Fix plastic skin and fake-looking AI portraits: prompt rules, edit locks, and a linter.

Quickstart · How it works · Examples · Install · FAQ

English · Русский · 简体中文 · Español · Português (BR) · 日本語 · 한국어

no-plastic: fix plastic skin and fake-looking AI portraits

AI portraits look fake for a short list of reasons: poreless skin, a mirror-symmetric face, light from nowhere, bokeh painted on, teal-and-orange grade, no grain. Each one is the model's default, and each has a sentence that overrides it. no-plastic is a Claude Code skill (with Cursor and Codex adapters) that rewrites your image prompt into a photographic brief with those sentences in it, writes edit prompts that change one thing and keep the person, and ships a zero-dependency linter that scores any prompt against the checklist. Written for Flux, Nano Banana, Midjourney and SDXL; the vocabulary is photography, so it carries to the next model too.

Brief, not tags	Edits that keep the face	Lint before you render
Subject, framing, camera, light, palette, texture, era, negatives: eight fields, three to five sentences, the anti-plastic rules folded in.	One change plus four locks (identity, realism, framing, seams), with templates for remove, background, outfit, hair, enhance and selfie-to-headshot.	scripts/lint_prompt.py scores coverage, flags the words that cause the plastic look, and prints a brief skeleton for the gaps. Python 3 stdlib only.

Quickstart

git clone https://github.com/uniqaiapp/no-plastic ~/.claude/skills/no-plastic
cd ~/.claude/skills/no-plastic
python3 scripts/lint_prompt.py examples/01-linkedin-headshot/before.txt examples/01-linkedin-headshot/after.txt

You should see the first prompt ("professional headshot of a beautiful woman, perfect skin, 8k...") score 0/100 with the words that sink it listed, and the rewritten brief score 100/100. Then in Claude Code type /no-plastic and paste a prompt, or just say "this looks plastic" with the prompt attached.

Demo

How it works

1. Classify. A generate prompt (text to image) or an edit prompt (an instruction on an existing photo: remove, change outfit, enhance, selfie to headshot). The linter auto-detects by the opening verb.
2. Lint. python3 scripts/lint_prompt.py --brief prompt.txt prints the eight-field brief with the phrases it found and ? where the prompt is silent; the plain report lists the flagged words (8k, masterpiece, flawless skin, cinematic lighting, tag soup, the SD1.5 negative ritual) and what each costs.
3. Critique and rewrite. The skill names the three biggest gaps in photographic terms and rewrites the prompt as a brief: subject and action; framing; camera and light; palette, texture and era; negatives. The anti-plastic rules are always in: natural skin texture with visible pores, slight asymmetry, a named light source with falloff, fine film grain, not over-processed, no retouched plastic look.
4. Adapt per model. Flux and Nano Banana get prose with the avoidances phrased positively (they ignore negative lists); Midjourney gets --style raw, a lower --stylize and a short --no; SDXL gets the brief with subject, light and texture in the first 77 tokens and a short negative field. See references/model-notes.md.
5. Pilot, then batch. Render one image, run the five-second review (forehead texture at 100 %, nose shadow, catchlights, hairline, background falloff), fix the brief rather than re-rolling, then render the set. This is the workflow UniqAI used to produce the illustrations on its own site; references/pilot-review.md has the job-list format and the decision rule.

loose request ──▶ lint (gaps + flags) ──▶ brief / edit with locks ──▶ per-model variant ──▶ re-lint ≥ 85 ──▶ pilot ──▶ review ──▶ batch

The scores in this README are produced by python3 scripts/lint_prompt.py on the files in examples/; tests/test_examples.py fails if they drift. The linter reads words, not pixels: a 100 is a complete brief, not a guaranteed image. That is what the pilot is for.

Examples

Nine pairs with reports in examples/. Two of them in full:

Before (examples/01-linkedin-headshot/before.txt, score 0/100)

Professional headshot of a beautiful woman, perfect skin, studio lighting, 8k, ultra realistic, masterpiece, sharp focus

After (examples/01-linkedin-headshot/after.txt, score 100/100)

This is the beginning of the README. Read the rest at the original. Read the full README ↗