ChatGPT Prompts for a Marketing Agency: Briefs, QA and Client Updates

An agency can use ChatGPT for briefs (turn call notes into a structured brief), QA (check copy against a checklist) and client updates (turn a task log into a short report), provided a person reviews each output. Keep out anything under a confidentiality clause, unreleased product details, customer data, account logins and the client's unpublished results.

Most agency prompt lists are written for individual marketers. An owner has different questions: what may the team paste in, who checks the output, and what happens if a client asks. The data rules decide which of the eight prompts your team may use on which client, so settle them first; every prompt below carries a sample output and the review step that goes with it.

Follow me on Instagram@sagnikteaches

Which client data stays out: a three-tier rule for the team

Pin a simple tiering to the wall. Tier one is public material: the client's published website copy, public product pages, social posts already live. This can go into ChatGPT freely. Tier two is working material: call notes, draft copy, campaign plans. This goes in only with the client's agreement, with names and figures replaced by placeholders where possible. Tier three is never entered: logins, contracts, unreleased launch details, customer lists, personal data about the client's customers, unpublished financials and anything the client has marked confidential.

Connect on LinkedInSagnik Bhattacharya

The account matters too. A ChatGPT Business workspace (two seats minimum, per OpenAI's help centre) keeps workspace content out of model training by default, and its admins decide which apps (formerly connectors) staff can switch on. Personal Plus or Pro accounts share chats for training until each person switches off "Improve the model for everyone", and the agency has no way to check that setting on someone's personal account. If anyone on the team uses a personal account for client work, the tiering above matters even more, and how to stop AI tools training on your business data lists the settings to check. A team on a business plan still follows the tiers; the plan reduces one risk, not all of them.

Subscribe on YouTube@sagnikteaches

Put the rule in writing for staff in about a page: the tiers, which account to use, the approval step and who to ask. A freelancer using a personal account is the most common leak in a small agency, so extend the same page to contractors and write it into their agreement.

Roles: who writes, who reviews, who signs off

AI changes who does what, and unclear roles are where mistakes slip through. A workable split for a small agency: the writer drafts and may use ChatGPT for options and structure; the account lead owns the brief and the client-facing message, so they use the brief and update prompts; the checker, a different person from the writer, runs the QA prompt and the human checklist; the owner or a director signs off anything touching scope, fees, claims or sensitive topics. For a team of three, one person may hold two roles, but the writer should never be the only checker of their own work.

Write that split into a one-page process, and name the person who approves exceptions, such as a rush job that skips a step. Exceptions logged in one line each give you evidence later of what was skipped and why.

Per-client headers: how the team keeps voices apart

Agencies serve clients with opposing tastes, and an assistant used carelessly will blur them. Give every account a header document of 150-250 words: audience in one sentence, tone in three adjectives with an example sentence, banned and required words, approved claims, mandatory disclaimers and the tier-two data rules. The header goes at the top of every prompt for that account, and in a shared ChatGPT Project if the team uses projects (OpenAI's help centre lists project sharing on every plan, and a Business workspace project holds up to 100 collaborators and 40 files).

A filled-in header for a driving school might read: "Audience: nervous learners and their parents. Tone: calm, encouraging, never jokey about accidents. Required: 'dual-controlled car', pass-rate claims only from the school's own published figure. Banned: 'guaranteed', 'easy'. Disclaimers: pricing excludes test fees." With that in place, the writer, the QA prompt and the update prompt all start from the same facts, and a new starter can produce an on-brand draft in their first week.

Keep each header in the client's own folder, not in a shared chat that mixes clients. If an account ends, archive the header with the final deliverables and delete any copies of client material you were only holding for working purposes, as the contract requires.

Briefing prompts

Prompt 1: call notes to a one-page creative brief

Briefs go wrong when the objective is vague, so the prompt forces the structure and forbids invention. Use your own notes with the client's name replaced by "the client".

Turn these call notes into a one-page creative brief for an independent bookshop's autumn campaign.
Sections: objective (one sentence), audience, single message, mandatory inclusions,
channels, deadline, budget range, success measure, open questions.
Use only the notes. Where the notes are silent, write "NOT SAID" and add it to open questions.
Notes: [paste]

A realistic output has nine neat sections, with "success measure: NOT SAID" appearing more often than you would like. That is the prompt working, not failing: it shows you what you forgot to ask on the call. The usual flaw is a tidied-up objective that sounds firmer than the client's actual words. Check: read the objective against your notes, and send the open questions list back to the client before work starts.

