Reusable Creative System
Turn recurring creative work into a reusable, controlled generation pipeline.
- Difficulty
- Advanced
- Time to result
- ~weeks to results
- Steps
- 8
- Confidence
- 98%
A Reusable Creative System converts a recurring asset-production task into a pipeline whose inputs, expert rules, transformations, model calls, and editing controls are explicitly connected. The builder begins with reference work and a narrow outcome, then uses analysis components to extract relevant visual or strategic properties. Those properties are combined with task-specific instructions and passed to suitable generation models. Conventional controls such as masking, blur, levels, and layout remain available when prompting cannot provide precision. After testing and refinement, the complex workflow is wrapped in a simplified interface that exposes only the inputs another user needs. This shifts creative effort from repeatedly making individual outputs to building a capability that can generate many consistent outputs across a team.
Origin
Extracted from Marketing Against The Grain through Lore's demonstrations of reusable angle generation, brand-style illustration, and YouTube thumbnail testing workflows.
Core principles
- 01Build processes rather than repeatedly prompting for individual assets.
- 02Preserve manual controls wherever precision matters.
- 03Encode brand knowledge and specialist judgment into the workflow.
- 04Design around a specific recurring use case instead of a universal solution.
- 05Expose complex systems through simple interfaces for non-specialists.
- 06Invest once so the workflow can serve many people repeatedly.
How to run it
- 1
Select a recurring production task
Choose an asset or transformation the team performs frequently enough to justify an upfront investment. Define the final production format and quality threshold.
Pro tip Start with a painful bottleneck that has clear examples of successful output.
Watch out Do not begin with a vague goal such as automating all creative work.
- 2
Capture references and constraints
Gather strong historical assets, brand guidelines, marketing briefs, test categories, and other specialist knowledge that determines what good looks like.
Pro tip Use several representative assets rather than relying on a single example.
Watch out Weak references will be reproduced and amplified by the system.
- 3
Decompose the workflow
Map the flow from left to right as inputs, analysis, transformations, generation, editing, and final outputs. Decide where deterministic tools should replace generative prompting.
Pro tip Give each stage one clear responsibility so it can be tested independently.
Watch out A single giant prompt hides failure points and makes improvement difficult.
- 4
Extract structured guidance
Use an LLM or specialist-authored component to describe the relevant properties of references in the form the downstream model needs. Encode expert criteria explicitly rather than assuming the model understands them.
Pro tip Write the analysis role and expected dimensions in detail.
Watch out Passing raw references directly to a generator may not preserve the properties that matter.
- 5
Generate controlled variants
Combine the structured guidance, user input, and task rules into detailed generation instructions. Produce multiple variants or compare models when output quality is probabilistic.
Pro tip Include purposeful variation categories rather than requesting arbitrary alternatives.
Watch out More outputs do not compensate for an incorrectly designed upstream process.
- 6
Restore precise editing control
Add conventional operations such as masks, layers, color correction, blur, or formatting wherever exact changes are needed. Keep these operations outside generative models when deterministic control is cheaper and more reliable.
Pro tip Use AI for interpretation and generation, but deterministic nodes for pixel-level requirements.
Watch out Prompting for tiny positional or visual corrections often triggers unwanted regeneration.
- 7
Productize and delegate
Hide the underlying node graph behind a simple interface containing only the inputs ordinary users should change. Make the validated system available to the wider team.
Pro tip Let specialists own the workflow while marketers or operators use the simplified design.
Watch out Exposing every internal control can recreate the complexity the system was intended to remove.
- 8
Measure and refine
Run the system on real work, identify the next limiting stage, and improve that component. Continue until a useful share of outputs reliably reaches production quality.
Pro tip Track the next concrete improvement rather than merely counting AI activity.
Watch out Expect an initial learning curve and some experiments that produce no useful result.
In the wild
A team supplies its strongest thumbnails, a marketing brief, and descriptions of test types such as visual hook, headline, pose, and gesture. An analysis component extracts the visual identity, a concatenation stage assembles detailed prompts, and an image model generates several purposeful test variants from the baseline thumbnail.
→ A recurring thumbnail-testing task becomes a reusable internal application that can produce multiple experiment candidates in minutes after roughly a week of initial workflow development.
A design team provides four illustrations with a distinctive style and constructs a system that accepts a short description of the desired scene. The complex graph is then converted into a simplified design interface so marketers can request new illustrations without operating the underlying workflow.
→ Non-designers can create timely illustrations while the outputs remain consistent with the specialist-defined brand style.
A source image is analyzed by an LLM, which proposes ten alternative views of the scene. Generation nodes turn those ideas into images, and replacing the source image allows the complete process to run again without rebuilding it.
→ One reusable machine generates batches of scene variations from different source images.
Common mistakes
Building a one-size-fits-all workflow
Broad systems rarely encode enough domain knowledge to reach the required production standard. Narrow the workflow to a repeated, specific use case.
Prompting away deterministic edits
Requests for exact layer, color, blur, or position changes can cause uncontrolled regeneration. Use conventional editing operations when the desired change is precise.
Expecting immediate efficiency
A reliable system requires learning, testing, and failed iterations before it produces leverage. Treat the initial slowdown as capability development rather than proof that the approach has failed.
Is it for you?
Best for
It is best for teams repeatedly producing similar assets such as thumbnails, campaign illustrations, product imagery, or video variants.
Not ideal for
It is not ideal for rare one-off tasks whose potential time savings cannot justify the workflow-building investment.
From the transcript
“The way we look at it is that the craft is shifting from kind of like making pixels into building process”
“And so the idea is that you're building your own toolbox and once you've built your own toolbox you can reuse it again again.”
“if you have something that you are doing on a very regular basis, you need to build a system for it.”
From the episode
Stop Prompting: Build an AI "Design App" Instead (Demo)