A typical brainstorm produces three variants of the same safe idea, then everyone agrees the middle one is “the strongest direction.” Real differentiation comes from forcing genuinely different stances and scoring them honestly — not from a louder whiteboard. This walks through a 5-step AI workflow that produces 12 differentiated campaign concepts in about 30 minutes of active work, scored and culled to a 3-4 concept shortlist you can bring straight to a stakeholder review.
TL;DR
- Write a tight one-paragraph brief: audience, product, goal + one metric, hard constraints, and an explicit “do not repeat these angles” list.
- Generate 12 concepts on 12 forced stances (emotional, contrarian, data-driven, satirical, etc.) so the model can’t collapse to the safe middle.
- Add a briefing layer (headline, hook, lead channel, one concrete execution) and score each concept 1-5 on three axes: on-brief, differentiation, executability.
- Cull to the top 3-4, keep the rejected eight as a scored appendix, and present both. The appendix is what makes you look strategic.
- Model pick (as of June 2026): GPT-5.5 for the widest, most unexpected divergence; Claude Sonnet 4.6 when brand voice and longer copy fidelity matter more. Both cost $20/mo on Plus / Pro respectively.
Who this is for
Marketers and creative leads who need to present multiple campaign directions, not just one. Indie founders launching without a full agency. Agency strategists doing the first pass before the larger team gets involved. Anyone whose “brainstorm meetings” reliably produce three rewrites of the same safe idea.
Pick your model first
The job here is divergent thinking — generating many genuinely different angles before you converge. Models behave differently on this. A June 2026 study on divergent thinking in LLM agents splits models into a high-ideation group that produces broad candidate sets and a narrow group that anchors hard on the first plausible subgoal. For early-stage “what if” ideation, the reasoning-heavy GPT-5.5 tends to surface more varied and unexpected angles, while Claude Sonnet 4.6 is strong on breadth and noticeably better at brand-sensitive, longer-form copy.
| Model (as of June 2026) | Plan / price | Best for in this workflow | Notes |
|---|---|---|---|
| GPT-5.5 (Thinking) | ChatGPT Plus $20/mo | Widest, most unexpected stance divergence | ~160 messages / 3 hrs on Plus; pick “Thinking” in the model picker |
| Claude Sonnet 4.6 | Claude Pro $20/mo | Brand-voice fidelity, longer execution write-ups | 1M-token context at standard pricing; Pro is 5x the free tier |
| Gemini 3.1 Pro | Google AI Pro $19.99/mo | Pulling in live references / Workspace docs | 1M context; useful if your brief lives in Google Docs |
Smaller, faster models (free-tier defaults, “mini” rungs) collapse to the safe middle on a 12-concept divergence task — they re-skin one idea twelve times. Stay on a flagship reasoning model. See the ChatGPT vs Claude vs Gemini comparison if you’re choosing a single subscription.
External references: OpenAI pricing · Anthropic plans.
Before you start: write the brief
Spend 20 minutes here, not 5. Garbage in, garbage out — a vague brief produces vague concepts and vaguer scores. A good campaign brief fits on one page and answers five things (Meltwater, HubSpot both converge on this):
- Audience: a 2-3 line persona, not “everyone.” Demographics plus one behavioral detail (what they currently do instead of buying you).
- Product: one paragraph. What it is, the one thing it does better than the alternative.
- Goal + one metric: SMART and singular. “Increase signups 20% in Q3” — not “increase awareness.” A single metric is what lets the model (and you) score honestly later.
- Hard constraints: budget, channels you can and can’t use, timing, and the legal/brand mandatories that “can’t change.”
- Do-not-repeat list: the angles from prior campaigns you are explicitly not recycling. This list matters more than people expect — it’s what stops the model quietly re-running last quarter.
The 5-step workflow
1. Paste the brief
Drop the brief into a fresh chat using the five headings above. Without this scaffolding the output is generic, because the model fills the gaps with the most average campaign it can imagine.
2. Generate 12 concepts on forced stances
The single most important prompt. Forcing distinct stances is what prevents twelve variants of “we’re great”:
Generate 12 campaign concepts for the brief above. Each must
take a different stance from this list:
emotional, functional, aspirational, contrarian,
community-driven, data-driven, urgency-based, founder-story,
educational, satirical, partnership-led, and platform-native.
Do not produce two concepts on the same stance.3. Add the briefing layer
For each of the 12 concepts, give me:
- A 5-word campaign headline
- A 1-line creative hook
- The lead channel (Instagram / YouTube / email / OOH / etc.)
