Upload one selfie. Get back a designed report card that scores your face across ten features, prints every score right onto your real photo, and hands you a ranked list of fixes you can actually act on. One prompt, about two minutes, and it runs entirely inside Claude Code.
The reason this lands is that it is honest and specific. Not "you look great," but "symmetry 7.8, skin 6.6, ceiling score 8.4, and here are the five changes that move the number, ranked by impact." It reads like a sharp friend with good taste and a good camera.
The stack: Claude Code (the brain, and the part that reads files and runs the build) + Higgsfield MCP (clean background removal so the report looks premium) + Python and Pillow (Claude installs and runs these to paint the scores onto your photo and typeset the report). No design skills, no Photoshop.
What you get
- An annotated face map. Your real photo, background removed and color corrected, with all ten zones scored right on the image: hairline, eyes, skin, cheeks, nose, lips, jaw, chin, symmetry, and overall harmony.
- A polished report card. A clean editorial page with your overall score, your ceiling score, a ten category breakdown with progress bars, and a Top 5 Improvements list ranked by impact. It saves as a PDF you can keep or share.
Why this runs in Claude Code, not the browser app
⚠️ This one needs Claude Code. The build saves your photo to a file, runs Python to composite the scores onto it, and writes a real HTML report you open and export to PDF. The browser app at claude.ai cannot touch files on your computer or run code, so it cannot produce the annotated image or the report. Install Claude Code and do the whole thing there.
Step 0 — the 5-minute setup
/mcp inside Claude Code and authenticate in the browser window that opens.claude mcp add --transport http higgsfield https://mcp.higgsfield.ai/mcpProjects/face-rating, and start Claude Code inside it.Step 1 — take a selfie worth scoring
The input photo sets the ceiling. A clean, front facing shot in soft light scores the real you. A dim, angled, filtered one just scores the photo.
✅ A good input photo is: front facing with eyes to the camera and head level, lit softly and evenly across the whole face (a window works), neutral or a natural smile with hair off the forehead, no filters or beauty mode or harsh shadows, shoulders up with your face filling most of the frame.
Step 2 — paste the prompt with your selfie attached
Attach your selfie in the message, then paste this exact prompt. It runs a strict pipeline: pull your photo to a file, remove the background with Higgsfield, paint the ten scores onto your real face, lock the scores, then build the report you can save as a PDF. It is tuned to stay honest and to keep your real face untouched, meaning background removal and exposure only, nothing that reshapes you.
FACIAL AESTHETIC REPORT — run this pipeline in strict order. Do not skip or merge stages.
INPUT: I have attached one photo of a face. That attached, background-removed photo is the
subject's REAL face — the source of truth for the whole report.
STEP 0 — GET THE PHOTO ONTO DISK:
Uploaded images are usually NOT saved as a file you can read directly. Extract the attached
image to a working file before doing anything else:
1. Find this session's transcript JSONL under ~/.claude/projects/<project-dir>/ (the newest
file, or grep it for a unique word from this prompt like "FACIAL AESTHETIC REPORT").
2. With Python, parse the JSONL, find the message "source" object of type "base64" with an
image media_type, base64-decode "data", and write it to your scratchpad as face_original.
3. Convert to PNG (sips or PIL). Confirm dimensions before continuing.
If you truly cannot recover it, ask me to save the photo to a path and stop — otherwise proceed
silently, no questions.
IDENTITY LOCK (non-negotiable, every stage):
The uploaded photo is the subject's REAL face — keep those exact pixels. Do NOT regenerate,
redraw, beautify, slim, smooth, re-age, or swap the face for a better-looking stranger. Preserve
bone structure, skin, features, expression, and proportions exactly. The ONLY edits allowed on
the face are: (1) background removed, (2) neutral exposure/white-balance correction if the shot
is over- or under-exposed. Nothing else about the face changes.
STAGE 1 — BACKGROUND REMOVAL (Higgsfield MCP):
Use the Higgsfield/creative MCP server. Flow: media_upload (get a presigned PUT URL) -> curl PUT
the image bytes (expect HTTP 200) -> media_confirm (get a media_id) -> remove_background with that
media_id and media_type "image" -> poll jobs_wait until completed -> download the result_url.
This yields a transparent RGBA cutout (face + shoulders, clean background). Verify the cutout is
real by compositing it over a bright solid color and viewing it — don't trust the raw preview.
Exposure/white-balance cleanup allowed; NO facial edits.
STAGE 2 — ANNOTATED FACE (build as a TRUE COMPOSITE, not a generative image):
Do NOT feed the face through a generative image model — those alter pixels and garble label text,
which violates the identity lock. Instead composite with Python + PIL (Pillow):
- Apply a GENTLE white-balance-only correction to the cutout: gray-world over opaque pixels
(mask = alpha>200), computed via PIL ImageStat (do NOT rely on numpy being installed), applied
at ~45% strength via per-channel Image.point scaling. Save a corrected transparent PNG and a
white-background JPG.
- Place the real corrected cutout on a clean off-white canvas with margins left and right.
Supersample ~2x for crisp text, downscale with LANCZOS at the end.
