ChatGPT Prompts for Copywriters: Research, Angles and Edits

A freelance copywriter gets the most from ChatGPT with prompts that follow the job: summarise research, mine customer language, generate angles, draft against a brief, then run edit passes for clarity, claims and voice. Give it real inputs at every step, and keep strategy and the final read for yourself.

Search for copywriting prompts and you find lists of fifty or a hundred. They are fine as inspiration and poor as a workflow, because a prompt without inputs returns the average of everything. Eight prompts are enough if they run in the order a job does and each is fed real material; the sample outputs below show what typically comes back and what a careful writer changes before anything reaches the client.

Follow me on Instagram@sagnikteaches

The order of work: why the research prompts come first

Most weak AI copy is weak at the input stage. If the only thing ChatGPT knows is "write a headline for a plumbing firm", it will write what every plumbing firm writes. A freelancer's advantage is access to the client, the customers and the proof. So the workflow starts there:

Connect on LinkedInSagnik Bhattacharya

1. Condense what you have (brief, existing copy, reviews, call notes).
2. Mine customer language.
3. Choose an angle with evidence behind it.
4. Draft against the brief.
5. Edit in separate passes, one concern at a time.

Each prompt below belongs to one of those steps. The examples use a generic client: a heating and cooling contractor selling maintenance plans to homeowners. Everything is illustrative.

Subscribe on YouTube@sagnikteaches

Prompt 1: condense the brief and find the gaps

Here is a client brief with names and figures replaced by placeholders:
[paste]
Do three things. (1) Summarise the brief in five lines: audience, offer, goal, constraint, deadline. (2) List what is missing or ambiguous that I need to ask the client before writing. (3) List any claims in the brief that would need proof before they can go in copy.

A realistic output starts with a summary: audience is homeowners with ageing boilers; offer is an annual plan; goal is 50 sign-ups before autumn; constraint is no price discounts; deadline is the 20th. The gaps list is the valuable part: "What does the plan include besides one visit? Is there an emergency call-out guarantee, and what is the response time? Who is the competitor the client fears most?" The claims list might flag "trusted by thousands" and "fastest response".

What you fix: send the questions to the client yourself and do not let the AI guess answers. A brief with unproven claims is also a protection issue, so ask for evidence or remove the claim.

Prompt 2: summarise existing material without losing the detail

Below are 30 anonymised customer reviews for a heating and cooling contractor. Group the comments by theme. For each theme give: a one-line description, how many reviews mention it, and two short verbatim quotes. Do not paraphrase the quotes. Put anything that appears only once under "singletons".
[paste]

Sample output (illustrative): "Arrival and communication (11 reviews): 'rang me when he was ten minutes away'; 'actually turned up in the window they gave me'. Explaining the problem (8): 'showed me the old part and told me what it did'. Price surprise (4): 'no hidden extras, the quote matched the invoice'." Counting matters. The AI sometimes miscounts, so count the biggest theme yourself, which takes two minutes and tells you whether to trust the rest.

Voice-of-customer mining

This step is the difference between copy that sounds like a company and copy that sounds like a customer, and voice-of-customer research from reviews and calls takes it further than one prompt can. The raw material is whatever real customer language you may lawfully use: reviews the client has permission to quote, survey answers, support emails with names removed, call notes. Do not scrape other businesses' reviews and quote them as if they were your client's.

Prompt 3: pull out phrases and worries

From the customer comments below, extract: (a) the exact phrases customers use to describe the problem before they buy, (b) the worries they had that nearly stopped them buying, (c) the words they use for the result afterwards. Quote the words exactly. Separate the three lists. Do not invent any phrase that is not in the text.
[paste]

A usable output gives you lists such as: problem phrases ("boiler making a noise", "cold house in January", "never know who to trust"), worries ("worried it would be a big upsell", "didn't want to be at home all day waiting"), and result words ("sorted", "peace of mind", "didn't have to think about it"). Those become headlines, subheads and objection-handling lines. The edit you make: check each phrase against the source text. The model occasionally smooths a quote into something nicer than what was said. A phrase you cannot find in the source is not a customer phrase.

