What does folio-visuals do?
Real photos and real screenshots always beat generated ones. Generate only what the student cannot shoot or screenshot. Never generate a fake version of their own face, a fake testimonial photo, or a company's real product with changed packaging presented as real.
Inputs
- portfolio.json (for the hero project ids) and the list of sections the site will have. PRD.md is written later by folio-prd.
- Reference images: their headshot, a real product pack, a style frame.
- The tool they have. Ask if unknown:
- ChatGPT images: free tier with a daily limit. Good for illustration, covers, icons, edits from an uploaded reference. It may not return the exact shape you ask for; ask for landscape and crop to the ratio in Canva.
- Gemini (image): free tier. Good at editing a photo while keeping the face and at consistent characters.
- Higgsfield: paid credits. Best for short video and cinematic shots; can be connected to Claude.
- FLUX on ComfyUI: free and local on an Apple Silicon Mac with 16 GB+. Unlimited tries, private.
- Canva / remove.bg: cut-outs, resizing, mockups. Free tiers change; tell them to check today's limit.
Step 1: Shot list
Make a table: file name | where it goes | type | must match | tool | real or generated.
File names follow assets/projects/<id>/cover.png, assets/me/.... Default needs: 1 hero image, one cover per hero project, 1 og-image (1200x630, saved as assets/og-image.png, with the name typed in Canva rather than generated) for link previews, optional icons and a 5 to 8 second cover loop. Put the list in priority order. On a free tier, say how many days it will take at today's limit.
Step 2: One prompt per shot, in nine slots (Motion is for video only)
Subject: who or what, specific (age, clothing, object, material)
Action: what is happening
Setting: where, with 2 concrete details
Light: direction and quality (soft window light from the left)
Lens: focal length, angle, depth (50mm, eye level, shallow)
Style: editorial photo / flat vector / 3D clay / risograph; palette in hex from refs/cards.md or design-system.md (if neither exists, propose two base colours and one accent and get a yes before writing prompts)
Frame: aspect ratio and where the subject sits (16:9, subject on right third, space for text on left)
Motion: video only: camera move, duration, loop (slow push-in, 6s, seamless loop)
Avoid: text in the image, extra fingers, logos, watermarks, plastic skin
Write the final prompt as one paragraph made from the slots. For ComfyUI or FLUX, keep it descriptive prose; for ChatGPT and Gemini, add "Use the attached image as the reference for ..." when a reference exists.
Recipes
- No headshot yet: do not generate one. Tell the student: phone at eye level, plain wall, face a window, chest up, take ten, pick one, save as assets/me/headshot.jpg. On a Classic page this photo is the hero image.
- Your headshot cleaned up: upload the real photo. "Keep the face, hair and expression exactly. Change only the background to and even out the light." Never generate a new face.
- Character or mascot sheet: "Character turnaround sheet: front, 3/4, side and back views of the same character, same outfit and proportions, plain background, each view labelled." Save it, then always attach it as the reference for every later shot of that character.
- Product shot: attach the real pack photo. "Keep the pack exactly as in the reference, including all text. Change only the scene: , , ." Any text that comes out wrong: fix it in code or in Canva, not by re-rolling.
- Project cover: a calm, abstract or object-led image in the site palette with space for the title; no fake UI with gibberish text. Better: a real screenshot in a device frame.
- Cover video loop: make the still first and approve it, then image-to-video: "Animate this image: , 6 seconds, seamless loop, no new objects." Export MP4 (H.264) under 4 MB and a poster image.
- Icons: "Set of 6 line icons, 2px stroke, rounded caps, on transparent, consistent grid: ."
Step 3: Consistency
Pin a style line (palette in hex, medium, light) and reuse it word for word in every prompt. Reuse the same reference image for any recurring character or product. Keep one aspect ratio per slot.
Step 4: Check before use
For each output: does the text read correctly, are hands and faces natural, does it match the palette, is it under 500 KB (compress at squoosh.app; if the file type changes, for example .png to .webp, update the path in portfolio.json to match), does it look like the same set as the others? List what to fix by hand.
Output
The shot list, one prompt per shot, the style line, and the check list. If the student has a connected image tool (Higgsfield or a ComfyUI MCP in Claude Code), offer to run the prompts; otherwise they paste them manually. End with: "Next: run folio-prd."