Lesson objective
Refine an AI-assisted draft in logged, single-change rounds and build deliberate variation into a brief so a campaign does not converge on the same ideas and phrasing as everyone else's.
Plain-language explanation
Why this matters now
Most drafts need more than one round. Refinement is where drafts improve, and also where claims creep back in and where a campaign quietly starts to sound like every other campaign made with the same kind of tool. Level 1 introduced homogenised creative and the cost of sameness. This lesson turns that into two working habits: logged refinement rounds, and a variation plan written into the brief.
Professional guidance (NIST AI 600-1, July 26, 2024; voluntary): NIST lists "Harmful Bias or Homogenization" among the twelve risks generative AI creates or makes worse. The profile is not written about marketing and binds no agency.
Nothing here says AI makes marketing more or less persuasive. The point is narrower: if everyone starts from similar suggestions, the work tends to converge, and an agency is paid for work that does not.
A question to sit with
Three agencies pitch the same fictional juice brand. All three used an assistant to generate first concepts from the same brief. Two decks open with nearly the same line about "sunshine in a bottle". What did those two teams skip?
What goes wrong
Changing five things at once. A writer rewrites the prompt, swaps the tone, adds a new claim and changes the length in one go. The next draft is different, and nobody knows why.
Refinement that adds claims. "Make it punchier" produces "the healthiest snack on the shelf". Each round needs the same closed-file rules as the first.
Refining the draft instead of the input. The same error returns every time because the cause is in the brief or the claims block, not in the wording.
Accepting the first idea set. The team picks from the model's first suggestions, which are the suggestions any team using a similar tool might get.
Mistaking polish for originality. A draft can be rated well written and still be close to what competitors are producing.
Losing the human concept. The creative team never writes its own ideas before prompting, so there is nothing distinct to compare against.
The practical answer
Part 1: refinement in logged rounds.
| Rule | Why |
|---|---|
| One change per round | So the effect of each change can be seen |
| Say what the change is for | "Shorter for mobile" is checkable; "better" is not |
| Keep the closed-file constraints in every round | So refinement cannot add claims |
| Fix inputs, not symptoms | If an error repeats, change the block that causes it |
| Log each round | Round, change, reason, result, kept or reverted |
| Stop at a set number | A limit, such as four rounds, prevents endless polishing; beyond it, a person rewrites |
A refinement log for one headline might read:
| Round | Change | Reason | Result |
|---|---|---|---|
| 1 | Base template | First draft | Too long; one claim strengthened |
| 2 | Length limit 40 characters | Mobile placement | Fits; strengthened claim still present |
| 3 | Claims block restated with exact wording | Stop strengthening | Claim now matches file |
| 4 | Tone rule added from brand guide | Too formal | Kept; sent to preparer check |
Part 2: variation by design.
Write variation into the brief before anyone prompts.
- Human concepts first. The creative team writes its own territories and lines before any tool is opened, and keeps them.
- Vary the inputs, not only the request. Change the audience fact being led with, the claim leading the message, the format, or the emotional register. Each is a deliberate lever.
- Ask for difference explicitly. Ask for options that each use a different claim and a different structure, and ask the model to say how each differs from the others.
- Screen for sameness. Remove options that repeat common category phrases. Keep a short list of phrases the team has seen too often in the category and treat them as banned for first drafts.
- Compare against the human set. For each AI-assisted option, ask whether the team's own concepts already cover it. If the AI set adds nothing distinct, the human set leads.
- Record the origin. Mark each concept in the record as human, AI-assisted, or AI-generated and edited, so the agency can answer honestly if a client asks.
What the productivity evidence can and cannot support. Research (Noy and Zhang, Science, July 2023; 453 professionals; short laboratory writing tasks): time fell by 40% and quality rose by 18%, and workers exposed to ChatGPT were twice as likely to report using it in their real job two weeks later and 1.6 times as likely two months later. That supports the idea that people keep using these tools once they try them. It does not show the effect on agency campaigns, which have claims, brand and approval steps the tasks lacked.
When a client asks whether ideas were AI-generated. Answer from the record: which concepts were human, which were AI-assisted, and who reviewed them. Module 12 covers what to tell clients about AI use in scopes and fees.
Remember this
- Refine one change at a time, say why, keep the constraints, and log each round.
- If an error repeats, fix the input block, not the wording.
- AI-assisted stories in one experiment were more similar to each other; treat sameness as a risk to manage.
- Write human concepts first, vary the inputs deliberately, and screen for category clichés.
- Mark each concept's origin so the agency can answer a client truthfully.
