Three-Part AI Creative Writing Test
Probe AI creativity with hooks, cross-domain analogy, and sequential storytelling
- Difficulty
- Easy
- Time to result
- ~days to results
- Steps
- 6
- Confidence
- 94%
This test evaluates an AI writing assistant with three distinct creative tasks while holding the product and audience constant. First, request multiple headline hooks, requiring a different creative device for each and explicitly banning buzzwords. Second, run a cross-domain analogy test: ask the model to explain a relevant concept through a surprising domain unrelated to marketing or business. Third, request a sequential advertisement made of short, independently memorable clips with a defined emotional progression. Review all three outputs for audience alignment, conceptual range, clarity, specificity, and whether they provide a strong first version worth developing. Because the tasks exercise compression, metaphor, and narrative sequencing, they reveal more than a single generic writing prompt. The result is an evaluation of creative assistance, not permission to publish unedited output.
Origin
Extracted from Marketing Against The Grain, where GPT-5 is tested on three creative writing tasks for a VP Marketing Content ICP.
Core principles
- 01Test creativity through multiple task types
- 02Hold the audience and product context consistent
- 03Demand variety rather than repeated phrasing
- 04Use surprising constraints to expose conceptual flexibility
- 05Treat outputs as first drafts for iteration
How to run it
- 1
Fix the evaluation context
Select one product, one audience profile, and a consistent set of relevant constraints for the full test.
Pro tip Reuse exactly the same context when comparing multiple models.
Watch out Changing products or audiences between models makes comparisons unreliable.
- 2
Run the hook test
Ask for multiple headlines, each using a different device such as contrast, curiosity, metaphor, paradox, taboo, or future casting.
Pro tip Require each device to be labeled so variety is easy to inspect.
Watch out Do not mistake ten near-identical headlines for creative breadth.
- 3
Apply the no-buzzwords constraint
Explicitly prohibit buzzwords and assess whether the hooks become more direct and concrete.
Pro tip Add a list of industry clichés when the model repeatedly uses them.
Watch out A no-buzzwords instruction does not automatically eliminate vague claims.
- 4
Run the cross-domain analogy test
Ask the model to explain a relevant concept through an unrelated but fitting domain that makes the idea tangible and memorable.
Pro tip Reject analogies that merely rename the original concept without clarifying its mechanism.
Watch out A vivid analogy can still be strategically irrelevant.
- 5
Run the sequential-ad test
Request short consecutive clips with memorable visual hooks, standalone coherence, and a deliberate emotional progression.
Pro tip Specify duration, visual action, voiceover, and transition requirements.
Watch out Models may default to clichés in the final transformation scene.
- 6
Score and iterate
Compare the outputs for audience relevance, novelty, clarity, variety, and editability, then develop the strongest material.
Pro tip Evaluate the quality of the first version and the ease of improving it.
Watch out Never equate a promising draft with finished creative work.
In the wild
A marketer gives GPT-5 a CRM product and a stored VP Marketing profile. The model generates ten device-based hooks, explains fewer metrics through a Michelin-star kitchen analogy, and scripts a multi-clip ad moving from dashboard anxiety to boardroom confidence.
→ The marketer observes the model's range across hooks, metaphor, and visual narrative while identifying material that still needs human revision.
Common mistakes
Testing only one prompt type
A model that writes acceptable headlines may still struggle with analogy, narrative progression, or visual specificity.
Leaving creative constraints vague
Without distinct devices, unrelated domains, or clip requirements, the outputs are difficult to compare meaningfully.
Treating the test as publication
The method evaluates and seeds creative work; it does not remove the need for editing, judgment, or production.
Is it for you?
Best for
It is best for marketers comparing AI models or testing whether a prompt setup improves creative performance.
Not ideal for
It is not ideal as a complete benchmark of factual accuracy, long-form writing, or production readiness.
From the transcript
“We're gonna figure out how good it is in comparison to the other AI assistants by giving it three creative write-in tasks.”
“We call this the cross-domain analogy test.”
“Gets you a first version that you could really iterate on.”
From the episode
Turn GPT-5 into Your Creative Writer with this one trick