Scaling Creative Stories Through Reusable Design Knowledge
For Magic Gifts, I organized designer examples into reusable knowledge and story constraints, helping the team generate stronger image candidates. Production improvements informed a later workflow tool.
On this page: 01 / 06The problem
01 · The problem
A beautiful image still needs a story
Creative designers had already made strong Magic Gifts through individual exploration and refinement. Their work established a quality reference, but finding each good direction took time.
When we tried to expand through batch generation, many images looked polished yet felt uneventful. A character and a setting were present, but there was little interaction or story. The relationships that made the designers’ work compelling were not consistently reaching the next generation attempt.
The design question
How could the next batch start from what we had already learned, while still leaving room for new stories?
02 · Reusable knowledge
Give the LLM a story skeleton
I collected strong work and prompts into a reusable library, used AI to help identify recurring patterns, and refined the system prompts and templates. The key was to connect the creative conditions.
Personality informs expression, clothing and behavior. A setting makes some activities plausible. An interaction gives those choices a reason to appear together. These relationships give the LLM a story skeleton, with room to develop the details.
- 01
Curate examples
Creative designers’ strong work provides a shared quality reference.
- 02
Find relationships
Extract how character, setting, action and interaction support each other.
- 03
Shape the next input
Use those relationships in prompts and templates, then review what they produce.
03 · Story example
A cool cat, a playful dolphin, one event
Consider a cool, technology-inspired cat at the seaside. Naming the character, place and style leaves the event unresolved. Here is how connected choices can give the scene a story.
Explanatory example
Select a decision to trace it in the story.
A serious tech cat in swim shorts and sunglasses surfs beside a mischievous dolphin, keeping its cool as the dolphin plays.
Too little structure leaves character and story to chance. Too much can flatten variation and liveliness. I adjusted the conditions by looking at the resulting candidates with the team.
04 · Generate & select
Spend human attention on the choices
For this task, we found that generating a batch and reviewing the candidates was faster than repeatedly editing individual images. We used the improved prompts for batch image generation, then selected with creative designers and the team.
Accepted images needed little or no manual editing before image-to-video generation. People’s time went into judging which scenes were coherent, interesting and worth taking forward.
- 01
Prepare
Examples + story conditions
- 02
Generate
A batch of still images
- 03
Select
Human creative review
- 04
Continue
Image-to-video + further review
Does this image work?
Is the scene coherent? Do character, outfit and action belong together? Is there an interaction or event worth developing? Can the image move into video with little or no editing?
Is this batch worth choosing from?
Are there several strong, distinct directions? Does the team see stories it wants to explore, with enough variety and interest to make selection worthwhile?
These summarize our review concerns. They were not a historical weighted scorecard; creative designers and the team made the final choices.
When candidates miss the brief
- Little story
- Revisit the event and the relationship between characters before generating more variations.
- Incoherent action
- Check whether the action fits the character, setting and intended event.
- Weak composition
- Look at the focal point and whether the important interaction is readable.
- Character drift
- Compare identity-defining features with the chosen references.
Diagnostic principles, not a record of controlled before-and-after experiments.
05 · Results & roles
More images worth taking forward
After the library and prompt improvements during production, we had more candidates the team could select for image-to-video generation.
Image acceptance rate · retrospective estimate
At least 50% relative improvement
Success meant a human-approved still image that could enter image-to-video generation with little or no manual editing.
This is my estimate from the production experience. Matched batch counts and a same-condition baseline have not been verified for this case study. It is a relative change, not a 50-percentage-point gain, and does not measure video acceptance or final release rate.
The improvement was a team outcome. Without a controlled comparison, I cannot isolate the contribution of the library and prompts from model changes, accumulated creative experience or other production conditions.
- My contribution
- Organized reusable knowledge, shaped prompt conditions, iterated from generation results and developed the later workflow tools with AI-assisted coding.
- Creative designers & reviewers
- Established creative directions and strong examples, contributed aesthetic judgment, and selected candidates together with the team.
- Technical collaborators
- Contributed to implementation and maintenance of the production chain. Final delivery depended on the team’s subsequent generation and review.
06 · After the retrospective
Turning experience into tools
After the production retrospective, I developed tools to bring experience reuse, batch generation, selection and post-processing together. This was a later phase: almost none of that production phase’s released work directly used the newly assembled workflow. Some subsequent automation adopted its ideas.
Import selected prompts
Keep the scene, character, style and source alongside a reusable prompt.
Allocate a batch
Set a target mix of existing selected prompts and newly composed tasks.
Check explicit constraints
Check prompt rules and record repairs or fallbacks.
Keep selections traceable
Save chosen images with their prompts and metadata for later processing.
The source mix allocates prompt tasks; it is not a percentage of AI creativity or autonomy. Rule checks inspect text constraints. Image quality and story still require human review.
Creative designers establish valuable directions through exploration. My work is to organize what can be reused so the next round starts with more direction. The test is whether the team gets worthwhile choices and can do better creative work with them.
The Lab contains separate interface experiments with simulated data. It does not reproduce the production workflow or the later tools.
Explore the Lab