Codified Taste Learning Loop
Turn subjective preferences into variables you can study, test, and refine.
- Difficulty
- Moderate
- Time to result
- ~weeks to results
- Steps
- 6
- Confidence
- 90%
The Codified Taste Learning Loop treats taste as observable evidence rather than an inexplicable gift. The creator collects compelling work and annotates why each example succeeds or fails. An AI model analyzes the collection to surface recurring structures, contrasts, and stylistic attributes, which are then written into a style guide. The creator uses those attributes as adjustable variables, generating multiple outputs while changing one feature at a time. Comparing the results reveals which variables drive the desired response and helps the creator sharpen both prompts and judgment. External galleries, high-performing community posts, or a respected creator's body of work can supply examples, but the goal is not direct imitation. The output is a living vocabulary for recognizing quality, explaining creative choices, and learning faster through repeated comparison.
Origin
Extracted from Marketing Against The Grain, where Steph Smith and the hosts described annotating preferences, distilling successful work, building style guides, and varying their components with AI.
Core principles
- 01Taste becomes teachable when preferences are captured with reasons.
- 02Strong examples reveal patterns more clearly than generic best practices.
- 03Comparative generation accelerates learning through controlled variation.
- 04AI can structure subjective observations without replacing human judgment.
- 05Novel creation comes from exploring possibilities, not copying an average.
How to run it
- 1
Build an example set
Collect work that produces a strong positive or negative reaction in the relevant medium. Include enough variety to distinguish enduring preferences from one-off novelty.
Pro tip Use galleries or communities that expose both the work and its prompt, context, or audience response.
Watch out Do not collect only examples that share the same creator or trend.
- 2
Annotate your reactions
For each example, state what attracted or repelled you and identify specific details that produced the reaction. Separate subject matter from composition, voice, structure, and execution.
Pro tip Write annotations before asking AI for an analysis so the model does not overwrite your initial judgment.
Watch out Labels such as 'good' or 'beautiful' are too vague to become useful variables.
- 3
Extract recurring patterns
Give the examples and annotations to an AI model and ask it to identify repeated attributes, contrasts, and mechanisms. Review the analysis and reject patterns that do not match your own observations.
Pro tip Ask the model to cite the examples supporting each proposed pattern.
Watch out Do not confuse frequently occurring features with the features that actually drive quality.
- 4
Create a variable-based style guide
Translate validated patterns into a style guide covering the dimensions that can be deliberately changed. Define preferred ranges or contrasts rather than prescribing one fixed output.
Pro tip Include positive rules, anti-patterns, and examples for each variable.
Watch out An overly rigid style guide can erase experimentation and produce repetitive work.
- 5
Run controlled variations
Generate alternatives while changing one or two style variables at a time. Compare the outputs to determine how each variable affects clarity, novelty, emotion, or platform fit.
Pro tip Keep the underlying brief constant so differences can be attributed to the changed variable.
Watch out Changing every variable simultaneously prevents meaningful learning.
- 6
Refine your judgment
Record which outputs succeeded, why they succeeded, and how your preferences changed. Feed this evidence back into the style guide and repeat the loop.
Pro tip Preserve rejected outputs because they help define the boundaries of your taste.
Watch out Do not delegate the final evaluation to engagement metrics or the model alone.
In the wild
Steph browsed Midjourney's showcased images, inspected an Asian-looking 3D sculpture, and discovered the term 'blind box design.' She incorporated it into a roughly 30-word prompt that could transform subjects such as Berlin's skyline into a consistent visual style for Internet Pipes.
→ A previously unknown design vocabulary became a reusable creative system for producing consistent branded imagery.
Kieran supplied a model with a body of Shaan Puri's writing and asked why it performed well on its platform. The model surfaced elements of wit and differentiated articulation that Kieran could examine and selectively integrate into his own learning rather than copying the finished content.
→ Tacit observations about effective writing became explicit areas for deliberate practice.
Common mistakes
Copying instead of learning
Reproducing a finished style skips the analysis that develops judgment. Extract mechanisms, then decide which ones genuinely fit your own work.
Letting AI define your preferences
The creator must first observe and annotate personal reactions. Otherwise the model will tend to supply generic internet averages.
Changing too many variables
Uncontrolled experimentation makes it impossible to identify what improved the output. Hold the brief steady and vary a limited number of attributes.
Is it for you?
Best for
It is best for creators, marketers, and designers who want to understand and strengthen their distinctive judgment.
Not ideal for
It is not ideal for people seeking one-click imitation without reviewing outputs or developing their own point of view.
From the transcript
“what did I like about this image why did I like this response from chat gbt or Claud better than this one”
“you would just save pictures and annotate why they're good after a while you can then feed all that to an AI model and actually…”
“what I've been doing is creating style guides for anything and then within the style guide starting to tweak the variables and seeing by tweaking…”
From the episode
How I Use Loom + Ai To Automate Everything I Do - Steph Smith