Talk to AI like a clear human (not magic words)

Lena’s Slack huddle ended at 9:04. The sticky on her laptop lid had three lines in blue Sharpie: inventory freeze Thursday, Midwest warehouse still on the old SKU list, Pat wants a one-pager before noon. At 9:12 she pasted those 47 words into ChatGPT and added eight more: “Can you summarize this meeting for my boss.” The reply was 400 words. It opened with “I hope this email finds you well.” It invented a Q3 alignment workshop. It assigned a follow-up to Raj, who had been on PTO since Friday. Pat’s 9:20 ping landed while she was still deleting the workshop. She had asked for three bullets. She had eight minutes. She got an essay.

This is Part 12 of Phase P, and Part 1 of Prompting for everyone. The chooser in Which AI product should I use? told you which door to open. AI setup from zero is accounts, cards, and where files go. If the artifact is a paragraph, start with AI for writing and questions. This series is what you type once the chat is open. No secret formulas, no jailbreak folklore, no “act as a Harvard professor.” Next in this series: Give context, constraints, and examples.

What you’ll learn

Why a prompt is a work request you would send a sharp intern
The anatomy of a useful ask: task, audience, facts you already have, format, and what to avoid
How Lena’s 8-word prompt turned a 3-bullet recap into a 400-word essay, and the 6-line rewrite that would have landed
A table of magic-word habits and the clear-human line that replaces each one
A paste-ready skeleton that works in ChatGPT, Claude, Gemini, or Grok without a token lecture

A prompt is a work request

Four cards: work request, role-play prefix, eight-word ask, and the intern test for a clear human prompt

Talk to the box the way you would Slack a sharp intern on their second week. You would name the job, name the reader, and say what “done” looks like. You would not write “Act as a Harvard professor of supply chain and produce a world-class summary using chain of thought.” That sentence is costume jewelry. The intern still does not know Pat, the 9:20, or that Raj is out.

A prompt is the text you send. Vendors sometimes call the craft “prompt engineering.” For this series, keep the intern picture. You are filing a work request. The model is fast and well-read and has never sat in your huddle. It will fill every gap with something that sounds like a meeting. Lena’s 400-word essay felt like a recap until she hit Raj’s name.

Role-play prefixes (“You are an expert…”, “Act as a senior VP…”) rarely beat a clear job, a named audience, and a length. As of writing (August 2026), ChatGPT, Claude, Gemini, and Grok already know how to write email. What they do not know is your org chart. The costume does not import Pat’s patience. It often hides the missing facts, because the prompt looks long and “professional” while still saying nothing a person could act on.

Jailbreak threads (ignore your rules, pretend the safety layer is off, “DAN” folklore) are a different hobby. You need three bullets before 9:20, not a model that broke its own policy. If a prompt would be embarrassing to send to a human colleague, it is too thin for the chat box too.

Rule of thumb: The intern test. If you would not send this message to a sharp person on day two, do not send it to ChatGPT. Add who you are, what you need, and what good looks like.

Anatomy of a useful ask

Four rows of a useful ask: task, audience, facts already on hand, then format plus what to avoid

Five pieces. You can skip one if it does not apply. You cannot skip all five and hope the model was in the room.

Task

The verb and the object. Recap the standup. Draft the one-pager. Rewrite this sticky into three bullets. OpenAI’s own prompting notes start the same way: outline the task with an action verb. “Help” and “look at this” are not tasks. “Write a 3-bullet recap of the pasted notes” is.

Audience

Pat, director, 90 seconds before a 9:20. Not “my boss.” Not “stakeholders.” Not “the team.” Audience tells the model how much they already know and how much flattery they will tolerate. Pat does not need a paragraph on why inventory freezes exist. Pat needs Thursday, Midwest, and the one-pager.

Facts you already have

Paste them. The sticky, the names, the times, who is out. Label the block “Facts, do not add any.” The model did not attend the huddle. If a fact is not in the paste, the polite-sounding extra is fiction. Lena’s Q3 workshop was fluent fiction. Raj’s action item was fluent fiction with a calendar attached.

