In-Context Brand Learning Loop
Condition AI on brand history and outcomes before asking for new creative
- Difficulty
- Advanced
- Time to result
- ~months to results
- Steps
- 6
- Confidence
- 93%
The In-Context Brand Learning Loop addresses the tendency of general-purpose models to produce average-looking creative. A brand assembles its historical posts, creative assets, style cues, and the outcomes associated with those assets. Successful and unsuccessful examples are both valuable because they help establish the boundaries of the brand's effective creative space. Instead of immediately fine-tuning a model, the team supplies the relevant archive as context and asks the model to infer the brand's patterns during the task. New candidates are then generated in a style informed by the company's actual history. Human reviewers check both brand consistency and novelty, because conditioning can otherwise produce increasingly similar versions of past winners. The approach is strongest when the archive is broad, metrics are comparable, and contextual limits permit sufficient evidence.
Origin
Darius Lam proposed this approach on Marketing Against The Grain while explaining how in-context learning could eventually make AI a stronger creative partner for established brands.
Core principles
- 01Generic models regress toward average outputs
- 02Brand history contains usable evidence about style and audience response
- 03Provide both successful and unsuccessful examples
- 04Condition generation on context before considering model fine-tuning
How to run it
- 1
Assemble the Archive
Gather representative posts, advertisements, campaign assets, briefs, and brand guidelines across a meaningful period.
Pro tip Include different channels and creative formats so the archive does not overrepresent one campaign.
Watch out Do not include assets the brand lacks permission to reuse for model processing.
- 2
Join Assets to Outcomes
Connect each asset with comparable performance data and any relevant audience, placement, timing, or spend context.
Pro tip Preserve contextual variables that may explain performance differences.
Watch out Raw likes or conversions are misleading when exposure and distribution differ substantially.
- 3
Label the Boundaries
Identify strong, weak, on-brand, off-brand, evergreen, and trend-dependent examples.
Pro tip Include reasons for human judgments where those reasons are known.
Watch out Training only on winners hides useful negative examples and may exaggerate survivorship bias.
- 4
Provide Task-Relevant Context
Select the archive subset most relevant to the current audience, product, channel, and objective, then place it in the model's context.
Pro tip Use retrieval to avoid flooding the model with unrelated history.
Watch out More context is not automatically better if it contains contradictory or stale examples.
- 5
Generate and Review
Ask for new candidates, then assess whether they retain the brand's style, satisfy current constraints, and add enough novelty.
Pro tip Compare outputs against both high performers and the full brand range.
Watch out Similarity to historical winners does not guarantee future performance.
- 6
Return New Evidence
Add approved assets and their eventual outcomes to the archive so future contextual learning reflects newer evidence.
Pro tip Retire data whose audience, platform, or brand assumptions are no longer valid.
Watch out An archive that never updates will reinforce stale creative conventions.
In the wild
A brand with three years of posts gathers both the creative assets and data showing which ones performed well or poorly. The model receives this history as context before generating new work in the brand's established style.
→ The outputs become more brand-specific than those produced by an unconditioned general model.
Common mistakes
Using Winners Alone
Without weak examples, the model receives less information about the boundaries and failure modes of the brand's creative space.
Ignoring Distribution Context
An asset's measured performance may reflect spend, placement, timing, or targeting rather than the creative itself.
Confusing Similarity With Creativity
Conditioning on past successes can generate more of the same rather than the unpredictable concept that starts a new creative direction.
Is it for you?
Best for
It is best for established brands with a substantial archive of content and trustworthy outcome data.
Not ideal for
It is not ideal for new brands without enough history or for teams whose historical metrics are noisy and poorly attributed.
From the transcript
“I think the key solution is what they call uh in-context learning.”
“if you are a brand that's been running for three years, you have three years worth of posts already, including data on posts that have…”
“the future is probably going to be you can put all of that past knowledge into the model and let the model then learn in…”
From the episode
This AI Startup Is Beating Apple and NVIDIA at Their Own Game