JeremyAI Loading the file…
For JeremyAI members
Give Jeremy the full picture.
Give JeremyAI your offer, your numbers, and what you want back. This guide helps you put them into one clear request, so Jeremy has the context to work from your business.
Example with made-up details: "How do I improve my show rate?" becomes:
Help me find out why booked calls aren't showing up for [what you sell]. Here are [your real numbers] and the reminder messages we send now. I want [what you need back]. Don't assume the reminders are the problem, and mark anything [you can't be sure of].
Gaps filled: 0 of 4. These details make the request more specific.
Where do you want to start?
Start with the builder when you have one problem in mind.
Already drafting in Claude or ChatGPT, or want to set up the full system behind this guide? That's under Advanced.
The whole idea in three steps
- Step 1Say what you want back
Your goal, and the specific thing you need now: a diagnosis, a rewrite, a plan.
- Step 2Give Jeremy the real thing
Paste the script, the messages, or the numbers, with the dates they cover. Don't describe them from memory.
- Step 3Leave blanks blank
Relevant gaps are marked [MISSING]. The prompt asks Jeremy to clarify what matters.
Common questions
Do I need Claude or another AI for this?
No. Use the quick builder and paste the result into JeremyAI. The other tools under Advanced are optional.
Do I need to pick a skill?
No. Jeremy picks the right skill automatically. If you want a specific skill, you can name it in your request. See How to Use JeremyAI Skills.
What if I don't know a number?
Leave the box empty. Relevant gaps are marked [MISSING], and the prompt asks Jeremy to clarify them rather than invent an answer. In step 3, you can allow a read-only lookup when a suitable connection is available.
Can Jeremy read my links and files?
The builder itself can't open links or files. Paste the text into the material box, or attach the file in JeremyAI next to the prompt. The prompt reminds Jeremy to check that it actually has the material.
Why did Jeremy's answer change after I added numbers?
That's expected. The first answer only uses what you gave it. See Why did JeremyAI change its answer?.
Is what I type saved anywhere?
Only in this tab, unless you turn on "Remember answers in this browser" in the builder. Nothing is sent anywhere until you paste it into JeremyAI yourself.
Something isn't working. Who do I contact?
Message support in the help center or email hey@utari.ai with the chat link, your prompt, and what you expected. See How to contact support.
About 5 minutes
Build a prompt for JeremyAI
Answer what you can and skip what you can't. The prompt marks relevant gaps and asks Jeremy to clarify what matters. Review it, then copy it into JeremyAI.
Off. Your answers stay in this tab only.
Don't paste passwords, API keys, or private client details.
1
What do you want to fix?
Pick the closest match. Then choose whether you want Jeremy to find the cause first, or build something you've already decided on.
Advanced · Optional
Already drafting in Claude, ChatGPT, or Gemini?
You can have that AI interview you first, then write the prompt for JeremyAI. The drafting AI prepares the prompt. JeremyAI does the business work. Use the quick builder if you already know what to include. An interview can help when the task or missing context needs more discussion.
Initial request→Clarify the gaps→Approve the brief→JeremyAI
The interview prepares the same kind of brief, but can ask follow-up questions. The quick builder only assembles what you type. Neither route installs the full system.
Choose the drafting tool
Use a dedicated chat for this task. The instruction below changes with your selection.
Start with your task and what you know
Copy the interview starter, replace its bracketed fields, and supply relevant material. Do not give the other AI a solution to defend when you need help diagnosing a problem.
Interview starterAnswer the important questions
The starter asks the drafting AI to collect material gaps, then confirm the goal, constraints, output format, and failure conditions. You approve that brief before it writes the prompt. Unknown information can remain unknown.
Keep the conversation focused on one decision. The aim is better preparation, not an endless questionnaire.
Review the prompt and hand it over
Check every number and claim. Remove unconfirmed assumptions presented as facts. Paste only the finished prompt into JeremyAI and separately provide any material it needs.
Text is not a workspace transfer. Project context, private links, earlier conversations, and attachments do not travel with the copied prompt. A filename alone does not supply a source.
For repeat use Save a dedicated prompt-drafting role
Use these instructions only in a dedicated drafting chat, project, or equivalent instruction area. Check that your tool actually saves and loads them. Do not replace your account-wide preferences just to prepare one brief.
Reusable drafting instructions
This shorter adaptation does not install specialist agents, the original app, independent reviewers, or automatic rule editing.