Format, and what to avoid

Format is a number plus a shape: 3 bullets, 80 words, no greeting, no sign-off. “Be concise” is a wish. “80 words” is a constraint. Avoid is the list of landmines you have already stepped on: invented meetings, action items for people who were not there, “circle back,” “I hope this email finds you well.” You can add a tone line if you have a sample sentence. One sample beats “make it professional.”

Next in this series, Give context, constraints, and examples stretches the facts block and shows few-shot samples. Today, get these five on the page before you hit Send.

Eight words vs six lines

Same huddle. Same sticky. Two asks. Watch what the eight-word version leaves for the model to invent.

The 8-word prompt (plus the 47-word sticky, dumped underneath with no label):

Can you summarize this meeting for my boss

What came back at 9:12: 400 words, a greeting, a fake Q3 alignment workshop, three “next steps” including Raj, and a close that sounded like a vendor newsletter. Lena spent the next eight minutes cutting. She still owed Pat a one-pager by noon. The essay had used her remaining runway.

She tried a second pass before the 9:20. A group-chat pin had a “mega prompt”: 14 lines of “You are a world-class chief of staff, think step by step, write with executive presence.” She pasted that on top of the same sticky. 280 words this time. Still a workshop. Still Raj. The costume made the prompt look expensive. It added zero facts and zero length.

The 6-line human prompt for the same email:

I am Lena, ops coordinator for the warehouse rollout.
Write a recap for Pat (director). Pat has 90 seconds before a 9:20.
Facts, do not add any others:
- Inventory freeze starts Thursday
- Midwest warehouse is still on the old SKU list
- Pat asked for a one-pager before noon
Standup ran 28 minutes. 11 people on the call. Raj is on PTO through Wednesday.
Format: exactly 3 bullets, 80 words or fewer, no greeting, no sign-off.
Avoid: invented meetings, action items for people who were not there, "circle back," "I hope this email finds you well."

What that ask can return (and what Lena typed by hand once she stopped arguing with the essay):

Inventory freeze starts Thursday. No inbound slot changes after that date unless Pat signs off.
Midwest is still on the old SKU list. That is the blocker for the one-pager, not a new workshop.
One-pager to Pat before noon. Raj is on PTO through Wednesday; do not assign him the follow-up.

Three bullets. Freeze, warehouse, one-pager. Raj appears only as a constraint, a fact the Sharpie never held. The 6-line version is longer to type once and shorter to use. The 8-word version is faster to type and expensive in the ten minutes you do not have.

Run the pair in a fresh chat if you want the receipts. Same facts block. First message: the eight words. New chat, six lines. Count words in the reply, then count names that were not in the paste.

Magic-word habits vs clear-human replacements

These lines show up in prompt packs, Twitter screenshots, and well-meaning Slack pins. Swap them. Keep the skeleton, not the folklore.

Magic-word habit	Clear-human replacement
Act as a Harvard professor / world-class expert	Name the job and the reader: “Recap for Pat, director, 90 seconds.”
You are a senior VP of communications	Paste one sentence that already sounds like you, or name the channel (Slack vs email).
Let’s think step by step / chain of thought	Split the work: “First list only the facts in the paste. Then write 3 bullets from that list.”
Be concise / keep it short	Give a number: 3 bullets, 80 words, 6 lines, no greeting.
Make it professional / punchy / executive	Name the audience and forbid the phrases you hate.
Ignore previous instructions / jailbreak lore	Stop. File a normal work request. Policy theater is not a recap.
A 14-line mega prompt with no facts	Six lines with a facts block. Length in the costume is not length in the brief.

You can still say “you are” when the job is a role: “play devil’s advocate on this plan.” Then add audience, facts, and format. The failure is using the role as a substitute for the brief. Lena’s mega prompt had fourteen role lines and zero “do not invent.” The model stayed in costume and kept inventing.