A quick sum on what this is worth

Reading 30 reviews carefully takes around 40 minutes. The prompt does a first pass in about two, and you spend another 15 verifying and choosing. So you save perhaps 20 minutes on a small set and far more on a set of 300 reviews or survey answers, which no one reads thoroughly by hand. The saving is real, but the bigger gain is that you start writing with the customer's words on the screen.

Prompt 4: generate angles, then pick one on evidence

I am writing a landing page for an annual maintenance plan from a heating and cooling contractor. Audience: homeowners with a boiler or air-conditioning unit over eight years old. Proof I have: [list true facts]. Customer worries: [from prompt 3].
Give me eight different angles for the page. For each angle: a one-line promise, the headline in under 10 words, and the proof from my list that supports it. If an angle has no proof in my list, mark it UNSUPPORTED.

The sample output tends to include angles such as "The winter breakdown you can avoid" (supported by call-out data, if you have it), "One visit, no upsell" (supported by the review theme), "A fixed price on the plan" (supported by the client's terms), and a few UNSUPPORTED ones like "Guaranteed lowest price". The UNSUPPORTED tag is the point of the prompt. It makes the model show its working and makes you choose angles you can defend.

What you fix: pick one or two angles yourself using the client's goals and your own judgement. If the page is for people who already know the client, a different angle suits than for cold traffic. AI does not know which, because you have not told it. For how business owners use similar prompts, ChatGPT prompts for small business owners is a useful comparison.

Prompt 5: draft against a structure you choose

Write a first draft of the landing page using this structure: headline, subhead, three short benefit sections, an objection-handling section, one call to action. Angle: "One visit, no upsell". Use these customer phrases where natural: [list]. Use only these facts: [list]. Tone: calm, specific, no exclamation marks. Length: 350 words. Where you need a fact I have not supplied, write [FACT NEEDED].

The "[FACT NEEDED]" instruction works well. Without it, ChatGPT will cheerfully write "over 20 years' experience" and similar details. With it, the draft shows its own holes, and each hole is a question for the client. A typical first draft is competent and a little flat; the headline is usually fine and the middle paragraphs often repeat the same point. Treat it as scaffolding. A copywriter's value is in what happens next.

The edit passes: one concern at a time

A single "make this better" prompt produces a vague rewrite. Separate passes are more controllable and easier to check. Run these in order, in the same chat so the context carries through.

Prompt 6: the clarity pass

Edit the draft for clarity only. Do not change the meaning or add claims. Shorten sentences over 25 words, replace abstract nouns with concrete ones, and flag any sentence where the reader would ask "meaning what?". Show the changes as a list of before and after for each edit.

Asking for before-and-after pairs stops silent rewrites. Sample item: "Before: 'We deliver peace of mind through proactive servicing.' After: 'We service your boiler before it breaks, so winter does not catch you out.'" You accept the second, reject another edit that has dropped a qualifier, and move on.

Prompt 7: the claims pass

List every factual or comparative claim in the draft (numbers, "best", "fastest", guarantees, comparisons, testimonials). For each: quote it, say what evidence would be needed, and mark it as SUPPLIED (I gave you the fact) or ASSUMED (you added it). Then rewrite each ASSUMED claim as a version that does not need proof.

This pass protects both you and the client. Advertising rules in many places require claims to be substantiated, and "fastest" or "most trusted" with no source is the usual offender. The fix is almost always a more modest statement the client can stand behind ("we aim to attend within one working day" if that is true, with the exceptions stated). Never let a pass like this tick everything green without you checking a sample.

Prompt 8: the voice pass

Compare this draft with the three samples of the client's own writing below. List five places where the draft's voice differs (word choice, sentence length, formality, humour). Then suggest edits that bring the draft closer without copying phrases from the samples.
[samples, with confidential detail removed]

You will get a list such as: "The client writes 'give us a ring'; the draft says 'contact our team'", "The client uses short, declarative sentences; the draft uses three subordinate clauses". Accept the edits that match the client, reject those that do not. If a client has no samples, ask for two emails they are happy with. For a deeper treatment of tone, see writing AI prompts that sound like your business.

One landing page, start to finish, with timings

Illustration: a freelancer takes a typical job: a landing page for a heating and cooling contractor, budget for one day. The time plan below is typical.

