It's 11pm and you're still
answering messages.
Here's the job I'd fix first. It's free, it works tonight, and there's nothing to install — you paste two prompts into an AI chat you already have.
No email required. Nothing to sign up for. Take it and go.
The AI drafts. You decide. Nothing goes out until you say so.
That's the whole idea. Most AI tools act, and you find out afterwards. This one writes the reply and waits for you. You stay the one who hits send — always.
What you're about to do
Three steps. About 25 minutes the first time, then two minutes a message after that.
Step 1 is the part people skip, and it's the part that makes the difference. An AI with no Brain writes like a brochure. An AI with your Brain writes like you.
Step 1 — build your Brain
Paste this into a new AI chat (Claude, ChatGPT, whichever you use) and send it. It interviews you — six sections, a few questions at a time, with examples so you know the shape. Answer casually. "Skip" is a fine answer.
At the end it hands you a finished document. That's your Brain. Keep it somewhere you can open in five seconds.
You are going to interview me to build my "Business Brain" — one plain document that holds everything an AI assistant needs to draft work as my business: facts, voice, customers, rules. Follow this process exactly. INTERVIEW RULES 1. Interview me ONE SECTION AT A TIME, in the order below. Never show me all the questions at once. 2. For each section: briefly say what it's for (one line), then ask 2–4 plain questions, each with a tiny example answer in parentheses so I know the shape you need. Then WAIT for my answers. 3. Write the document in MY words and MY language. If I answer casually, keep it casual — do not polish me into corporate voice. If I answer in Thai, the Brain is in Thai. If I mix languages, ask which one my customers get. 4. NEVER invent or assume a fact. If I skip a question or my answer is vague, put [FILL LATER: what's missing] in that spot — do not guess. 5. If an answer seems to conflict with an earlier one (e.g. a price changes), point it out and ask which is right. 6. After each section, show me just that section drafted, ask "anything wrong or missing?" — then move on. THE SIX SECTIONS Section 1 — WHO WE ARE: what the business is, who runs it, what it's known for. (2–3 sentences of truth, not marketing.) Section 2 — WHAT WE SELL: products/services with real current prices, plus the policies I repeat all day — deposits, lead times, delivery, refunds. Push me for exact numbers; if a price "depends," ask what it depends on and write that. Section 3 — HOW WE TALK: my tone in one line; phrases I actually use; phrases I never use; emoji rules; language rules. Then ask me to paste 2–3 REAL replies I've sent and was happy with — file them under "Examples of us at our best." Section 4 — WHO OUR CUSTOMERS ARE: who writes to us, what they usually want, what they worry about. Section 5 — FACTS THAT CHANGE: this month's availability, current promotions, anything temporarily true. Remind me this section needs a date and gets updated often. Section 6 — HARD BOUNDARIES: rules the AI must NEVER break in a draft. Ask: "What's the worst thing a wrong message could do to your business?" — then help me turn each fear into a never-rule (e.g. "Never confirm a date without checking with me," "Never admit fault for anything unverified," "If a customer is angry, draft a holding reply only and flag it for me"). Suggest common ones I haven't thought of, but I decide what goes in. ASSEMBLY When all six sections are done, output my complete Business Brain as one clean markdown document in EXACTLY this structure: # Business Brain — [my business name] *Last updated: [today's date]* ## 1. Who we are ## 2. What we sell — real prices and policies ## 3. How we talk — and how we never talk ### Examples of us at our best ## 4. Who our customers are ## 5. Facts that change ## 6. Hard boundaries — the AI must never: Then, below the document: - List every [FILL LATER] gap in one short list so I can close them. - Offer the test: "Want to hear it work? Say 'test' and I'll introduce your business to a stranger using ONLY this Brain — in your voice." If I say test, do exactly that, using nothing beyond the document. Begin the interview now with Section 1.
A note on what this is doing: it never invents a fact about your business. If you skip something, it writes [FILL LATER] rather than guessing. That's deliberate. A confident wrong answer is worse than an obvious gap.
Step 2 — the Reply Drafter
Now the job itself. Paste this into a new chat, drop your Brain into the first slot and the customer's message into the second.
Below are (1) my Business Brain and (2) a message from a customer. Draft the reply I would send, following these rules: 1. Use ONLY facts from the Business Brain. If the customer asks something the Brain doesn't answer, do not guess — write [CHECK: what I need to confirm] in that spot instead. 2. Match the voice in "How we talk" exactly, including the language the customer wrote in. 3. Make no promises about dates, availability, or prices unless the Brain explicitly confirms them. When unsure, offer to confirm and get back to them. 4. Keep it the length I'd actually send in chat — short. 5. This is a DRAFT for my review. Do not add a signature block or anything I'd have to delete. After the draft, add one line: "Check before firing:" listing anything I should verify. === BUSINESS BRAIN === [REPLACE: paste your whole Business Brain here] === CUSTOMER MESSAGE === [REPLACE: paste the customer's message here]
You'll see two kinds of brackets, and they're not the same thing:
- [REPLACE: ...] — slots you fill in before sending.
- [CHECK: ...] — markers the AI writes inside its draft to flag what it wasn't sure about. You never type these. You resolve them before you send.
That second one matters more than it looks. It's the AI admitting what it doesn't know instead of covering for itself.
Step 3 — the Gate
The draft comes back. Don't send it yet. Read it against three questions — it takes about twenty seconds:
- Facts — is everything in here actually true? Did it resolve every [CHECK:]?
- Tone — does this sound like me, or like a machine doing an impression of me?
- Promise — did it commit me to anything I can't deliver?
Then one of three verdicts: fire it, fix it, or bin it.
If you bin a lot of drafts, that's not the AI failing. That's your Brain being thin. Add the line that would have prevented it, and that mistake never happens again.
Try this before you trust any of it
Ask your AI anything. Then ask it to critique its own answer — to find the flaws in what it just told you.
It will almost always find real ones.
That's the thirty-second version of why this method exists. The AI sounds equally sure whether it's right or wrong, which is exactly why a human stays at the gate.
What this is part of
What you just ran is one job out of seven. The full kit — The AI You Don't Have to Trust, $97, one-off — is the method written out properly plus the other six: the end-of-day sweep, the polite chaser, the review responder, the enquiry triage, the one-thing-three-posts, and the skeleton for building your own.
35 files: a 90-page playbook, a 27-slide briefing you can read in twenty minutes, printable daily tools, and an optional pack for people who want to run this from their own machine. No software, no accounts, nothing phoning home.
Delivered by email within minutes. If it's not for you, email me within three days of delivery and I'll refund you in full — keep the files, you can't un-read a PDF.
Honest note: this is new. Nobody has bought it yet, so there are no testimonials on this page and there won't be invented ones. What you just used for free is the real thing — judge the rest by that.
Get the kit $97→ The mirror test — thirty seconds, and it explains why the gate exists.
→ The 20-Second Judge — the three questions to run on the draft you just got.
I'm Jee. I work solo out of Chiang Mai and I build these systems for small businesses — the AI drafts, the owner decides, nothing sends itself.
Questions, or the prompt did something strange? hello@gniusos.com — it reaches me, not a queue.