As of writing, a clear human prompt in ChatGPT, Claude, Gemini, or Grok beats a costume prompt in all four. If one box still pads, tighten the format line. Do not open a fifth vendor. That is the chooser’s “one chat until it fails twice” rule, applied to the ask rather than the logo.

A paste-ready skeleton

Keep this in a note. Fill the brackets. Delete any line that does not apply. Add “act as” only when the job is a role-play (a mock interview, a red-team pass). Paste into whichever chat you already opened in the chooser.

Task: [verb + object, e.g. recap the standup / draft the one-pager]
Audience: [name, role, how much time they have]
Facts I already have (do not invent beyond this list):
- [fact 1]
- [fact 2]
- [fact 3]
Format: [3 bullets / short email / table]
Length: [80 words / 6 lines / 1 screen]
Tone: [plain / warm / blunt]. Match this sample if I paste one.
Avoid:
- invented names, dates, numbers, meetings
- action items for people who were not in the facts
- [phrases you refuse, e.g. circle back]
Channel: [Slack / email / doc / chat reply only]

Fill it once for a real job this week. Then steal your own filled version as the template, not a stranger’s mega prompt. The skeleton is longer than eight words on purpose. The output should be shorter than 400.

You do not need to know tokens

A token is the vendor’s billing chip for a scrap of text. You do not need that word to write a good ask. If you later care about API cost, model choice, or why a long paste gets expensive, that lives in Practical AI. For this series, the unit is the work request and the minutes until Pat pings.

A 6-line prompt that returns 3 bullets is cheap in attention. An 8-word prompt that returns 400 words of fiction is expensive in the 9:12 to 9:20 window, even on a Free plan. Paste size still matters. Dump a 40-page PDF to get three bullets and you pay in rereading, far more than in chips. That is a context problem, which is the next part. Today: clearer beats magical, and a number beats “please be concise.”

I think the token anxiety is usually a displaced brief problem. Eight words fail, so people collect prompt packs and add costumes instead of facts. Skip the pack. Fill the skeleton. If the reply is still wrong, say what is wrong in the next message (Part 14 will cover that loop). Do not open a second product to run the same eight words in a different skin.

Common mistakes

Sending an 8-word prompt for a job you would brief in six lines to a person.
Using “act as a professor” as a substitute for audience, length, and facts.
Pasting nothing. The model was not in the huddle. The sticky has to go in the box.
Accepting a 400-word essay because the first sentence sounded confident.
Collecting prompt packs instead of one skeleton in a note.
Pasting the same thin ask into four vendors, which is how Dana in Part 1 burned 47 minutes on one thank-you.

How to practice this week

Pick one email or Slack recap you already sent. Write the 8-word version. Write the 6-line version from the skeleton. Run both in a fresh chat with the same facts. Keep the version that names Pat (or your Pat) and a number. Save the filled skeleton. Next in this series: Give context, constraints, and examples. The chooser stays at which door. Setup, if the account is still a mess, is AI setup from zero. More paths: Learn.

Quick recap

Treat the prompt as a work request. Run the intern test before you hit Send.
Fill task, audience, facts, format, and avoid. A number beats “be concise.”
Paste the sticky. The box was not in the huddle. Fluency is not a source.
Drop Harvard, jailbreaks, and 14-line costumes with no facts. Keep one skeleton.
You do not need tokens to write the ask. Cost detail lives in Practical AI if you want it later.

Sources

Research and further reading used for this article:

OpenAI Academy: Prompting fundamentals (task, context, and describe the output; no magic-word ritual)
OpenAI Help Center (product behavior moves; confirm in the app the week you write)
Anthropic: Getting started with Claude (clear asks on Claude’s surfaces)
Google Gemini and Grok (same intern brief, different doors)
AMS: Which AI product should I use? (pick a door before you polish the ask)
AMS: AI for writing and questions (the job family this prompt style serves)
AMS: Practical AI (tokens, cost, and model choice if you need that later)

Keep going

Same lessons in your feed

Short diagrams and hooks on Instagram, X, and Facebook.

@analyticsmadesimple Follow @AnalyticsSimply Follow Analytics Made Simple Follow