Claude skill taps ProductAI to make AI portraits look less plastic

Why it matters

Shape is using Claude Skills as a distribution layer for ProductAI, turning a narrow creative workflow into a credit-consuming path back to its product-photo studio.

A downloadable Claude skill puts ProductAI behind a workflow that rewrites glossy AI portraits and product renders into rougher, camera-like images while trying to keep product packaging intact.

The Vercel-hosted landing page describes the tool as "realistic-shot.skill" for Claude and ProductAI. It is free to download, but it requires Claude, a ProductAI account and the ProductAI MCP connector. Generation still consumes ProductAI credits at a listed rate of 1 credit per variant. The page does not name a creator or company behind the skill.

The product claim is narrow, and that is the point

The skill's argument is that AI image tools have made a recognizable beauty pass too easy: smooth skin, warm glow, painted color and studio-specular lighting. The page calls the target "plastic" AI portrait generation and says the skill turns photos or renders into files with pores, sensor grain, real light and unretouched texture.

Under the hood, the page says the workflow has 4 steps. A user drops in an image, including a render, model shot, product photo or layered PSD. The skill flattens and classifies it. Claude then writes an "evidence-first prompt" that names the synthetic artifacts to remove and the physical details to add. The image-to-image generation runs on "nanobananapro" through the user's ProductAI connector. The result is then checked at 100% zoom, with skin texture inspected and product labels compared against the reference.

That last step is the commercial center of the product. For brands, a generated product shot is useless if the bottle label mutates, the logo slides, or the packaging typography becomes plausible nonsense. The skill page states a "product fidelity law": imperfections go on the scene and never on the label. It says packaging shape, label layout, typography, logo placement and brand colors are supposed to stay locked to the reference, with a prompt suffix: "no changes to the product design whatsoever."

That is still a claim about a workflow, not independent evidence that the system reliably enforces label fidelity. The page does not disclose a benchmark, automated OCR test, human review protocol, failure rate or side-by-side dataset. It says labels are verified "character-for-character," but does not explain whether that check is automated, performed by Claude's visual inspection, or left to the user as part of the skill's instructions.

Why this runs through Claude

Claude Skills package repeatable instructions and supporting files so Claude can apply them when a task calls for them. That format is useful for a company like ProductAI because the hard sell is not another prompt box. The sell is a workflow: upload image, describe the physical evidence, call a connected image system, inspect the result, iterate by clause.

The install instructions make that strategy explicit. Users download the .skill file, save it in Claude, connect ProductAI's MCP connector, upload an image and ask Claude to "make this a realistic shot." The page says subsequent edits can be expressed as small clauses such as "more grain," "less frizz" or "cooler light," with the skill changing that part of the prompt and rerunning the generation.

For ProductAI, Claude becomes the operator layer. ProductAI still supplies the account, credits and image generation path. Claude supplies the conversational interface and the repeatable creative procedure. The skill acts as a small distribution surface for ProductAI, letting users try a specific use case without first learning ProductAI's broader studio.

ProductAI's own site presents the core product as an AI studio for e-commerce and marketing visuals, with adaptive templates, short video creation, background replacement, object removal, upscaling and a Simple API. ProductAI says users can start with 20 free credits, then pay $8 per month for Basic, $16 per month for Standard or $49 per month for Pro. ProductAI also claims "30.000 registred users," 830+ reviews across Trustpilot, Product Hunt and G2, and trust from 75+ brands worldwide, with logos shown for companies including Google, Adobe, Adidas, Bosch, Orange and Vinted. Those user, review and brand numbers are ProductAI's claims and are not timestamped on the page.

A demo aimed at the obvious failure mode

Most AI product-photo tools compete on speed, price and style range. This skill competes on a more specific failure mode: generated commercial images often look polished in the thumbnail and wrong at the point of sale. Human skin turns waxy. Hair becomes too clean. Fabric lacks pilling and fiber. Product packaging drifts just enough to make a shot unusable.

The realism list on the skill page is practical rather than cinematic. It asks for vellus hair, under-eye texture, dry lip patches, flyaways, frizz, baby hairs, sensor grain, shadow noise, imperfect white balance, wool fibers, fingerprints, dust, micro-scratches and contact shadows. That catalog reads like a production checklist written by people who have watched AI images fail under crop and zoom.

That checklist fits the paying market for ProductAI. The target customer is not someone making a single pretty portrait. It is brands and operators that need many assets across SKUs, channels and campaigns. ProductAI's site says its API supports production workflows, bulk generation, webhooks, asset library management and custom templates for larger teams. In that context, preserving a label is a revenue feature. So is making a generated model shot look less like it came from the same default image-model preset as everyone else's ad.

The open question is how much of this is enforcement and how much is prompt discipline. The skill does not disclose whether ProductAI has a dedicated label-protection mechanism, whether "nanobananapro" is ProductAI's own model route or an external model wrapped by ProductAI, or how the zoom check is implemented. It also does not disclose funding, valuation, revenue, headcount or founder ownership for ProductAI.

The approach is instructive. The AI image market is crowded with generators that promise better pictures. This project packages taste, error-checking and product constraints as a Claude skill, then routes the expensive part through ProductAI credits. If that workflow holds up outside the demo, it gives ProductAI a sharper wedge than generic product photography: make commercial AI images look less synthetic without breaking the thing the customer is actually trying to sell.