About three hours of your time for a page that otherwise takes a day. The point is not only speed. The claims pass and the evidence-tagged angles mean fewer client rounds of "where did you get that?". Some freelancers pass part of the saving to the client, others use it for a second deliverable. Turning hours saved into work clients pay for discusses that choice.

The same workflow on a security installer's email sequence

To show the prompts are not tied to one sector, here is a compressed run on another illustrative job: a three-email follow-up sequence for a security installer that has just sent a quote for a camera and alarm system to a small shop owner.

Research step: the freelancer asks the client for six anonymised past customer emails and the questions prospects asked before saying yes. Prompt 3 on those returns the worries in the customers' own words: "will it go off when the cat walks by", "who do I call at 2 a.m.", "how long are you in the shop for". Those three worries become the three emails, one worry each.

Angle step: prompt 4 is trimmed to three angles because an email sequence needs less. The freelancer picks "answer the worry before they ask it", supported by the real questions, not an invented benefit.

Draft step: the structure for each email is a subject line, two short paragraphs and one reply-friendly question. A sample first subject line from the draft: "Will the alarm go off when your cat walks by?" It is better than the model's earlier suggestion, "Secure your premises today", which the freelancer rejects because it is exactly what every installer writes.

Edit step: the claims pass flags "we respond to alarms within 30 minutes", which the freelancer had not been given. The client confirms the real figure is "typically within the hour during business hours, with an out-of-hours line", so the copy says that. The saving here is not time. It is that a wrong promise never reaches the customer.

Adapting the prompts to other deliverables

The eight prompts are a skeleton. Small changes make them fit other work.

• Product pages. Swap the structure in prompt 5 for "title, one-line summary, five bullets of features with the benefit, a short paragraph on who it is for, a specification table". Keep the [FACT NEEDED] rule, because product pages are where invented specifications cause returns and complaints.
• Ads. Ask for 15 headlines under the platform's character limit, grouped by angle, then choose five. Check the limits on the platform's own help page, since they change, and run the claims pass on every line.
• Case studies. You supply an interview transcript with the client's permission and names removed; prompt for a structure of situation, problem, what changed, result. The model must not add a result you did not hear. Ask it to mark every number with the line of the transcript it came from.
• Newsletters and social posts. Use prompt 8 against a sample of the client's last ten posts, then draft in batches of five, each with a different hook.
• Proposals. The same research prompts condense the client's needs into a summary; keep the pricing and scope entirely your own.

Headline testing without kidding yourself

A model can produce 30 headlines in seconds, which makes it easy to confuse volume with quality. A sensible routine: ask for 20 headlines across four different angles, cut to five by reading them against the customer phrases from prompt 3, and then read the five aloud as if you were the customer. Anything that sounds like an advert rather than a person is out. If the client runs split tests, take the final two into a test and let the data decide. If they do not, resist claiming that one headline will outperform another. You do not know, and you can say so honestly: "Here is the one I would use and why."

How to judge whether the copy is any good

Before delivery, put the page through five simple checks that do not depend on the tool:

1. The stranger test. Show the first screen to someone outside the business for five seconds. Can they say what is offered and for whom?
2. The evidence test. Highlight every claim. Each has a source on file or has been softened.
3. The swap test. Replace the client's name with a competitor's. If the copy still works, it is too generic.
4. The read-aloud test. Anything you stumble over, cut or rewrite.
5. The ask test. There is one clear next step, and it is easy.

The swap test deserves emphasis. It is the quickest way to detect AI averageness: if your landing page could belong to any of the client's competitors, the research step was too thin, and the fix is more customer language and more proof, not another round of polish.

Pricing and talking about it with clients

If you price by project, using ChatGPT lowers your cost, not necessarily your fee. The value you charge for is the strategy, the evidence and the judgement, which the prompts do not provide. If you price by the hour, say so plainly when part of the work is faster, and consider moving some of the saved time into research or testing, which clients value. Either way, an honest one-line description in your proposal works better than silence: "I use AI tools for research summaries and editing checks; strategy, drafting decisions and final copy are mine, and client material is handled under the terms above." That sentence needs the confidentiality habits in this tutorial to be true.

