ChatGPT and Claude Prompts for Business: A Simple Formula and 25 Templates

1. Task: say what you want made, for whom, and why

Start with one sentence that names the output, the audience and the goal. OpenAI's ChatGPT guide puts it as background, the specific task and what success looks like.

Why: "write an ad" leaves the model to choose the offer, the audience, the platform and the goal. It picks the most common answer for each.

BEFORE
Write a Facebook ad for my business.

AFTER
Write 3 Facebook ad primary texts for [business] selling [offer] at [price]
to [who it's for] in [city]. The goal is booked [consults / calls],
not likes. Each ad takes a different angle: [angle 1], [angle 2], [angle 3].

2. Context: paste the material, don't describe it

Paste the offer with price and what's included, three customer reviews, your call notes, last week's numbers. Too long? Attach the file.

With long material, put it at the top and your question at the end. Anthropic's docs say queries at the end can improve response quality by up to 30 percent in its tests with long, multi-document inputs.

Why: "my customers are busy professionals" is your summary, and the model guesses the rest. The customer's own words give it the details a summary drops.

BEFORE
My customers are busy professionals who want results.

AFTER
<context>
Offer: [paste the offer, price, what's included, guarantee]
What customers said, word for word: [paste 3 to 5 reviews or call notes]
What they worried about before buying: [paste]
</context>

Using only the context above, write the 3 ads.

3. Example: show one output you liked

Paste one past output you were happy with, label it, and say "match this shape and level of detail." No good one yet? A bad one labeled "avoid this" helps too.

Anthropic calls examples one of the most reliable ways to steer format, tone and structure. Here's the kind of line I paste when I want a campaign summary in my format.

Why: one example carries the date format, the separator, the comparison number and the length. You'd never think to write all of those as instructions.

<example_of_good>
2026-08-14 | HealthProCEO / Video 12 - Gift For You | SCALE
| CPL $5.41 vs $12 book benchmark | raising budget 25% Monday,
| holding creative, one test angle queued behind it
</example_of_good>

Match this shape and this level of detail for every campaign
in the data I pasted above. One line each. No commentary between them.

4. Format: never leave the shape open

Ask for a table with named columns, a numbered list in the order you'll use it, or one paragraph under a word limit. Say what to do, not only what to avoid: Anthropic suggests telling Claude what to do instead of what not to do, and notes that the style of your prompt can shape the style of the reply.

Why: an open prompt gets an essay. A shaped prompt gets something you can paste into your CRM, ad account or email.

BEFORE
Give me some ideas for follow-up.

AFTER
Return a table with these columns: Day | Channel | Message | Goal.
Days 1, 3 and 7 only. Each message under 300 characters.
Write in plain sentences a person would text.

5. Add-ons: a role and rules

Role: give it a job, not a compliment. "Act as my ad account auditor" sets a standard. "You are a world-class genius" narrows nothing. Anthropic notes even a single sentence of role makes a difference.

Rules: put the lines it can't cross before the ask, and repeat the important ones at the end. Always include permission to say "I don't know". Anthropic lists that as the first way to reduce hallucinations.

Why: a model will give you a confident false number before it admits a gap, unless you tell it a gap is allowed.

Act as my [job title, like ad account auditor].

Rules:
- Use only numbers and facts I give you. If one is missing, write "missing".
- Don't invent links, reviews, results or client names.
- "I don't know" is an acceptable answer.
- Plain words, short sentences, no dashes.

[your task, context, example and format]

Reminder: use only numbers I gave you.

6. Add-on: make it interview you first

Before: "Write me a follow-up message for leads who went quiet." You get something polite and generic.

After: the same ask plus the interview below. Now it asks how long they've been quiet, what they last said, what the offer was, and whether they booked and no-showed or never booked. Only you know those answers, and they're the gaps it was averaging over.

OpenAI's ChatGPT guide suggests the same move: have ChatGPT ask you questions, then write up a clean prompt.

Why: the model didn't get smarter. The input changed.

Before you answer, interview me one question at a time.
Push back on any answer that is vague. Don't move on until my answer
has a specific number, name, date or example.
Ask at most 6 questions. Then do the task in the format above.

7. Add-on: make it check its own work

You have an answer. Now find out which parts were guessed. Append the pass below to anything with a number or a claim in it.

