How to use an AI well, and what the blueprint already does for you

The AI only sees what's on its desk

Picture a desk. What's on it is everything the AI knows right now: what you typed, what you attached, a little memory. That's its whole world. If the thing you need isn't on the desk, it guesses. If everything you own is piled on the desk, it guesses which thing you meant.

So the job isn't writing a cleverer prompt. The job is putting the right things on the desk and nothing else. Jake Van Clief puts it this way:

"Every word in a prompt is a question, and something has to answer it."

The ideas on this page are his, in our words. What the blueprint does with them is ours. His channel: youtube.com/@JEVanClief.

Four things a good setup has

• One short map the AI reads first. A file at the top of the folder that says what this place is, where things go, and what to read for each job. One screen, no more. The most common mistake is a giant map with everything in it, which puts the whole drawer back on the desk. The blueprint writes this map from your answers at setup.
• One AI, many folders. You don't need a "research AI" and a "writing AI" and an "email AI." They're the same model reading different instructions. One good AI plus a good set of folders does the work. The folder it opens decides what it becomes for that job. The blueprint works with the AI you already have, plus a folder of instructions for each job it knows. The AI opens one only when your request matches.
• Your method in one place, your facts in another. How you do things carries from job to job. Facts about one client or one project stay with that client. Mix them and the AI starts promising everyone the same date. Keep each fact in one file and point everything else at it. Two copies drift. The blueprint keeps the two apart.
• A voice file. AI writing sounds like AI because nobody gave it your voice. Fix that with one file: a few plain rules it can check (no hype, sign it with your first name) and three or four things you actually wrote. Samples beat adjectives. The AI copies a sample far better than it obeys a word like "punchy." It copies your bad habits too, so pick samples you're proud of. Setup asks how you want it to sound and builds this file, and in week one you can hand it three or four things you wrote and have it learn from those.

Where you put your eyes

• Check early, where a mistake is cheap. Nothing moves to the next step until you've looked at the last file. Look at the start, where a wrong name costs a ten-second edit, not at the end, where it costs a rebuild. You don't watch every step. You put your eyes where your judgment changes the result. That's why setup shows you the folders before it builds them.
• Say what must never happen without you. Sending, spending, deleting. A line in the map saying "stop and wait for me" is a soft rule, and the AI almost always follows it. At setup the blueprint names those three, asks what else belongs on the list, and writes the list where every chat reads it. Before it does any of those things, it asks you.
• "Done" isn't the test. Open the file. Finished work should land as a plain file you can open without the AI. If you can open it, read it and change it, the setup works. Everything the blueprint makes is a file in your folder. Drafts stay drafts until you send them yourself.
• Anything with one right answer becomes a small script. Renaming files by date, pulling numbers from a report, resizing photos, finding every stray dash in a draft. If you've done a step the same way three times and the answer doesn't depend on taste, ask the AI to write a script. Check it once against something you already did by hand, then trust it. A script that's wrong is wrong the same way every time, which is why one check is enough. The blueprint already uses scripts for its one-answer jobs: renaming photos by date taken, converting files, summing a bank export by category.

How to ask for something

1. Say the outcome and who it's for. "A one-page summary of this contract, for me, so I can decide whether to sign."
2. Name the file or link that matters. "Read the PDF in this folder called lease-2026.pdf." In a big folder, not every file reaches the desk. Name the one.
3. Say where the result should land. "Save it next to the original." "Put it in drafts with today's date and the client's name in the file name."
4. New job, new chat. An old chat drags the last job onto the desk. Start fresh for anything unrelated.
5. Correct it in plain words, right there. "Too formal. I'd never say that. And don't promise a date we haven't confirmed." Every correction is a decision only you can make.
6. Then write the correction down. After the job, put your corrections in a file so you never retype them. When the same fix comes up twice, change the file, not the prompt. The blueprint won't start this list on its own, but you can ask it for one.
7. Keep the prompt short once the steps are written down. "Check my email." "The recording's in, take it from here." When the work lives in the folders, a prompt is a button. That's why the blueprint's first-week asks are each one sentence.

Small habits that save a lot

• Let it interview you. "I want help with my weekly report. Ask me one question at a time: what comes in, what I decide, what goes out, who gets it, where I stop to check."
• Tell it to say so when it doesn't know. One rule in the map: if the notes don't confirm something, say so instead of filling it in.
• Watch what it reads. Give one real request in a fresh chat and watch which files it opens. If it opened the right few and skipped the rest, good. If it wandered through everything, the map is too vague. Fix the one line for that job.
• Ask the three freshness questions. In a brand new chat: "What is this folder? What's in progress? What's next?" Wrong answers mean the map is out of date.
• Read what it wrote before you trust it. A file it drafts for you may carry a rule you never said, or turn one example into a law.
• Use an export when one exists. "Here's the export from my booking system" beats having the AI click around a website.
• Write rules like the note you'd leave whoever covers your vacation. Odd little rules included. That's exactly the material the AI would never work out on its own.
• Start from a job that hurts. Not "I want a second brain." One job you repeat that has a person at the end: the weekly report, the newsletter, new-client onboarding. Build the smallest thing that does today's version, and add a folder only when the work asks for one.
• Skip the plumbing in week one. Connectors, custom tools and automations all bolt on later. A lot of setup advice online comes from people whose favorite part is the setup. The blueprint arrives with its tools installed, so your first week is a few plain asks, listed on the main page.
• Count the whole cost. The subscription, the setup, the time checking its work, the time waiting. A cheaper tool that eats your Saturday costs you a Saturday.