Optional handoff Use an already configured JeremyAI MCP connection
Creating a project does not establish an MCP connection. First confirm that your client has a working JeremyAI connection and can make a test request. Copy and paste remains the fallback.
After approving the final brief, use this handoff instruction. This page does not perform the connection or send the request.
Approved-prompt handoff
Send the approved prompt above to JeremyAI through the available JeremyAI MCP tool and retrieve its response. Use only the context and material I approved for this handoff. Keep JeremyAI's response separate from your commentary. If the connection is missing or the request fails, report that and give me the approved prompt to copy manually. Do not present your own answer as a response from JeremyAI.
Don't put connection tokens or API keys in a brief. For how the connection works and where conversations appear, see Connect an AI client to JeremyAI.
Advanced · One working session
Set up the full system behind this guide
This guide is a simplified front end for a longer framework, the Master Prompt System. If you want the whole thing, including a nine-step prompt builder and reviewer agents, this walkthrough sets it up in Claude. You don't need any of this to use the quick builder.
Written for Claude. The steps below follow the original file. If you use another AI, the adaptation route maps the same setup to it. Expect to check what your tool can and can't do.
Original route: Claude
This walkthrough summarizes Part 1. The source document remains the reference for exact instructions and appendix relationships.
Choose your Claude environment
The source provides three columns in its Surface Map. Choose the environment you actually use, then confirm its current capabilities before creating anything.
claude.ai ProjectThe source maps operating instructions to Project Instructions and reference material to project knowledge. It describes manual rule updates and a sequential review fallback.
CoworkThe source maps work to a connected folder and supported skills or files. Make the session-start loading steps explicit.
Claude CodeThe source uses a project folder, instruction files, skill and agent configuration, and optional MCP tools. Review permissions and paths before applying commands.
The map describes the supplied framework, not a live capability test. See and Claude's project instructions.
Supply the original file and start the walkthrough
Save the file below and attach it, or copy its full text into a new setup conversation. Ask the AI to identify any sections it cannot read. Supplying a long file is not proof that every section was read.
Read MASTER PROMPT SYSTEM end to end. Run Part 1 as my setup guide, one stage at a time. Ask before you write any file.
This command is quoted from the source. Keep permissions and confirmation steps enabled. Ask for an explicit plan before file, agent, skill, or setting changes.
Fill placeholders as each stage needs them
Use real business details. Leave unavailable values marked rather than inventing them. Names such as
{{BUSINESS_NAME}},{{AUDIENCE}},{{BRAND_VOICE}}, and{{VAULT}}belong to the setup, not facts for a customer-facing prompt.Complete and check all six stages
One stage per turn. Review the proposed changes, verify the result, and approve the next stage. Save any manual updates in the location established during setup.
- Stage 1
The operating brain
Create the loader,
CORE.md, andRULES.mdwith the original loading order, task loop, writing gate, and dispatch table. Do not claim rules update themselves if you must save them manually.Check: in a new session inside the same configured project or workspace, follow its loading procedure and ask what happens before a task. Confirm that the saved rules are actually used. - Stage 2
Prompting patterns
Set up Reverse Prompting (A2) and Prompt Contracts (A3). Preserve the source's five-question pattern and its goal, constraints, format, and failure conditions.
Check: test the vague task in the source. The system should clarify the task, wait for answers, and seek approval of the contract before building. - Stage 3
Multi-agent patterns
Set up verification, consensus, and discussion patterns. Use independent agents only where the environment provides them. Clearly label a sequential role review or separate-chat review as a fallback. Optional MCP orchestration needs its own capability checks.
Check: record which mechanism ran and whether reviewers had separate context. A few labeled viewpoints in one chat do not establish independent verification. - Stage 4
Specialist agents
Use the source's orchestrator, reflection, design, website, writing, MCP, and content references. Adapt named tools and paths with approval; do not silently drop specialist workflows.
Check: use the source's homepage test. The configured route should begin with the required discovery inputs rather than immediately produce a page. - Stage 5
Build the Master Prompt Builder
Build the nine-step app from Part 2, with code-based prompt assembly. The source mentions a separate reference build; it is not included in this guide. Without that file, follow the build specification. This website's quick form is not that reference build.
Check: run all 14 acceptance tests in §2.13. Test the question-bank fallback with live AI unavailable. Record unsupported features rather than declaring them complete. - Stage 6
Test a real task
Work through the app, use references, build and export the prompt, then try it in the receiving AI. Review the output and record corrections using the source's reflection step.
