without sitting down to write one. It's a draft. Jeremy AI doesn't send or share anything.
What you need before you start
A Jeremy AI account. Nothing to connect to try the sample.
A transcript of you (or your ops lead) talking through the process. The sample is a short, made-up walkthrough of launching a new gym client in five steps. It's spoken by "Dana K, ops lead."
The rules for what the process guide must never do. The sample uses HARD RULES: the speaker is the narrator, not the stated owner, so the owner is "not provided" unless the transcript names one. Every table has 4 columns or fewer. Cells are 12 words or fewer. Use only what the transcript says. Use exact numbers only. Never calculate a new number like a total time. Never invent a step, tool, owner, time or deadline. Write "not provided" if something is missing. No claims about what most teams experience. Never address you by name.
How to set it up
You use the default Jeremy AI worker. There's nothing to build.
Open Jeremy AI. From the dashboard, type in the main box. Check the worker selector says Jeremy AI.
Paste the prompt below. Swap your own transcript in where the sample one is.
Click Send message.
Wait about a minute.
When it says the file is saved, read the chat summary first.
Open the file card and read the process summary.
Scroll to the steps table. Check the order, the tool column and the done-when column against what you said.
Read the open questions and the claims check.
The prompt
This is the exact prompt from the walkthrough. The company, the speaker, the transcript and every number are SAMPLE and made up. Replace them with yours.
Act as my "Process Guide Writer" worker for SAMPLE Lumen Growth, a paid-ads agency for gym owners (all data below is made up). Turn the SAMPLE screen-recording transcript below into a team process guide. Create ONE file named process-guide-new-client-launch.md using the file tool (reply with a short summary only) with: 1) a PROCESS SUMMARY table with exactly 3 columns (item, detail, source), covering purpose, owner, tools and time, 2) a STEPS table with exactly 4 columns (step, action, tool, done when), one row per step in the order spoken, 3) an OPEN QUESTIONS list for anything the transcript leaves unclear, 4) a CLAIMS CHECK table with exactly 3 columns (claim, source, status). HARD RULES: the speaker is the narrator, not the stated owner, so write the process guide owner as "not provided" unless the transcript names one; every table has 4 columns or fewer, short cell text of 12 words or fewer; use ONLY what the transcript says, exact numbers only, never calculate a new number such as a total time, never invent a step, tool, owner, time or deadline, and write "not provided" if something is missing; no general claims beyond the transcript and no claims about what most teams experience; never address me by name anywhere in the chat reply or the file. Drafts only, nothing is sent or shared.
TRANSCRIPT (SAMPLE, spoken by Dana K, ops lead):
"Okay, this is how we launch a new gym client. First, open the client's intake form in Notion and copy the number of locations. Second, in Meta Ads Manager make one campaign per location, and name it gym name dash city. Third, upload the 2 best ads from the client's past ads, and stop the rest. Fourth, in GoHighLevel make one booking page per location and test that a booking lands in the calendar. Fifth, send the client the launch checklist in Slack. Step two usually takes me about 20 minutes per location. Who checks the booking test, I'm not sure yet, we'll decide that. Done."
Prompting tips and how to adapt it
Paste the whole transcript. The more of your own words it has, the less it has to guess.
Say who's speaking. The prompt says the speaker is the narrator, not the stated owner. That's why the owner shows as "not provided." Unless you said who owns it, nobody does.
Ask for the steps in spoken order. One row per step, in the order you said them. That keeps it faithful to how you actually work.
Keep cells short. The prompt says 12 words or fewer. Short cells fit the file panel with nothing cut off.
Ban new numbers. The prompt says never calculate a new number like a total time. In the sample, only "about 20 minutes per location" for step two shows up. There's no launch total, because the speaker never gave one.
Ask for open questions. Gaps in your transcript become a to-do list instead of made-up filler.
Say what's not provided. Tell it to write "not provided" for anything missing.
Ask for the claims check. It's how you confirm every line in a couple of minutes.
Adapt it. Swap in any process: onboarding, reporting, client offboarding or a weekly review. Add or remove columns, but keep the table to 4 columns or fewer.
The result to expect
One Markdown file named process-guide-new-client-launch.md, titled "Process Guide: New Gym Client Launch." With the sample transcript you should see:
A 3-column process summary. Purpose is "Launch a new gym client's ads and booking setup." Owner is "not provided," because the speaker never named one. Tools lists Notion, Meta Ads Manager, GoHighLevel and Slack. Time says step two takes about 20 minutes per location.
A 4-column steps table with five rows, in spoken order. Open the intake form and copy the number of locations. Make one campaign per location, named gym name dash city. Upload the 2 best ads from the client's past ads and stop the rest. Make one booking page per location and test that a booking lands in the calendar. Send the client the launch checklist.
An open questions list. Who checks the booking test? The speaker said undecided. What criteria define the "2 best ads"? Time estimates for steps other than step two are not provided. Who sends the launch checklist is not provided. Total launch time is not provided, and not calculated.
A 3-column claims check. The campaign per location, the 2 best ads, the booking page per location and the launch checklist are "Stated by speaker." Step two's 20 minutes is marked "Speaker estimate."
The chat reply summarizes the file and offers quick follow-ups, like adding more detail to the steps or tightening the open questions list. Nothing is sent or shared.
These are sample results from a made-up agency. Wording changes from run to run. Read every line before you hand it to your team.
Check the tool and done-when columns
This part matters. In the sample, the speaker said "in Meta Ads Manager" for step two, but never named a tool for step three. The process guide lists Meta Ads Manager for step three too. It's a reasonable guess, because those ads live in Meta Ads Manager. But it's an inferred tool, not a spoken one.
The "done when" column works the same way. Your transcript rarely says what "done" looks like, so Jeremy AI fills in a sensible check.
So always check the tool column and the done-when column against your own words. If you didn't say it, either confirm it's right or change it to "not provided." You can also tighten the prompt: "Only list a tool if the speaker names it for that step. Otherwise write 'not provided.'"
Common mistakes and fixes
A tool shows up that you never said. Check the tool column. Say: "Only list a tool the speaker names for that step. Otherwise write 'not provided.'"
The owner is filled in. Say: "The speaker is the narrator, not the owner. Write 'not provided' unless the transcript names one."
A total time appears. Delete it. Remind it: "Never calculate a new number such as a total time."
A step is invented or reordered. Say: "One row per step, in the order spoken. Don't add steps."
The done-when column states something you never said. Rewrite it, or say: "Write 'not provided' unless the speaker says what done looks like."
A table has more than 4 columns. Remind it: "Every table has 4 columns or fewer."
A cell is too long. Say: "Keep every cell to 12 words or fewer."
A general claim sneaks in. Delete it. Remind it: "No claims about what most teams experience."
You hand it out as is. Don't. Read it once, fix the columns, then share it with your team yourself.