The assumptions list gives you the most for the least time, because it shows every gap the model filled. OpenAI's ChatGPT guide also suggests asking it to quote directly from the provided material when accuracy matters.

Why: a check before you ship is cheaper than a correction after a client reads it.

Now check your answer. Don't rewrite it yet.
1. List every assumption you made, most doubtful first.
2. Redo every number in the open and show the math.
3. For each claim, quote the line in my material that supports it.
   If nothing supports it, mark it UNSUPPORTED.
4. Name the one line that would do the most damage if it were wrong.

8. Templates 1 to 5: marketing and ads

Fill every bracket with your real material before you send. Each template has the 4 parts built in, and the line under the title says when to use it. For how to judge the ads these produce, see what makes a good Facebook ad, and for hooks on video, how to write hooks for Reels.

Why: marketing is where generic output costs the most, because it reads like everyone else's ad.

1. AD ANGLES FROM CUSTOMER WORDS
When: before you write a new ad and you have reviews or call notes.

Act as my direct-response copywriter.
Here is what customers said, word for word: [paste 5 to 10 reviews or call notes]
Our offer: [offer, price, what's included]
Find 5 ad angles in their words. For each: the angle in one line, the exact
customer phrase it came from, and one hook under 12 words.
Return a table. Don't invent quotes.

2. AD COPY VARIATIONS
When: you have one offer that works and need versions to test.

Write [3] Facebook and Instagram primary texts for [offer] at [price] for [audience].
Match the tone and length of this ad that worked: [paste it]
One per angle: [angle 1], [angle 2], [angle 3].
Each: a hook line, 2 to 4 short sentences, one call to action: [book / call / message].
No claims I didn't give you. Label each by angle.

3. LANDING PAGE REVIEW
When: a page gets clicks but not enough leads.

Act as a conversion reviewer. Here is my page copy: [paste]
The ad that sends traffic: [paste]  Visitor: [who clicks]  Goal: [form / call / booking]
List the 5 biggest reasons a visitor would leave without acting, most costly first.
For each: the line on the page, why it loses them, and a rewrite.
Table format. If you need data you don't have, say so.

4. A MONTH OF CONTENT FROM REAL QUESTIONS
When: planning posts and you're out of ideas.

Here are questions customers asked us this month: [paste from DMs, calls, emails]
Our business: [what you sell, to whom]
Turn them into 12 post ideas. For each: the question, a hook under 12 words,
the one point the post makes, and the format [Reel / carousel / post].
Return a table. Keep the customer's wording in the hooks.

5. COMPETITOR OFFER COMPARISON
When: you've saved competitor ads or pages and want a gap to aim at.

Here are [3] competitor ads or offers: [paste each, labeled by name]
Here is ours: [paste]
Compare them in a table: offer, price if shown, promise, proof, call to action.
Then list 3 things none of them say that we can prove. Use only what's pasted.

9. Templates 6 to 10: sales and follow-up

These work for text, email or DM. If you run your follow-up inside GoHighLevel, GoHighLevel workflows shows where these messages go once they're written.

Why: follow-up is repetitive, so a template saves the most time here, and the example part keeps it sounding like you.

6. QUIET LEAD FOLLOW-UP
When: leads asked about your offer and stopped replying.

Act as my front desk lead.
Leads went quiet [number] days ago. Their last message: [paste]
Offer: [offer, price]  What changed since: [new availability / bonus / nothing]
This text got replies before: [paste]
Write 3 follow-up texts for day [1], [3] and [7]. Under 300 characters each,
one question each, no pressure lines. Interview me first if anything above is missing.

7. SALES CALL PREP
When: 10 minutes before a call with a prospect.

Here's what I know about the prospect: [form answers, notes, website text]
What I sell: [offer, price range]
Give me: 1) 3 likely goals, 2) 3 likely objections, 3) 5 questions to ask,
in that order, as short bullets. Mark anything you're guessing as GUESS.

8. OBJECTION ANSWERS
When: you hear the same objection every week.

The objection, in the prospect's words: [paste]
What's true about our offer: [price, guarantee, results you can prove, timeline]
Write 3 replies I can say on a call: one that asks a question,
one that reframes, one that uses proof. Under 50 words each.
Use only the facts above.