Check: verify that the exported brief preserves your answers and that the receiving AI has the actual material. Test the handoff instead of assuming an export transfers the workspace.
Use the saved setup for later tasks
Return to the configured project or workspace, load its saved instructions as established, and reopen the built app. Prepare one brief, review it, then give JeremyAI the finished prompt and supporting material. Save confirmed corrections where the next session will actually load them.
Advanced alternative Adapt the full setup to another environment
Same source. An explicit capability map.
Part 1 can serve as the setup reference outside Claude. Use an adapter to preserve its structure while identifying what runs natively, what needs different configuration, and what requires an approved fallback.
This adapter is an addition in this guide. It is not wording from the original framework and has not been validated as a complete installation across the selected products.
| Keep from Part 1 | Confirm in the receiving environment |
|---|---|
| Loader, CORE, and RULES | Where instructions and references live; who updates them; how a new session loads them. |
| Prompting and specialist workflows | Supported skills or reusable instructions; required inputs; routing and permission controls. |
| Agent review patterns | Separate agents and context where available. Clearly identified fallbacks where not available. |
| Optional MCP orchestration | Both client and server support, tool access, permissions, and an actual test request. |
| The nine-step app | HTML execution or a local-file route, export, storage, optional live AI, and question-bank mode. |
| The checks | Stage-level verification plus §2.13 acceptance tests. Log missing and unverified behavior. |
Attach the unchanged source file and paste this adapter into the setup conversation. Approve the proposed map before configuration begins.
Full-system adapter
If the whole file will not fit, work through the required sections and appendices with an explicit inventory. Do not silently substitute a summary for an unread source. The conversational drafting route remains a lighter alternative, not the full setup.
A 3-minute read
How it works
You don't need this page to use the builder. It explains what each part of the prompt does, and what to do when an answer still misses.
What's in the prompt
The builder puts your answers into these sections. Each one answers a question Jeremy would otherwise have to guess at.
ROLEThe kind of help you need, like sales-copy review or ad analysis.DESIRED OUTCOMEYour goal, and whether you want the cause found first or something built.KNOWN FACTSYour supplied details and numbers. The prompt asks Jeremy to flag conflicts.ASSUMPTIONS / UNKNOWNSThings you believe but haven't checked, and relevant gaps. The prompt asks Jeremy to keep these separate from facts.THE MATERIALThe actual script, messages, or numbers you pasted.WHAT I WANT BACKThe specific answer you need now, plus a few pointers for that type of problem.OFF-LIMITS / GUIDELINESYour limits, plus rules like "don't invent numbers" and "write plainly."BARRIERSAnything that could still throw Jeremy off, listed so it deals with it first.
Why answers get better
- Jeremy works from your materialThe actual script and dated numbers give Jeremy something specific to review.
- Blanks are labeled, not filledA [MISSING] label makes a gap visible. The prompt tells Jeremy to ask about it when it could change the answer.
- Cause before cure"Find the cause first" asks Jeremy to weigh a few explanations before it recommends a fix.
- It checks its own answerThe prompt asks Jeremy to compare its answer with what you asked for and say what it couldn't verify.
When the answer still misses
The answer is generic or ignores something I gave it.
Check the chat actually has the material (a filename isn't the file). Then point to the exact thing it missed and ask for a revision in the same chat. If Jeremy keeps missing things you supplied, contact support with the chat link, your prompt, and what you expected.
Jeremy couldn't read my connected data.
Leave the [MISSING] label in place and give Jeremy the number yourself, or check the connection separately. See Meta CLI Setup and Connect JeremyAI to GoHighLevel.
Jeremy jumped to a fix I didn't want.
Switch step 1 to "Find the cause first" and add the fix you don't want under Off-limits. Then ask Jeremy to weigh other explanations.
The answer changed after I added numbers.
That's how it should work. See Why did JeremyAI change its answer?.
Copy or save didn't work.
The guide shows the text in a box you can select and copy by hand. Downloads depend on your browser settings.
Where this comes from
This guide is a simplified front end for a longer framework, the Master Prompt System, written for setting up Claude. Its 24 reference files are below. You don't need them to use the builder. They're here for anyone setting up the full system under Advanced.
Advanced · The original file
MASTER-PROMPT-SYSTEM.md
You don't have to read this. It's the original framework file, written for Claude to follow during the full setup. Skim it if you're curious. Don't paste it into JeremyAI as a business request.