Prompt 2: stress-test the brief before the team sees it

Ask ChatGPT to play a sceptical creative lead.

Act as a creative lead who has been given the brief below. List the five questions you would ask
before starting, the two places where the brief contradicts itself, and any instruction a team
could read two ways. Do not rewrite the brief.

A sample answer might flag that the brief asks for a "premium feel" while limiting imagery to free stock photos, or says "short and punchy" while mandating three legal lines. Check: decide which flagged items are real. Models over-flag, so expect two or three of the five questions to be noise, and keep the ones you recognise from past projects.

Prompt 3: variations for a single message, not a campaign

Agencies get into trouble when ChatGPT produces a whole campaign nobody asked for. Constrain it: one message, six headline options, each under 40 characters, none making a claim the client has not approved.

Single message: "Our autumn reading list is out". Write six headline options under 40 characters.
No superlatives, no price claims, no claims about awards or reviews.
Tone notes: dry, bookish, warm. Avoid these words: [client's banned words].

Expect usable-looking lines such as "Autumn's reading list has landed". Check the claims line: any headline implying "bestselling" or "most-loved" must be stripped unless the client can prove it. Keep the client's banned-word list in your per-client header so the team stops reintroducing words the client dislikes.

QA prompts and a reusable checklist

Prompt 4: a copy QA pass against a written checklist

QA prompts work only when you supply the checklist. "Check this for errors" produces generic praise. A named checklist produces named failures.

Check the copy below against this checklist and answer each item PASS, FAIL or UNSURE with the exact line it refers to:
1. Brand name and product names spelled as in the style sheet below.
2. No claim beyond the approved-claims list below.
3. Every price, date and offer matches the facts below.
4. Tone matches the tone notes.
5. One clear call to action; link text describes the destination.
6. No sentence over 28 words.
Style sheet, approved claims and facts: [paste]
Copy: [paste]

A realistic output lists item 3 as FAIL ("'ends Friday' conflicts with the stated end date, Sunday") and item 6 as FAIL for two quoted sentences. The valuable behaviour is UNSURE: you want it to admit doubt instead of guessing. Check: for every FAIL, confirm it yourself; for every PASS on items 2 and 3, a person still compares the live asset to the source, because ChatGPT only checks the text you pasted.

The agency QA checklist itself

Here is the checklist behind the prompt, written for people. Copy it into whichever project tool the team uses.

• Source facts match. Dates, prices, offers and contact details checked against the client's confirmation, not the previous draft.
• Claims are approved. Each claim traceable to the client's approved list or a source they supplied.
• Names spelled right. Client, products, people and places, against the style sheet.
• Links work. Clicked on a phone and a desktop; UTM tags present where agreed.
• Accessibility basics. Alt text written, link text meaningful, no text baked into images that matters to meaning.
• Tone and legal lines. Required disclosures and disclaimers in place, in the approved wording.
• Formatting in the real channel. Previewed in the actual email client or ad platform, not only the document.
• Second pair of eyes. Someone other than the writer signs off, with initials and date.

Use ChatGPT for the first, third and sixth items on pasted text; a person does the rest. The split is the point: the model reads text, people check the world.

Prompt 5: compare final artwork copy with the approved copy

Last-minute edits are where errors enter. Paste approved copy and the final text from the layout and ask for differences only.

Compare TEXT A (approved) and TEXT B (final from layout). List every difference as a table:
location, A says, B says. Do not comment on quality. If they are identical, say "identical".

A good output is a short table showing, say, a changed comma and "from $12" replaced with "from $15" (a real catch). Because models can miss subtle character changes, run a standard text-compare tool as well for anything legal. Check: confirm by eye that any "identical" verdict is true on the numbers.

Client update prompts

Prompt 6: weekly update from the task log

Clients mostly want three things: what moved, what is blocked on them, and what is next. Feed the log, not the raw ad-platform export.

Write a weekly update for a florist client from the log below.
Sections: Delivered this week, Needs your decision (each as a question), Next week.
No metrics that are not in the log. Plain language, no jargon, under 180 words.
Log: [paste]

From a nine-line log, an illustrative output reads: "Delivered this week: the spring bouquet email went out on Tuesday; four social posts published; the order-form typo fixed. Needs your decision: can we use the customer photo you sent on Instagram? Next week: draft the summer offer; set up the wedding enquiry form." Two things to fix in drafts like this. ChatGPT tends to add colour the log never had ("the email performed well"), so delete any phrase you cannot point to in the log. And it often drops the deadline from a decision question, as it did here; add "by Thursday", because a question without a date is the one clients leave unanswered. Check: confirm each delivered item in the log appears once, with nothing extra.

