How do I build a reusable AI creative workflow my team can run?
Write the job down once in FLORA: a prompt template with fixed style rules and named slots, the model and settings, and the review step. Save it as a FLORA skill or a file your agents read, and anyone on the team can run it by filling the slots. Our three-season bakery header set ran in one call for $0.165.
Key takeaways
- A reusable workflow is four things written down: a template, fixed settings, the inputs that change and a review step.
- Our template produced three seasonal 21:9 email headers in one FLORA call for $0.165, $0.055 each.
- The whole set was done in 39 seconds, because a batch runs every variant in parallel.
- flora_create_skill saves the instructions as a workspace skill that every teammate's agent can load by name.
- Fixing light, color and layout in the template kept the three headers looking like one campaign.
At a glance
| What it does | Turns a one-off creative job into a template anyone can rerun |
|---|---|
| Works in | Claude, Claude Code, Cursor, Codex and the FLORA canvas |
| Models | Any FLORA model; GPT Image 2.5 Flare in our run |
| Inputs | A prompt template plus the values that change per run |
| Outputs | A consistent set of images on a FLORA canvas |
| Typical cost | $0.055 per GPT Image 2.5 Flare image |
| Typical time | Under 40 seconds for a three-image set |
What you need
- The FLORA MCP in your team's agents. Setup for Claude Code is in the Claude Code install guide.
- One finished piece of work you want to repeat, such as an approved email header
- A shared FLORA workspace and project for the team's runs
Key features
- Workspace skills: flora_create_skill saves written instructions to the workspace, and flora_discover_skills finds them by name.
- FLORA's own skills: Ready-made skills such as flora-batch-generate and flora-refine-loop cover common loops.
- Shared project: Every run lands on the same canvas, so the team sees past results before making new ones.
- Prices up front: Each run can be quoted in dollars before it spends anything.
Best for
- Recurring campaign assets such as monthly or seasonal email headers
- Teams where several people make the same kind of image
- Handing a proven look from a designer to the rest of marketing
- Agencies running one format across many clients
Step by step
Step 1: Start from one result you approved
Pick a finished image the team liked. Write down what made it work: light, color, framing, where the empty space sits and what is never in frame. Those become the fixed part of the template.
Step 2: Split the prompt into fixed rules and slots
Everything that should stay the same is written once. Everything that changes per run gets a named slot. Keep slots few; each one is a place the result can drift.
Seasonal email header for a neighborhood bakery: a {bake} on a {surface}, {light}, muted natural color, true-to-life texture, the {bake} on the right third with empty space on the left for a headline, no text.Step 3: Fix the model and settings
Name the model id, aspect ratio, quality and resolution in the workflow so nobody has to choose. A teammate who changes the model has started a different workflow.
Model: t2i-gpt-image-2-5-flare. aspect_ratio 21:9, quality high, resolution 1k.Step 4: Save it where agents can find it
Save the template, settings and review checklist as a FLORA workspace skill with flora_create_skill, or as a markdown file in a shared repo. Give it a plain name, such as bakery-email-header, so a teammate can say "run bakery-email-header for winter".
Step 5: Run it and review the set
Fill the slots for each variant and send them in one batch. Review on the canvas against the checklist, and add anything that went wrong to the fixed rules so the next run avoids it.
Run bakery-email-header for spring (rhubarb tart, white marble counter, soft spring morning light), summer (peach galette, weathered wooden table, warm late summer afternoon light) and autumn (apple pie, dark slate board, low overcast autumn light).
What did a reusable creative workflow look like in FLORA?
On October 16, 2026, we ran the bakery header template above for three seasons from Claude Code. The agent filled the three slots for each season and sent all three in one flora_create_generations call with model t2i-gpt-image-2-5-flare, aspect_ratio 21:9, quality high and resolution 1k, then read the results with flora_list_generations.
The three runs took 39, 29 and 27 seconds and cost $0.055 each, $0.165 for the set. Each came back at 1520 x 656 pixels. All three put the bake on the right with open space on the left, and the light changed by season the way the slots asked: bright and cool for spring, warm and dappled for summer, grey and dim for autumn.
What went wrong: all three added props we did not ask for, a vase and rhubarb stalks in spring, loose peaches in summer, apples and fallen leaves in autumn. The autumn pie also sat beside the slate board instead of on it, cut off at the right edge. Both fixes went into the fixed rules: "nothing else in frame except the {bake}" and "the whole {bake} in frame, sitting on the {surface}". Each correction is written once, and every later run inherits it.
- Variants
- 3 seasons
- Model
- t2i-gpt-image-2-5-flare, 21:9, 1K
- Tools called
- flora_create_generations, flora_list_generations
- Cost
- $0.165 for three headers
- Time
- 27 to 39 seconds, run in parallel
| Option | What it holds | How a teammate runs it |
|---|---|---|
| Workspace skill | Written instructions, template and checklist | Asks their agent to run the skill by name |
| File in a repo | Template, settings and inputs as text or CSV | Points a coding agent at the file |
| FLORA canvas | The approved results and the prompts behind them | Branches a new run from an earlier one |
| FLORA technique | A saved multi-step visual workflow | Runs it from the canvas on a new input |
Workspace skill or repo file: pros and cons
Pros
- Workspace skill: lives inside FLORA, so any teammate's agent finds it with flora_discover_skills.
- Workspace skill: works in chat clients with no files, such as Claude on the web.
- Repo file: versioned with git, so changes to the rules are reviewed like code.
Cons
- Workspace skill: edits replace the instructions, so keep your own history of changes.
- Repo file: only agents with access to the repo can read it.
- Both: a vague template still drifts; the rules need to name light, color and layout.
What should a reusable workflow include?
Enough that someone who never saw the original can get the same result. It usually fits on one page.
- The prompt template, with slots named in plain words.
- The model id and every setting: aspect ratio, quality, resolution.
- One or two approved examples, as canvas links or image URLs.
- A review checklist: what a good result has and what to reject.
- The FLORA project the results go into.
How is this different from a saved prompt?
A saved prompt holds words. A workflow also holds the model, the settings, the inputs that change and the check at the end, which are the parts people get wrong when they reuse a prompt by copy and paste. Because agents read the workflow and call FLORA themselves, a teammate only supplies the slot values.
Which FLORA skills are worth building on?
FLORA publishes skills any agent can load with flora_discover_skills. Use them as the engine and keep your own workflow as the house rules on top.
- flora-batch-generate: one image per row of a list, in parallel.
- flora-refine-loop: generate, judge against a goal, adjust the prompt, stop.
- flora-image-resize: one approved image into other aspect ratios.
- flora-script-to-video: script to shots to clips, with an approval step between stages.
Frequently asked questions
Does every teammate need to know prompting?
No. The template holds the prompting. A teammate fills the slots in plain words, such as "winter, gingerbread, flour-dusted bench", and the agent does the rest.
What did the three-header run cost?
$0.165: three GPT Image 2.5 Flare images at $0.055 each, run in one batch.
Can the workflow include video?
Yes. Add a step that animates the approved still with a video model such as Seedance 2.5. Quote it first, since a 5-second 720p clip costs about $2.48.