- Draw labeled callout boxes (rounded white rects, thin charcoal border, colored accent bar) with
thin connector lines + a small target dot on each facial zone. Each box: zone NAME + score X.X/10.
Color the score/accent by band: green >=7.5, amber 6.5-7.4, red <6.5.
- Zones (10): Hairline, Eye Area, Forehead Skin, Left Cheek, Nose Tip, Mouth/Lips, Jaw Angle,
Beard/Chin Grooming, Mid-Face Center Symmetry, Overall Harmony.
- Top-right: "Overall Attractiveness Score X.X/10" and "Ceiling Score X.X/10".
- Bottom: one honest 2-sentence summary (word-wrapped to the canvas width, no overflow).
Output a single annotated PNG. View it and fix any overprinted text or misaimed connectors before
moving on.
STAGE 3 — SCORE THE FACE (text only, lock BEFORE building the report):
Score each out of 10 honestly, no flattery, no cruelty — specific and constructive:
Symmetry; Hairline & Hair Style; Eye Shape, Color & Spacing; Nose Harmony; Cheekbone Structure;
Lip Proportions; Skin Texture & Tone; Jawline & Chin; Grooming Quality; Overall Facial Harmony.
Lock as text: overall X.X/10, ceiling score, a 2-3 sentence summary, a one-line note per category,
and the Top 5 Improvements ranked by impact (each: impact label HIGH/MEDIUM/LOW, title, 2-3 sentence
why, estimated point gain). These EXACT values get typeset into the report — the report must never
invent or change a number. Distinguish what's fixable (skin, lighting, grooming, styling, expression)
from fixed bone structure, and call out when the photo's lighting/expression is underselling the face.
STAGE 4 — COMPILE THE REPORT (single self-contained HTML file, PDF-ready):
Portrait editorial report card, embed the Stage 2 annotated image as a base64 data URI (one image
only — the annotated face, full-width in its own panel; do NOT show two side-by-side images).
- Masthead: "Attractiveness Potential Report" in an elegant serif (Playfair Display via Google
Fonts, with a serif fallback stack), small-caps subtitle, thin divider rule.
- Hero: LEFT = large overall score "X.X / 10" with an above/below-average label and the ceiling
score; RIGHT = the 2-3 sentence summary.
- Category Breakdown: one row per category — name, one-line note, a colored progress bar
(green >=7.5, amber 6.5-7.4, red <6.5, width = score/10), and the numeric score.
- Top 5 Improvements: numbered cards ranked by impact — impact badge, title, 2-3 sentence
description, estimated point gain.
- Footer: a one-line note that this is a subjective aesthetic read for personal use (directional,
not clinical) and that the photo is real, background-removed and white-balance-corrected only,
features unaltered.
STYLE: white background, charcoal text, thin hairline dividers, serif for the display title only,
clean sans-serif for body and scores, premium editorial feel, no gradients, no dark backgrounds.
Include @media print + @page portrait CSS so Cmd+P -> Save as PDF looks right.
OUTPUT: the Stage 2 annotated image, plus the final report as a single downloadable self-contained
HTML file I can view and export to PDF, with all Stage 3 scores and copy typeset exactly. Save the
deliverables to a clear folder and hand me the paths. Verify the HTML renders correctly in a browser
(all sections, bars, and cards) before delivering. Work autonomously — no questions unless the photo
genuinely can't be recovered.💡 It is built to protect your identity. The prompt hard locks your real pixels: it only removes the background and corrects exposure, and it never smooths, slims, or redraws your face. Every score is read off the real you, not a prettied up stranger.
Step 3 — read it right, then iterate
🧭 Use it for the patterns, not the decimals. This is an AI aesthetic read, not a clinical assessment. The scores drift a little run to run and the model leans kind. Take it the way you would take a blunt friend's opinion: the fix list is where the real value is.
Most of what moves a face is free: better light, hair off the face, matte skin, groomed brows, a genuine smile. Do those, shoot it again in a month, and watch the same few tips turn into a higher number. To push the report further, just talk to Claude in plain English: "make the scores stricter," "add a 30-day glow-up plan," "export the report portrait for stories."
Things that catch people
- You used the browser app. claude.ai cannot run this. Use Claude Code.
- Higgsfield is not connected. If the background removal step fails, run the
claude mcp addcommand again and confirm the connection with/mcpinside Claude Code. - The selfie is working against you. Orange indoor light and shine tank the skin score. Shoot in soft daylight before you blame your face.
- You took the number personally. It is directional. The ranked fixes are the point.
What it costs
About 10 to 20 cents in Higgsfield credits for the one background removal, plus your Claude subscription. That is the whole bill.
Pro move: turn it into a skill
When you are done, tell Claude: "Save everything we did here as a skill so future face ratings start with all of this baked in." It packages the pipeline, the identity lock, the scoring bands, and the report design into one reusable skill. Next time you just say "rate this face" and it already knows the whole build.
Built by Dylan Watkins. I build AI systems like this and film how. If you want this or an AI system like it set up for you or your brand, DM me RATE on Instagram @dylan_j_watkins.