- One specific execution idea — a real piece of content,
not an abstraction. Name the format.4. Score on three axes
Score each concept 1-5 on:
- on-brief: does it serve the goal + metric?
- differentiation: vs the do-not-repeat list and competitors
- executability: can we actually ship it in budget and timing?
Be brutal. The point is to surface real winners, not flatter.
Return a table sorted by total score.If every concept scores 4-5, the model is flattering you. Re-prompt: “Which two of these would I be embarrassed to pitch, and why?“
5. Cull, justify, and assemble
Keep the top 3-4 by score. For each, ask for a one-paragraph rationale: why this beats the eight you rejected. Then assemble two things for the deck — the shortlist with rationales, and an appendix of the 8-9 considered-and-rejected concepts with their scores. Save the brief + the four prompts as a template; next campaign you swap only the brief paragraph.
The 30-minute timeline
A real launch morning, with the brief already written:
| Time | Step | Output |
|---|---|---|
| 0:00 | Paste brief | Model has full context |
| 0:05 | 12 stances generated | 12 raw concepts |
| 0:12 | Briefing layer added | Headline + hook + channel + execution each |
| 0:22 | 3-axis scoring | Ranked table |
| 0:30 | Cull + rationales | Top 3-4 + appendix in the deck |
Thirty minutes of active work replaces the four-hour brainstorm. The 2pm stakeholder meeting then gets a far sharper set of options, and becomes about combining and stretching the shortlist rather than generating from a blank wall.
Run it once on a campaign you already shipped
The fastest way to trust the workflow is to run it backward. Take a campaign you launched last quarter, write its brief, and run all five steps. Then ask three questions:
- Did the AI surface angles you hadn’t considered? If yes, that’s the workflow’s value. If no, your stance list wasn’t strict enough — sharpen the labels.
- Scoring the 12 concepts as if briefing fresh, were the top 3 stronger than what you actually ran? “Stronger for 1 of 3” is still a win.
- Did the differentiation survive scoring, or did the top 3 all collapse back to the safe middle? If they collapsed, your rubric is over-weighting executability — re-balance toward differentiation.
Common mistakes
- Asking for “10 campaign ideas.” You get 10 variants of the safe one. Always force stance differentiation.
- Skipping the constraints section. The model invents concepts that violate budget or channel reality, and you burn time discarding them.
- Not scoring. Without a rubric every idea feels equally good and you default to the loudest advocate’s pick.
- Omitting the do-not-repeat list. You quietly recycle last quarter’s campaign and don’t notice until someone in the room does.
- Treating the AI’s scores as final. They’re a first pass. Your context — internal politics, brand fit, channel relationships — overrides.
- Showing only the top 3. The scored appendix of rejected concepts is what makes you look strategic and pre-answers “what about ___?”
Tailor the stance list to your category
The generic 12-stance list works everywhere, but a category-specific list works better. Heavily regulated industries are the clearest case: in pharma, finance, or B2G, satirical and contrarian stances often can’t ship, so swap them for “compliance-as-trust,” “third-party-validation,” or “education-led” stances that fit the constraint. Keep a “considered angles” log per campaign; after four campaigns you’ll see your own stance habits — always reaching for founder-story, never picking platform-native — and can correct for them deliberately.
FAQ
- Which model should I use?: As of June 2026, GPT-5.5 (ChatGPT Plus, $20/mo) tends to produce the widest, most unexpected divergence for early ideation; Claude Sonnet 4.6 (Claude Pro, $20/mo) is strong on breadth and better for brand-sensitive, longer copy. Both handle a 12-concept divergence comfortably. Avoid free-tier and “mini” models — they compress to the safe middle.
- Can it replace brainstorm meetings?: Not entirely. Use AI for divergence and the meeting for convergence and stretching. The 12 concepts go into the meeting; the room picks and combines the final winner.
- My industry is regulated and half the stances don’t apply — now what?: Build a tailored stance list. Replace the stances that can’t legally ship (often satirical, contrarian) with category-appropriate ones like third-party-validation or education-led. The scoring step then does the rest.
- How long should the brief be?: One page. Audience + product + goal/metric + constraints + do-not-repeat list. A strong brief fits on 1-3 pages and is the single doc everyone returns to; one page is plenty for ideation.
- What about visual concepts, not just copy?: Add a step for the top 3: generate a visual mood board per concept. See the AI ad creative tutorial for the visual brief format.
- Will the same stance always win?: No. Different audiences and goals favor different stances; the workflow surfaces what fits this brief and the scoring catches it. If founder-story wins every time, your rubric is anchored on it.