flora-mcp-skills: Skill for Claude Code

flora-brand-ad-pipeline

Attribution. Pass skill: "flora-brand-ad-pipeline" on every FLORA call you make while running this skill — execute included — along with a skill_run_id you invent once when the run starts and reuse for the rest of it. Both are reporting only: they change nothing about the call or its result.

What it is

Not a creative assistant. A conveyor. A product line goes in one end; paused ads, an archived master and a hosted CDN copy come out the other, with one record tying every ad back to the generation that made it.

FLORA is one stage of four. Most of what makes this reliable is not the generation — it is refusing to advance a stage that has not been proven, and keeping enough state on disk that a re-run repairs the pipeline instead of duplicating it.

The law

The brand is an input, not an instruction.

Fixed colours, a price that must always be visible, product reference shots: these decay the moment they are carried as prose an agent retypes each run. Agents paraphrase. A paraphrased brand drifts, and it drifts silently across every item in a batch — nobody sees it until a human opens the ads.

So the brand never lives in this file, and never lives in a prompt you compose:

• The treatment lives in a saved FLORA technique. A technique's steps, prompts and model choices are fixed by its author, so it gives the same treatment every time and the agent cannot rewrite the middle of it. Whatever can move out of prose and into a technique, move.
• The product lives in a reference image. Reference-locked, never re-described.
• The values live in brand.json. Colours, price format, placement sizes, product reference ids. See reference/brand.md. Read it; do not inline it.
• The price is composited, not generated. No image model renders a price string reliably, and "price always visible" is a compliance requirement, not a style note. Overlay it deterministically and the question becomes a boolean.