9. PROPOSAL SUMMARY EMAIL
When: after a sales call, before you send pricing.

Call notes: [paste]  Package: [what's included]  Price: [price]  Start date: [date]
Write a follow-up email under 200 words: what they told me they want,
what we'll do, the price, and one clear next step: [sign / book / reply].
Use their words for their goals. No hype.

10. NO-SHOW REBOOK MESSAGES
When: someone booked and didn't show.

Appointment type: [consult / call / demo]  Missed on: [date]
Booking link: [link]  Our tone: [paste one message we sent that got a reply]
Write a same-day text, a next-day text and a day-4 email.
Texts under 250 characters. No guilt, one link each.

10. Templates 11 to 15: client work and reporting

These suit agencies and any business that reports to clients. For the judgment calls behind an ad report, see the Meta ads audit guide.

Why: client work is where invented numbers do the most damage, so every template here says what to do when data is missing.

11. WEEKLY AD REPORT
When: you have last week's campaign numbers and a client or boss to update.

Act as my ad account analyst.
Numbers: [paste export or table]  Target cost per lead: [$]
Match this line format: [paste one past report line you liked]
One line per campaign with a call: SCALE, HOLD or FIX, and the reason.
Use only the numbers above. If one is missing, write "missing". Never estimate.

12. CLIENT UPDATE EMAIL
When: every week or month, from your rough notes.

Notes: [paste]  Client: [name, what they care about most]
Sections in this order: Wins, Numbers, What we're doing next, What we need from you.
Under 200 words, plain language. Flag anything in my notes that looks wrong
instead of fixing it. End with one question. Draft only.

13. MEETING NOTES TO ACTION ITEMS
When: right after a call, from a transcript or messy notes.

Transcript or notes: [paste]
Return a table: Action | Owner | Due date | Quote from the call that supports it.
If no owner or date was said, write "not set". Then 3 lines on decisions made.
Don't add actions nobody said.

14. CLIENT ONBOARDING QUESTIONS
When: a new client signs and you need the right information on day one.

What we deliver: [service]  What we need to start: [access, assets, approvals]
What went wrong with past onboardings: [paste]
Write a welcome email under 150 words and a questionnaire of at most 12 questions,
grouped by topic, each answerable in one line.

15. REPLY TO AN UNHAPPY CLIENT
When: a client sends a frustrated message and you want to answer calmly.

Their message: [paste]  What happened, the facts: [paste]
What we can offer: [fix, timeline, credit or nothing]
Write a reply under 150 words: acknowledge the specific problem, state the facts
without blame, give the next step with a date. No excuses, no promises beyond what's above.
Give 2 versions: one warmer, one shorter.

11. Templates 16 to 20: operations and SOPs

SOPs are the easiest win for most owners: talk through a task once, and the model writes it down. To see how these fit into one system, read business operating system.

Why: a written process is what lets you hand work to someone else, and dictating it takes a fraction of the time typing does.

16. SOP FROM A VOICE MEMO OR SCREEN RECORDING
When: you do a task yourself and want someone else to do it next time.

Transcript of me doing the task: [paste]
Write an SOP: purpose in one line, tools needed, numbered steps
(one action each), what good looks like at the end, and common mistakes.
Where I skipped a step or said "you know", ask me instead of guessing.

17. CHECKLIST FROM AN SOP
When: the SOP exists but people skip steps.

SOP: [paste]
Turn it into a checklist of at most 15 items a person can tick in order.
Each item starts with a verb and has a visible result, like "invoice sent in [tool]".
Mark the 3 items that cause the most damage if skipped.

18. JOB POST AND INTERVIEW SCORECARD
When: hiring for a role you've described out loud but never written down.

Role: [title]  What they'll do each week: [paste]  Pay: [range]  Must-haves: [list]
Write a job post under 300 words and a scorecard with 5 criteria,
each with one interview question and what a strong answer includes.
No buzzwords. Don't invent benefits.

19. WEEKLY PRIORITIES FROM A MESSY LIST
When: Monday morning, when the to-do list is longer than the week.

My list: [paste everything]  Goal this quarter: [one goal]
Hours I have this week: [number]
Sort into: Do this week (max 5), Delegate (who: [names]), Later, Drop.
One line on why for each item in "Do". Ask if any item is unclear.