Client confidentiality rules for a freelancer

This is the part clients ask about, so have answers ready.

• Read the contract and the NDA. Some forbid uploading material to third-party tools. If yours does, do not.
• Use placeholders. Replace the client's name, product names, prices and unreleased details with [Client], [Product], [Price]. Put the real ones in at the end.
• Keep client work apart. One ChatGPT Project per client keeps context separate. OpenAI says project instructions apply only within that project and override global custom instructions, so you can set client-specific rules without leaking them into other work.
• Know your settings. OpenAI's help page notes that information from your use of custom instructions can be used to improve model performance unless you opt out, so keep client facts out of the instructions field and check your data controls.
• Never reuse one client's material as an example for another. Even rewritten, it can reveal strategy.
• Delete when done. Remove working chats and uploaded files after delivery, and keep deliverables in your own storage.

For the longer discussion, whether customer data is safe in ChatGPT is the place to start.

Four ways this workflow goes wrong

• Angles without evidence. The model produces a clever angle you like and the client cannot support. Always tag proof, as in prompt 4.
• Quote drift. A customer's clumsy sentence becomes an elegant one on the way through. Check quotes against the source before they go on a page in quotation marks.
• Stacked edits. After six passes the copy is smooth and lifeless. Stop after the three above, then make your own final changes.
• The same page for everyone. If you reuse your prompts unchanged across clients in one sector, the pages converge. Change the inputs: the customer language, the proof, the voice samples.

Making the prompts reusable

Put the eight prompts in a document with the placeholders marked, and keep a one-line note of what each is for. Store the document in your own files so it does not depend on any one tool. When a prompt keeps needing the same correction, fix the prompt: for instance, if the claims pass keeps missing comparative claims, add "including words like 'better', 'leading' and 'most'". Treat the prompts as part of your craft, the same as your swipe file and your editing checklist. If you want help turning your own process into a set of prompts and a confidentiality routine, that is what my AI implementation consultation covers.

Questions freelance copywriters ask about ChatGPT

Should I tell clients I use ChatGPT for copywriting?

Check your contract and the client's own rules, and be honest if asked. Many clients care about the result and about confidentiality, not the tool. A short line in your terms about which tools you use and how client material is handled avoids awkward conversations later.

Can I paste a client's brief into ChatGPT?

Only if the contract allows it and the brief holds nothing confidential. Strip names, unreleased product details and figures, and use placeholders. If the client forbids third-party AI tools, do not use them on that job, however convenient.

Will ChatGPT copy sound the same as everyone else's?

It will if you prompt with a topic alone. Give it real customer language, a specific audience and one claim that is true, then edit hard. The sameness comes from missing inputs, which the research prompts below supply.

Which parts of copywriting should I keep for myself?

The strategy and the final sentence. Deciding the angle, choosing the claim you can prove, and reading the draft aloud before it goes out are your jobs. Use ChatGPT for options, summarising raw material and catching problems.

Further reads

• How Freelancers Use AI for Invoices, Proposals and Chasers — The freelancer's admin side: invoices and proposals.
• Best AI for Copywriting Ads: Variants, Testing and Client Sign-Off — Ad variants, testing and client sign-off.
• How Business Coaches Use AI Without Losing the Personal Touch — Another solo professional keeping a personal voice.
• AI Software for Copywriting: A Freelancer's Monthly Tool Budget — What a freelancer's AI tool stack costs each month.
• ChatGPT Prompts for Freelancers: Scope Creep, Chasers and Updates — Client emails and scope-creep replies for freelancers.
• How to Map Your Customer Journey and Find Where AI Helps — Map the journey your copy has to serve.
• AI Tools for Copywriting: A Freelancer's Draft-to-Delivery Workflow — One landing-page job walked from brief to delivery: client pack, AI draft, human edit, claims log, plagiarism check and an honest AI-use line.
• Best AI Model for Copywriting: How to Test ChatGPT, Claude and Gemini on One Brief — An owner's brief-test scorecard: one brief, one voice sample and one edit round through ChatGPT, Claude and Gemini, and where each needs heavy editing.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI Help Center pages on ChatGPT custom instructions and Projects in ChatGPT; no other external data used. All prompts and outputs are illustrative.