Marketing agency example
Kestrel Media Partners is a fictional agency pitching Sunfold Juice, a fictional brand. The creative director, Leon Marsh, writes a variation plan into the brief. First, the two-person creative team spends an hour writing its own territories: "the ten-minute kitchen" and "grown two valleys over". Then the writer, Priya Addo, prompts with the approved template. The first AI set of eight lines includes "sunshine in a bottle" and "nature's goodness", both on the team's list of category phrases seen too often, so both are screened out. Priya then varies the inputs in three deliberate rounds: leading with the audience fact "most buyers shop on Sunday", leading with claim C2 "cold-pressed within 24 hours of picking", and switching the format to a two-line dialogue. She logs each round. In round two the draft says "cold-pressed within hours", which is vaguer than C2 and could be read as a faster process, so she restates C2 exactly in the claims block and reruns. After four rounds she has eleven options. Leon compares them with the human territories. Three AI-assisted options add something the team had not written; the other eight repeat the human ideas or each other. The pitch leads with "the ten-minute kitchen", a human concept, supported by one AI-assisted line that was rewritten by hand. The record marks each concept as human, AI-assisted or AI-generated and edited, so when Sunfold's marketing lead asks which ideas came from AI, Leon answers in one sentence and can show the refinement log behind it.
Responsible-use warning
Do not change several things in one refinement round, and never drop the closed-file constraints because a draft "just needs polishing". Polishing rounds are where claims get stronger. Do not describe AI ideas as making a campaign more creative or more persuasive: the experiment used short fiction, and the finding that matters for agencies is that AI-assisted work in that study was more alike. Keep the team's own concepts and a record of where each idea came from.
Human-review requirement
The creative director or another named agency reviewer owns the variation plan and the refinement log. The reviewer checks that the team wrote its own concepts before prompting, that each refinement round made one stated change with the constraints intact, that repeating errors were fixed in the input blocks, that category clichés were screened, and that each concept carries an origin mark. The reviewer compares the AI-assisted set with the human set, decides which leads, and approves only versions that passed the preparer check.
Practical exercise
Use a fictional brand, brief and claims file only.
- Write a fictional brief for a drinks, food or household brand with three approved claims.
- Spend fifteen minutes writing three concept territories and five lines of your own before any tool is used.
- List six phrases you consider overused in that category.
- Write a first prompt with the closed-file constraints from Lesson 2, or write a mock first AI set of eight lines.
- Screen the set against your overused-phrase list and remove any matches.
- Plan three variation rounds, each changing one input: the lead audience fact, the lead claim, or the format.
- For each round, log the change, the reason, the result and whether you kept it.
- Identify any round where a claim was strengthened, and fix the block that caused it.
- Compare the AI-assisted options with your own concepts and list which add something distinct.
- Mark every final option as human, AI-assisted, or AI-generated and edited.
- Set a round limit for your team and write what happens when it is reached.
- Ask a colleague to read your final options without the log and guess which were human. Compare their guesses with your origin marks.
- Write two sentences you could give a client who asks where the ideas came from, and save the log in your workbook.
Knowledge check
-
Why make only one change per refinement round?
Answer guidance: So the effect of each change can be seen, and a problem can be traced to the change or input that caused it.
-
What did Doshi and Hauser find about stories written with AI ideas?
Answer guidance: They were rated more creative and better written, but were more similar to each other than stories by humans alone. The study used short fiction, not advertising.
-
Why should the creative team write its own concepts before prompting?
Answer guidance: To have distinct ideas to compare the AI set against, so the agency can tell whether AI suggestions add anything or simply converge on common ideas.
-
Does this lesson claim AI makes campaigns more persuasive?
Answer guidance: No. The sources do not show that, and the variation plan is a practice to counter sameness, not a way to raise campaign results.
Sources and further reading
- Science Advances (AAAS); authors Doshi AR and Hauser OP — Generative AI enhances individual creativity but reduces the collective diversity of novel content (2024-07-12). Accessed 2026-09-15.
- Science (American Association for the Advancement of Science); authors Shakked Noy and Whitney Zhang (MIT) — Experimental evidence on the productivity effects of generative artificial intelligence (2023-07-13). Accessed 2026-09-15.
- National Institute of Standards and Technology — Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1 (2024-07-26). Accessed 2026-09-15.
- AI Coalition Network — Certification Educational Standard: Human Review of AI-Assisted Professional Work. Accessed 2026-09-15.
- Generative AI enhances individual creativity but reduces the collective diversity of novel content, Science Advances, AAAS, 2024