Prompt 7: explain a dip in results without inventing causes

Clients will ask why numbers dropped. The risk is a confident invented story. Make ChatGPT separate fact from hypothesis.

Here are this month's and last month's figures for a driving school's enquiry campaign: [table].
Write three sentences describing what changed, using only the table.
Then list up to four possible reasons under the heading "To investigate", each with the data we'd need to confirm it.

Sample output states "enquiries fell from 62 to 48 while spend rose from $900 to $1,000", then proposes "seasonal demand", "ad fatigue" and "tracking change" with the check for each. You decide what goes to the client; the "To investigate" list is internal until you have evidence.

Prompt 8: the difficult message

Late delivery, missed target, scope creep. Ask for a draft that is honest, brief and offers a next step.

Draft a 120-word email telling a client the launch will move from the 14th to the 21st because their approvals arrived late.
Be factual, not defensive. State the new date, what we need from them, and what happens if that slips again.
Do not apologise more than once.

An illustrative draft: "Hi [first name], the launch will now go live on the 21st rather than the 14th. The homepage and banner approvals reached us on the 10th, four days after the agreed date, and the remaining build and QA need five working days. To hold the 21st we need sign-off on the two email drafts by the 16th. If that slips, the next launch slot is the 28th. Sorry for the change; the revised plan is attached." What to fix: "four days after the agreed date" is stated as fact, so check the approval dates in your records before it goes, and "the next launch slot is the 28th" is a commitment the account lead must confirm with the team first.

Check against the contract: if you promise "no extra cost" or imply fault, it is a commercial decision, not a drafting one. The owner or account lead approves anything that touches scope, fees or blame.

A client-update template the team can fill in

Standardise the shape so clients know where to look. Subject: "[Client] weekly update, w/c [date]". Body: a two-line summary; Delivered (3-5 bullets, each starting with a verb); Needs your decision (numbered questions, each with a date it is needed by); Numbers (only those agreed in the reporting plan, with the source named); Next week (3 bullets). Close with one line on blockers.

A filled-in example for an illustrative photography studio: "Summary: gallery pages are live and the spring offer email is scheduled. Delivered: published 6 gallery pages; fixed the booking-form error; wrote the spring offer email. Needs your decision: 1. Which of the two subject lines for the spring email, by Wednesday? Numbers: 214 form views, 19 enquiries (from the booking form's own report). Next week: send the spring email; draft two social posts; review mobile speed." Notice it holds no adjectives and no claims; every line is something a client can verify.

A leaked launch date, a broken form and an added adverb

A junior pastes a client's unreleased launch date into a personal ChatGPT account to get tagline ideas. Nothing visible happens, which is the problem: it breaches the confidentiality clause quietly. The fix is the three-tier rule and an agreed account for work, with a short monthly reminder.

A QA prompt returns "all PASS" on a landing page, and the live page has a broken form. The checker only saw pasted text. The fix is the checklist split above: model for text, person for links, forms and layout.

A client update says "traffic is up strongly", but the log only said traffic was up 4 per cent. ChatGPT added the adverb. The fix is the "no metrics or claims not in the log" line, plus a check that every adjective is backed by a number.

A bookshop's autumn campaign, from call notes to QA

Picture a six-person agency, with made-up but realistic figures, taking on an autumn campaign for an independent bookshop: three emails, eight social posts and a landing page, budget $3,000, launch in three weeks. The account lead writes call notes in the usual messy bullets and runs Prompt 1. The brief comes back with "success measure: NOT SAID" and "deadline for approvals: NOT SAID". Two questions to the client later, the success measure is "40 event ticket sales" and approvals are due within two working days.

Prompt 2 then flags that the brief demands a "playful" tone and also lists three formal legal lines for event terms. The team agrees to keep the legal lines in the footer only. Prompt 3 gives six headline options for the first email; two imply "bestselling" and are dropped. Copy is drafted by a writer, not by ChatGPT, because the owner wants a distinct voice, and the QA prompt runs on each piece against the checklist. Over the three emails, it flags nine items; a person confirms seven as real, including a price that appeared as $12 in the layout and $15 in the approved copy. The agency estimates the whole cycle saved around five hours of back-and-forth against its usual run, but the figure is an estimate, not a measurement; a real agency should time it.

