The working instruction for an AI assistant: 7 mandatory sections
Your assistant writes generic text not because "AI is raw" but because half the instruction is missing. The full 7-part skeleton, a step-by-step build guide, and a scaffold generator.
01 — The diagnosis
Why does your assistant write "like everyone"?
Working with AI is built on showing quality examples and setting frames. Otherwise the model takes "something average from the internet" — that's its nature, not a defect.
When an assistant produces faceless text, I almost always find one of three causes: the instruction has no reference examples (only a description like "write lively"); no anti-examples — the model doesn't know what drives you mad; no criteria and self-check — the model doesn't know what "good" means and hands you its first draft.
All three are cured by structure. Across many focus-group cohorts on AI adoption, a seven-part skeleton crystallized — let's take it apart.
02 — The skeleton
What are the 7 sections of a working instruction?
Here's the full skeleton, with wording you can take:
1. Who is the assistant?
"You are a copywriter who writes texts the way I do. Your job
is to write so that I publish immediately instead of editing.
You study my manner of speaking, understand HOW I think, and
write from that understanding."
2. Who am I?
"I am [name]. [How you think, what you love, how you persuade,
what matters to you in texts]."Then the materials: 3. Style analysis (a separate txt file, not inside the instruction) and 4. Reference examples — your texts you actually like. With a mandatory caveat:
IMPORTANT: these examples are for understanding the approach,
tone and structure. Do NOT copy them word for word. Do NOT reuse
the same phrasings. Use them as a style compass, not a template.Without that caveat, ChatGPT will start bluntly copying phrases from your references — tested.
03 — The filters
Anti-examples, criteria and the self-check: what everyone skips
Three sections almost nobody has — and exactly the ones that separate an assistant from a draft generator.
5. Anti-examples. What you NEVER do and what infuriates you — with the why:
— DO NOT use: "Let's dive in", "In today's world",
"As we all know". Empty templates.
— DO NOT write long paragraphs with nested thoughts.
One thought — one paragraph.
— DO NOT end with a question unless I expect a real answer.6. Quality criteria. The text is good if:
— The headline reveals a problem people don't suspect they have
— The reader smiled at least once
— There's specificity, not abstraction
— It reads in one breath7. The self-check — crucial, at the very end of the instruction:
Before handing me the text:
1. Check it against the quality criteria above
2. Make sure you used no anti-examples
3. Check that it sounds like me
4. If something's off — rewrite without askingThe self-check makes the model run its own text through your filters before you ever see it. One paragraph — and half your edits disappear.
04 — The hack
Why must the instruction be written in your own voice?
A hack I discovered in practice: the instruction must be written in your manner of speaking.
"The assistant shall maintain a conversational style with elements of expertise…" → The assistant writes like a manual. It literally copies the tone of the document it reads.
"I can't stand corporate-speak. Write like you're telling a friend over coffee — with examples and self-irony" → The assistant starts thinking like you: the instruction's tone becomes the texts' tone.
Hence my advice: write every section by hand, in your own words. Ask AI to "generate the instruction" and you'll get templates and clichés — and an assistant that writes in templates and clichés. This is the one place where 40 minutes of manual work pays off every single day.
05 — The build
How do you assemble and install the instruction, start to finish?
From zero to a working assistant — six steps. Move one at a time:
Find at least 10 texts that demonstrate the style you want: your own posts you like, or texts by authors you want to echo. This step is critical: mess it up and the assistant writes generic "like everyone" text.
Ask AI to dissect the references: rhythm, sentence length, vocabulary, structure, devices. Save the result as a separate txt file, "Style Analysis" — it lives next to the instruction, not inside it.
Walk the skeleton from sections 02–03: the role, you, references with the caveat, anti-examples with the why, criteria, self-check. By hand, in your language — you can scaffold the draft in the generator below.
Hand your blocks to AI with the task: "merge my blocks into one coherent instruction, changing nothing in meaning or wording — add only short transitions, and write in my manner of speaking." The model glues; you verify your phrasings survived.
ChatGPT: GPTs tab → Create → paste the instruction into Instructions; too long — upload as a txt via the paperclip, along with the style-analysis file. Claude: create a Project and put the instruction into Project knowledge. Name it simply: "Voice — [name]."
Give the assistant three real-life tasks. Every edit you had to make goes back into the instruction: an edit → a new anti-example or criterion. Two or three iterations — and edits almost vanish.
06 — The generator
Where to start: scaffold your instruction in 2 minutes
Fill in four fields — get a skeleton you'll finish in your own words:
An assistant isn't "a longer prompt" — it's a 7-part instruction written in your voice: role, you, analysis, references, anti-examples, criteria, self-check. Scaffold it in the generator, finish it by hand, walk the six build steps. One evening of work — and your assistant stops being a draft generator.
FAQ
Which section of the instruction matters most?
The pair "reference examples + anti-examples." References show what right looks like (showing always beats describing); anti-examples define what's banned and why. Without the former the assistant writes "like everyone"; without the latter it keeps stepping on what infuriates you.
Why can't I just ask AI to write the instruction for me?
Because it will produce templates and clichés — and an assistant reading a template instruction writes template text. Write the sections by hand, in your own words: an instruction that sounds like you teaches the model to think like you. Bring AI in only for the final merge — with no right to change your wording.
Where do I install it — GPTs or a Claude Project?
Whichever you prefer: in ChatGPT — the GPTs tab → Create, instruction in the Instructions field (long one — upload as txt via the paperclip); in Claude — a Project with the instruction in Project knowledge. Same mechanics: the instruction and style file live in the system instead of being pasted into every chat.
What if the assistant still makes mistakes?
Feed the edits back into the instruction: every repeated mistake becomes a new anti-example or criterion. Two or three iterations on real tasks bring it to "I publish almost without edits." The instruction is a living document, not carved in stone.
