Originality
A distinct idea with a reason to exist.
Rooklet Research
Creative intelligence.
Built around ads.
Connecting product truth, original ideas, and the craft of a great script.
Explore the approachSee the thinking behind the ad.
01 Tension
A city that never stops.
02 Product role
Headphones change the moment.
03 Payoff
Her own pocket of quiet.
The creative idea: Make quiet a feeling you can see.
Fictional ad · illustrating our research approach.
Our research direction
Teach the thinking
behind the ad.
A line of copy is only the surface. Our proposed training approach connects the decisions underneath.
- 01 / Understand
Start with evidence.
Study the product’s role, the creative mechanism, and how the story unfolds.
- 02 / Connect
Make craft teachable.
Link ideas, beats, revisions, and critiques. Keep every example connected to its source.
- 03 / Compare
Test what adds value.
Compare prompting, retrieval, and fine-tuning using the same reviewed foundation.
Inside an example
Beyond
the transcript.
The words tell us what happens.
The relationships explain why it works.
Anatomy of an idea
- Product role
- Headphones create a personal moment in a busy commute.
- Creative mechanism
- The city keeps rushing. Her attention becomes still.
- Story & execution
- Rush → a drifting leaf → a moment of calm.
The learning target: the relationship between product, idea, and execution.
The evaluation question
Better creative.
On the same brief.
Can advertising-specific learning outperform a strong frontier baseline?
- Unseen briefs
- Blind review
- Matched budgets
Creative quality
Illustrative · not measuredExample rubric score · 0–100
Product truth
A meaningful role, grounded in the facts.
Execution
Specific action. Clear beats. A payoff.
Creative control
Take feedback. Keep the idea intact.
A closer look
The thinking.
The details.
Dataset design, model choices, and what a result would actually mean.
What would the model learn?
The proposed records connect a brief and supported product facts to a concrete idea, the relationships that make it work, ordered beats, and execution details. Source observations, interpretations, new creative choices, and unknowns stay distinct. Separately reviewed alternatives, revisions, and critiques can teach different creative tasks without rewriting the original evidence.
This is a dataset design, not a claim about a completed training corpus. Dataset size, final composition, and source-use permissions are not specified in this preview.
Why a custom model—and how would we train it?
The hypothesis is that reviewed advertising examples can improve idea development and script craft. The same versioned records can support prompting, retrieval-assisted generation, and fine-tuning, so we can compare what each approach actually adds.
No base model, architecture, or training recipe has been selected. The creative workflow does not require a separate model for every step. Specialization has to demonstrate value against a strong general-purpose baseline.
What makes the comparison fair?
Use held-out briefs, equal numbers of presented alternatives, matched revision tasks, and a common spending cap. Keep related campaigns, reuploads, and derived examples in the same data partition. Separate brief-only tests from tests that supply references.
Reviewers see the work before the model name. Preserve disagreement, allow ties, and validate model judges against human judgment. A published result should include exact model versions, prompts, sample size, review criteria, uncertainty, cost, and latency—and state whether it measures a model alone or the complete system.
What has been established so far?
This page describes the proposed approach. Model selection and training are paused. A custom production model and an advantage over frontier models have not been validated. The chart uses invented design placeholders.
Creative preference is not evidence of campaign performance. Even a strong script does not establish a successful generated video, audience response, conversion lift, or return on ad spend.
Research roadmap · October 2026
Training paused. Benchmark figures are placeholders.