20. DECISION MEMO: TWO OPTIONS
When: you're stuck between two choices, like two tools or two hires.

Option A: [paste details]  Option B: [paste details]
What matters most to me, in order: [cost, time, quality, risk]
Return a table comparing both on those points, then a recommendation
in 3 sentences, then what would change your answer. Say "I don't know"
where my information isn't enough.

12. Templates 21 to 25: writing in your own voice

This is the fix for AI writing that sounds like AI. Adjectives like "friendly and professional" don't describe a voice. Your past writing does. Anthropic's docs also note that the formatting style of your prompt can influence the response, so a prompt full of bullet points tends to get bullet points back.

Why: the model can only copy a voice it has seen. Give it 3 to 5 real samples and a list of words you'd never use.

21. BUILD YOUR VOICE GUIDE
When: once, before any of the other writing templates.

Here are [5] things I wrote that sound like me: [paste emails, posts, texts]
Describe my voice so another writer could copy it: sentence length,
how I open, words I use often, words I never use, how I end.
Quote my samples as evidence. Under 250 words. I'll save this.

22. REWRITE A DRAFT IN MY VOICE
When: a draft is right on content but doesn't sound like you.

My voice guide: [paste from template 21]
Two samples of my writing: [paste]
Draft: [paste]
Rewrite it in my voice. Keep every fact, cut anything I wouldn't say.
Don't add stories, results or numbers. Show the rewrite, then list what you cut.

23. REMOVE THE AI TELLS
When: a final edit pass on anything written with AI.

Text: [paste]
Find and fix: dashes used as pauses, "it's not X, it's Y" lines, groups of three
used for rhythm, questions answered in a fragment, filler openers like
"in today's world", and words from this list: [your banned words].
Return the edited text, then a list of each change.

24. VOICE MEMO TO POST
When: you talked through an idea and want a post from it.

Transcript: [paste]  Platform: [LinkedIn / Instagram caption / Facebook]
My voice guide: [paste]
Write one post under [150] words using my phrases from the transcript.
Open with the most specific line I said. No hashtags unless I ask.
Give 2 openings to choose from.

25. NEWSLETTER FROM THE WEEK'S NOTES
When: you owe your list an email and have notes but no draft.

This week's notes: [paste wins, lessons, questions from customers]
Last newsletter I liked: [paste]
Write this week's email in the same shape and length: subject line options (3),
opening line, one main idea with one example from my notes, one call to action: [link].
Use only what's in my notes.

13. Save the winners so you stop retyping them

Once a prompt works 3 times, save the parts that don't change: role, rules, example and format. Then you only type the new material.

In ChatGPT: custom instructions apply to all your chats and are available on all plans. A Project holds its own files and instructions, and its instructions override your global ones inside that Project.

In Claude: account-wide instructions live in Settings, and project instructions apply only inside one Project. Anthropic's help center says free accounts can create up to five Projects, and its personalization article lists project instructions under paid plans (checked 15 September 2026). For a job that needs the same steps every time, a Claude skill opens by itself when the job comes up.

In Claude Code: standing rules for a folder go in a CLAUDE.md file.

Why: the saved version stays the same every run. A prompt you retype from memory drifts.

Save this as project instructions, then type only the new material.

Act as my [role].
Rules: use only facts and numbers I give you. If something is missing,
say "missing" and ask. "I don't know" is always acceptable.
Any send, post or spend waits for my yes in a separate message.
Voice: [paste your voice guide from template 21]
Format: [the table or structure you always want]
Example of good: [paste one]
End every answer with the one thing that needs me.

14. Optional, for the API: what changes in code

Skip this step unless you or your developer call ChatGPT or Claude from code. In the chat apps none of it applies.

In the API, the parts that never change go in the system prompt or developer message, and the new material goes in the user message. OpenAI says developer messages are prioritized ahead of user messages. Anthropic suggests setting the role in the system prompt.

Two things to check in the vendor docs before relying on old tutorials. Prefilling Claude's reply returns an error on Claude 4.6 and later models, per Anthropic's migration notes. And settings like temperature and thinking controls differ by model and change often.

Why: API techniques are the part of prompt engineering that goes stale fastest. The 4 parts don't.