Modular Prompting Batch Engine
Combine reusable prompt modules, split batches, and generate assets in parallel.
- Difficulty
- Advanced
- Time to result
- ~weeks to results
- Steps
- 7
- Confidence
- 97%
The Modular Prompting Batch Engine converts a successful creative recipe into a reusable workflow. Inputs such as model attributes, clothing references, shot type, brand rules, and a user request are kept as separate modules. A concatenation step combines the relevant text and images into a master generation request. An LLM then creates multiple short prompts in a prescribed format, placing a delimiter such as a star between them. An array operation splits the response at each delimiter, and list selectors route the resulting prompts to image generators simultaneously. Stable system instructions govern formatting and recurring behavior, while variable inputs control each job. Once the base workflow works, it can be duplicated and adapted for camera angles, action poses, close-ups, advertisements, or other variations, turning one controlled asset into a scalable creative library.
Origin
Rory Flynn built the workflow to address the cost and delay of photographing thousands of fast-fashion SKUs. An early virtual photo-shoot system became the foundation for broader batch-generation engines used for product and advertising assets.
Core principles
- 01Build reusable modules instead of isolated prompts.
- 02Separate stable rules from changing user inputs.
- 03Use delimiters to convert one model response into an array.
- 04Run independent prompts simultaneously.
- 05Treat one validated asset as the seed for many variations.
How to run it
- 1
Define the batch job
Choose the base asset and the dimensions that should vary, such as camera angle, pose, crop, or advertising concept. Specify which product details must remain fixed.
Pro tip Start with one narrow, commercially useful batch rather than every possible variation.
Watch out Unbounded variation creates more review work than useful output.
- 2
Build modular inputs
Create separate controls for the subject, product, environment, shot type, perspective, and other variables. Include source images when exact visual context matters.
Pro tip Model the interface like a character creator with explicit options.
Watch out Do not combine all controls into one opaque text field.
- 3
Add stable instructions
Write a system prompt that defines recurring tasks such as describing uploaded clothing, formatting prompts, or preserving visual requirements. Keep job-specific requests outside it.
Pro tip Reuse stable instructions across every run and improve them from observed failures.
Watch out A weak system prompt multiplies errors across the entire batch.
- 4
Concatenate the context
Combine the selected modules, image descriptions, and reference assets into the master input for generation. Ensure both textual and visual context reach the model when needed.
Pro tip Keep the concatenation order consistent across runs.
Watch out Missing product context can cause logos, garments, or other details to drift.
- 5
Generate delimited prompts
Ask the LLM for a fixed number of concise prompts and require a unique delimiter between each one. Define constraints such as full-body framing or a maximum word count.
Pro tip Choose a delimiter that is unlikely to occur naturally in the prompt text.
Watch out Inconsistent formatting can break downstream splitting.
- 6
Split and route the array
Split the LLM output at each delimiter so every prompt becomes a separate array item. Route those items to image-generation nodes for simultaneous execution.
Pro tip Test the split with a small batch before increasing volume.
Watch out Do not scale until every array item maps cleanly to one generation request.
- 7
Curate and extend
Review the outputs, retain useful assets, and duplicate the engine for new variation families. Improve modules and system instructions based on recurring defects.
Pro tip Treat curation quality as part of the system rather than an afterthought.
Watch out Generating hundreds of assets without a review process merely relocates the bottleneck.
In the wild
Rory's workflow creates an artificial model, combines that model with real clothing references, and produces six prompts for front, side, three-quarter, and other full-body views. Stars separate the generated prompts, an array operation splits them, and the workflow runs the camera angles simultaneously. Duplicated branches then request action poses, closer crops, and lower-half shots.
→ One clothing upload produces a varied virtual photo shoot without repeatedly writing or running individual prompts.
A product image, user request, formatting instructions, and visual brand profile are combined into one workflow. The LLM creates ten precisely formatted ad prompts, splits them, and sends them directly to multiple generators. The same engine can be expanded to dozens or hundreds of concepts.
→ The business rapidly creates a large, brand-directed library for creative testing.
Common mistakes
Scaling before validating one output
A batch engine magnifies defects, so the base prompt and product representation must work before parallel generation begins.
Mixing rules with requests
Putting stable formatting rules and temporary creative requests into one prompt makes the workflow difficult to reuse and maintain.
Ignoring the curation bottleneck
Mass generation creates a new need to rank, review, and select outputs; volume alone is not a finished creative process.
Is it for you?
Best for
It is best for teams producing many variations of product photography, advertisements, campaign concepts, or other structured creative assets.
Not ideal for
It is not ideal when every output requires bespoke art direction or when the underlying generator cannot maintain critical product details.
From the transcript
“this is really just modular prompting.”
“every time you see a star in here that means you know separate new prompt.”
“each individual one is a prompt and then we can go run them simultaneously.”
From the episode
AI Tools to Replace Your $10k+ Creative Agency