The lesson is not that the AI wrote the campaign. It improved the questions asked at the start and caught slips at the end. The creative work in the middle stayed human.

What to measure so you know it is working

Owners often adopt prompts on enthusiasm and then cannot say whether they help. Pick four numbers and log them monthly in a shared sheet. First, time from call to approved brief. Second, number of rounds of client amendments per deliverable. Third, QA fails caught internally versus found by the client after sending. Fourth, hours of writer time per standard deliverable. All four are ordinary agency metrics and need no new tools.

Illustrative reading: if rounds of amendments fall from 2.4 to 1.8 per email and client-found errors drop to zero for three months, the brief and QA prompts are paying off. If rounds stay the same while time per deliverable rises, the prompts are adding review work without improving quality, and you should simplify before expanding. Treat any improvement as provisional for a quarter, because a good month can come from an easy client mix rather than a better process.

Running it as a team routine

Start with two prompts, not eight: the brief prompt and the QA prompt, because they have the clearest before-and-after. Run them for a month on two or three accounts the client has agreed to. Record time spent per brief and the number of QA fails caught before the client saw them; keep those numbers in a simple sheet. A change you cannot measure is a change you cannot defend to a sceptical account manager.

Name an owner for the prompt library, usually an account director. Each prompt gets a version date, the checklist it relies on and the review step. Retest each against a sample every quarter, since model behaviour shifts. For the larger picture of what AI does well in an agency, see which agency tasks AI handles well and which it doesn't; for automation beyond prompts, seven agency workflows to build first; and for what changes in pricing, agency retainers and packages after AI.

Telling clients what you do with AI

A short statement avoids awkward conversations. Something like: "We use an AI assistant for first drafts, summaries and checklist checks. We do not put your confidential material, customer data or logins into it. Every deliverable is reviewed by a person before it reaches you." Add it to your proposals and the master agreement, and offer a copy to clients who ask for an AI policy. Clients with stricter rules may specify tools or require approval per project; record that in the account's header document so every new team member sees it on day one.

Owner questions about ChatGPT inside an agency

Which client data should never go into ChatGPT?

Anything under a confidentiality clause, unreleased product details, customer lists, ad-account logins, financial results the client has not published and anything personal about the client's customers. When unsure, describe the situation in general terms or use a placeholder instead of the real material.

Do I have to tell clients my agency uses ChatGPT?

Check your contracts and any client AI policy first. Many clients now ask, and some require approval. Telling them voluntarily, with a line on what stays out, is easier than explaining it after they find out. A short clause in your master agreement usually does the job.

Should juniors use ChatGPT for first drafts of briefs?

Yes, if a senior reviews the brief against the client's call notes before it reaches the creative team. The risk is not the draft, it is a brief that sounds plausible and misstates the objective. Make review a named step, not an assumption.

Can ChatGPT do the QA instead of a person?

No. It is a fast second reader that catches typos, tone drift and checklist gaps, but it cannot see the live link, the final artwork or the client's approved claims. Use it to prepare the checker, then have a person confirm each item against the real asset.

Further reads

• How Does Claude Help a Marketing Agency? Six Jobs and Their Limits — How Claude handles six agency jobs, for comparison.
• AI Tools for Agencies: What a Five-Person Team Pays Each Month — What a five-person agency pays for AI tools each month.
• How Marketing Agencies Use AI to Automate Client Reporting — Automating the monthly client report beyond prompts.
• Hire an AI Specialist or Upskill Your Agency Team? — Hire a specialist or train the team you have.
• How to Mention AI Use in Client Contracts and Proposals — Contract wording for telling clients about AI use.
• AI Media Planning for Small Agencies: Channel Mix and Budget Splits — Using AI for channel mix and budget splits.
• AI PR Pitch Emails: Personalise for Journalists With a Review Step — A five-step pitch workflow for small PR agencies: beat note, hook line, three-line pitch, manager sign-off and a follow-up cadence that stops on time.
• AI Agents for PR: What They Do and How an Agency Supervises Them — A PR agency supervision method for media lists, monitoring and pitch drafts, with a strict approval-before-send gate.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI enterprise privacy page, and OpenAI help articles on ChatGPT Business, data controls and projects, checked 4 November 2026.