flora-mcp-skills: Skill for Claude Code

flora-pdp-deck

Self-contained. The deck builder is at the foot of this file; write it out and run it.

What it is

Not an image generator. A consistency engine. The product exists; the job is putting it in front of a camera many times without it drifting between frames.

The product is reference-locked. The world and the pose are generated. A seam moving, a strap ending somewhere new, a colour drifting — that is a fail.

What you can get from what you have

Start here. What the user hands you decides what you can honestly promise, and the limits are not obvious.

one product photo        1 clean plate + N on-model or context frames + a deck
  + a model frame        the same, with a real person holding or wearing it
several real photographs one plate per photograph, plus the above

A single reference cannot be rotated by asking. The SHOT: line changes framing; it does not turn the product. Measured — four plates from one photo, byte-identical prompt block, only the shot line varying:

mean absolute grey difference, 0 = identical
  three-quarter  vs  rear 180 deg      3.21    <- opposite faces requested
  front          vs  three-quarter     6.39
  each plate     vs  the reference     8.06 - 8.74

The plates differed from each other less than each differed from its own source. Four requested angles, four near-duplicates. On an asymmetric product it shows directly: a wax drip specified on one side stayed on that side in the frame asked to show the opposite face.

So from one photo: generate one plate, not four. Then spend the budget on on-model and context frames, where the composition genuinely changes and the model has real work to do. Say this to the user before firing. Four billed duplicates is the expensive way to learn it.

Intake — the image is in the chat, not in FLORA

The user drops a photo into the conversation. It is not in FLORA. Nothing generates until you put it there, and this is the step that stalls a run.