Tonebook  /  Color Guide

Color Analysis With ChatGPT: The Exact Prompt (and Why It's Unreliable)

Reviewed by the Tonebook color team · Updated June 2026

Quick answer

ChatGPT can do a rough color analysis — paste the prompt below with a daylight selfie. The warm/cool undertone signal is often plausible; the specific 12-season assignment is inconsistent: the same photo can yield Light Spring one run and True Summer the next. For a stable result, use a purpose-built model like Tonebook.

The copy-paste ChatGPT color-analysis prompt

This prompt is structured to extract all three axes that define a season: undertone (warm, cool, neutral, or olive), value (light, medium, or deep), and chroma (clear/bright or soft/muted). Paste it verbatim after attaching your photo:

Prompt to copy: "You are a Sci·ART-trained personal color analyst. Look at this photo and tell me: (1) my skin undertone — warm, cool, neutral, or olive; (2) my value — light, medium, or deep; (3) my chroma — bright/clear or soft/muted; (4) your best-guess season from the 12-season system (Bright/Light/True Spring, Light/True/Soft Summer, Soft/True/Deep Autumn, Deep/True/Bright Winter), with a runner-up season and brief reason for both."

You need a vision-capable version of ChatGPT (GPT-4o or later; the free text-only tier won't read photos). The prompt is intentionally structured: asking for all three axes first forces the model to reason step by step before committing to a season, which improves output quality somewhat — though not consistently.

Step-by-step: how to feed ChatGPT your photo and hex colors

  1. Take a daylight selfie. Stand near a window in indirect natural light. Avoid warm bulbs, ring lights, and filters. Hair back, face and neck fully visible.
  2. Open ChatGPT (vision-capable model). Click the image icon and attach your selfie. Do not use a filtered or heavily edited photo — lighting corrections in the photo corrupt the undertone signal.
  3. Paste the prompt above. Send it as a single message with the photo attached.
  4. Add your hex colors for a richer palette check (optional). After the first reply, follow up: "My vein color hex is approximately #7B9E9E and my eye color is approximately #8B7355. Does that change any of your three-axis readings?" Piping in measured hex values gives the model a harder data point to anchor on.
  5. Run it twice. Ask the same question a second time with a fresh session. If the season changes between runs, treat both outputs as unreliable and verify with a dedicated tool.

What ChatGPT gets right: general undertone direction

The warm/cool split is where ChatGPT performs best. A photo with clear golden skin tones will usually yield "warm undertone" across most runs; a photo with obvious pink or rosy tones will typically land on "cool." This is because the underlying vision model has seen enough fashion and beauty content to associate warm skin with gold-leaning hues and cool skin with pink-leaning ones.

The value axis (light vs deep) is also usually accurate — it is visually obvious even to a general model. These two observations together make ChatGPT a reasonable free triage tool: if you're completely unsure whether you're warm or cool, a photo-prompt session can help you form a working hypothesis before doing more rigorous testing.

Where ChatGPT fails: inconsistency, no 12-season grounding, lighting blindness

ChatGPT has no stable 12-season classification logic. The Sci·ART system (derived from Carole Jackson's Color Me Beautiful in the 1980s and refined by Kathryn Kalisz's seasonal classifications) requires reading three axes simultaneously and mapping them to named seasons with consistent decision boundaries. A general language model generates text probabilistically — the same photo, the same prompt, on two separate runs can produce different seasons.

Failure modeWhy it happensImpact
Season instabilityProbabilistic sampling; no fixed classification layerHigh — season can flip between runs
Lighting blindnessCannot correct for warm/cool ambient light in photoMedium — warm bulb pushes warm reading
No runner-up logicNo trained proximity model between adjacent seasonsMedium — misses borderline cases entirely
Chroma axis is weakBright vs muted is subtle; general models conflate it with saturationHigh — chroma determines Bright vs True vs Soft within a season
Deep/dark skin biasFewer training examples for Fitzpatrick V–VI in fashion datasetsHigh for deeper complexions

The lighting problem is the most insidious. If you took your selfie under a warm incandescent bulb, ChatGPT will often read "warm undertone" even for a Cool Winter. A purpose-built model normalizes for ambient color temperature before making any undertone call. ChatGPT has no access to EXIF data and no color-correction step.

ChatGPT vs a purpose-built model: a consistency framing

The meaningful comparison isn't accuracy on a single run — it's repeatability. A well-designed color-analysis model applies the same measurement pipeline to every photo: sample skin pixels, correct for ambient light, compute the undertone axis, compute value and chroma, then map all three to the 12-season grid. The output for the same photo is deterministic.

ChatGPT

Tonebook (dedicated model)

Why a dedicated 12-season AI beats a general LLM

The 12 Sci·ART seasons — Bright, Light, and True Spring; Light, True, and Soft Summer; Soft, True, and Deep Autumn; Deep, True, and Bright Winter — each occupy a specific region of the undertone/value/chroma space. Reliably placing a person in that space requires a model trained to measure those three axes from pixel data, not one trained to generate plausible-sounding text about beauty topics.

Tonebook's approach: read a single unfiltered selfie, normalize the color temperature of the ambient light, measure undertone (warm, cool, neutral, olive), value (the depth axis), and chroma (the clarity axis), then return a primary season and a runner-up with a confidence delta. If you're borderline between Soft Summer and True Summer, you'll see that uncertainty named — not a confident wrong answer.

How Tonebook helps

If you've already run the ChatGPT prompt and have a working hypothesis, Tonebook is the fastest way to validate or correct it. Upload one selfie — no filters, daylight preferred — and the model returns your season, your runner-up, and a palette in your specific colors. The first full analysis is free and takes under a minute. It works across all skin tones, including deeper complexions where general models have historically struggled.

Get a consistent 12-season result

Tonebook reads one selfie, corrects for lighting, and places you in the Sci·ART 12-season system with a runner-up and confidence score. Inclusive across Fitzpatrick I–VI. First analysis free.

Get Tonebook for iPhone

Common questions

Is the free ChatGPT color analysis method accurate?

ChatGPT can offer a rough warm/cool direction, but it cannot consistently output a specific season from the 12-season Sci·ART system. Results vary significantly with photo quality, lighting, and how the prompt is worded. It is useful as a starting-point sanity check, not a definitive analysis.

Can ChatGPT tell me my exact color season?

Not reliably. ChatGPT is a general language model with no stable undertone-measurement logic. Ask the same photo twice and you may get Light Spring one time and True Summer the next. A purpose-built model trained on 12-season color theory will give consistent, repeatable results.

Why does ChatGPT give different answers each time?

Large language models sample from a probability distribution at every response, so identical inputs produce different outputs. Color season assignment requires deterministic pixel measurement, not probabilistic text generation — which is why a dedicated vision model outperforms a chat interface.

What is the best AI tool for color analysis?

Tonebook is purpose-built for 12-season color analysis: it reads one selfie, corrects for ambient light, measures your undertone (warm, cool, neutral, or olive), and places you in the Sci·ART system with a runner-up season and confidence delta. The first analysis is free.

Does ChatGPT support the 12-season system?

ChatGPT has read about the 12-season Sci·ART system (descended from Carole Jackson's Color Me Beautiful) and can describe it, but it has no trained visual classification layer for it. Descriptions and actual season assignments are not the same thing.