Cindy Zhu.
Free guides
Design & visuals Claude

Turn phone pics into studio product shots

Hey, it's Cindy ๐ŸŒฑ You commented STUDIOS, so here is the whole setup from the reel: the product photography skill you install in Claude. Phone pics in, store-ready images out.
btw the waitlist for AI System Club is now open! If you're keen on using AI to automate your content, video editing, job search, business or build AI agents, feel free to sign up (only if you want to hehe). Ps. the waitlist gets first dibs:)
๐Ÿ’ก The division of labour, up front. Claude does not generate the image. Anthropic's docs are direct about it: Claude "is an image understanding model only. It can interpret and analyze images, but it cannot generate, produce, edit, manipulate, or create images" (Claude vision docs). What Claude is genuinely great at is the part that decides whether a product shot looks expensive: reading your actual photo (material, light, angle, label) and writing photography direction an image model can follow. Claude art-directs. An image model shoots. Your real photos go in as reference, which is why the labels and stitching stay yours instead of turning into AI mush.
the shape of it

How the whole thing runs

  1. ๐Ÿ“ฑ Shoot 3 to 6 phone photos. Any background. You want the hero angle, one straight-on front so the label reads clearly, one from above, and a close-up of the detail you care about (stitching, texture, foil, seam).
  2. ๐Ÿ“ฆ Install the skill once. Copy the SKILL.md below into a folder, zip it, and upload it to Claude. From then on it triggers whenever you ask for product shots.
  3. ๐Ÿง  Drop the photos into Claude and pick a shot type. White background, lifestyle scene, or editorial flat lay. Claude reads your photos first, then writes a full photography brief with the details it can see locked in.
  4. ๐Ÿ–ผ Render it in an image model that takes reference images. You paste Claude's prompt plus your original photos into a model that accepts reference images, so it is editing your product rather than inventing a lookalike.
  5. ๐Ÿ”Ž Check the label, then ship it. Zoom in on text and logos before anything goes on a product page. The five second check at the end of this guide catches almost everything.

do this first

How to shoot the phone pics

The output can only be as accurate as your reference. Ninety percent of "the AI changed my product" complaints come from feeding it three blurry photos of the same angle in yellow kitchen light.

Window light, no flash Put the product on a table near a window, turn the overhead lights off, and shoot with the light coming from the side. Flash flattens texture and throws a hard shadow the model then tries to copy.
Cover every side that matters Front, three quarter, top down, and one macro of the detail. The model can only keep what it can see, so an unphotographed side gets invented.
Make the label readable One shot where the text is sharp, square on, and fills a decent chunk of the frame. This is the single photo that decides whether your brand name survives.
Keep the frame boring A plain wall, a bedsheet, a benchtop. Clutter behind the product gives the model more to hallucinate around, and you are replacing the background anyway.

copy this

The product photography skill

This is the whole thing. A skill is a folder with one file in it, and this is that file. Copy everything in the box into a file named SKILL.md, inside a folder named product-photography.

๐Ÿ“„ SKILL.md (copy all of it)
---
name: product-photography
description: Turns phone photos of a physical product into studio-grade photography prompts and shot specs. Use when I upload product photos and ask for studio shots, white background shots, lifestyle images, flat lays, or ecommerce photos.
---
# Product photography direction

## What you do
You are a commercial product photographer and art director. The user gives you real phone
photos of a real product. You do not generate images. You write the photography brief and the
final render prompt that an image model will follow, using their photos as reference.

## Step 1: read the photos before writing anything
Look at every photo the user uploaded and report back, in plain language:
- What the product is, and what it is made of (glass, matte plastic, brushed metal, ceramic,
  fabric, paper stock, leather).
- Every piece of text and every logo you can read, transcribed exactly, including capitalisation.
- The finish: glossy, matte, satin, translucent, textured, embossed.
- The real colours, described as a photographer would (warm cream, cool charcoal, sage green),
  not as hex codes.
- The physical details that must survive: stitching, seams, spouts, caps, labels, ridges, weave.
- The strongest single feature of this product, the thing a shot should be built around.
If a detail is not visible in any photo, say so and ask for one more angle instead of guessing.

## Step 2: confirm the shot type
Ask which of the three the user wants, and let them pick more than one:
1. White background (ecommerce and marketplace listings)
2. Lifestyle scene (social, ads, website hero)
3. Editorial flat lay (launches, press, gifting content)

## Step 3: write the render prompt
Write one complete prompt per requested shot type. Each prompt is a single paragraph of
photography direction, written for an image model that is receiving the user's real photos as
reference images. Every prompt must contain, in this order:
- The instruction to keep the product exactly as shown in the reference photos, preserving all
  text, logos, proportions, materials and colour, and to change only the environment, lighting
  and camera.
- The exact label text transcribed from the photos, in quotes, so the model has it in writing.
- The set: surface, background, and any props (props only for lifestyle and flat lay).
- The lighting: source, direction, quality (soft or hard), and where the shadow falls.
- The camera: angle, approximate focal length, depth of field.
- The output spec: aspect ratio and where the negative space sits for text overlays.
Do not use the words "photorealistic", "8k", "masterpiece" or "high quality". Describe the actual
scene instead. Write in full sentences, not comma-separated keyword soup.

## Step 4: hand over the checklist
End every response with a short verification list of the specific details in this product that
the user must zoom in on before publishing: the exact label text, the logo shape, the closure or
cap, the count of any repeated element (buttons, stripes, ridges), and the material finish.

## Rules
- Never invent a claim, certification, award badge, ingredient or price that is not in the photos.
- Never change the product shape to make it prettier. Accuracy beats beauty every time.
- If asked for a person holding the product, keep the product size correct relative to the hand.
- Keep the tone of the brief practical. The user is shipping to a product page, not entering a
  competition.
get it in there

Install the skill in Claude

  1. ๐Ÿ“ Make the folder. Create a folder called product-photography and put the SKILL.md file inside it. That folder is the skill.
  2. ๐Ÿ—œ Zip it. Right click the folder and compress it, so you have product-photography.zip. On Mac that is "Compress", on Windows "Send to, compressed folder".
  3. โž• Upload it. In the Claude app or on claude.ai, open Customize, then Skills, then Add, then Upload a skill, and pick the zip. Skills are available on every plan, Free included. Anthropic's own walkthrough is here.
  4. ๐Ÿ’ป Or drop it into Claude Code. If you use Claude Code, put the product-photography folder in ~/.claude/skills/ and it loads on the next session. No zip needed.
  5. ๐Ÿงช Test it. Start a new chat, upload your photos, and say "run my product photography skill on these, white background version". If it does not trigger by itself, name it in the message once and it will.
The description line at the top of the SKILL.md is what makes Claude reach for the skill on its own, so keep it specific about when to use it. If you rename the skill, avoid putting "claude" or "anthropic" in the name, because the upload gets rejected. More on how skills work in the official skills overview.

the three looks

The three shot types, with the prompts

These are the prompts the skill writes for you, with the bracketed parts filled in from your photos. They also work on their own if you want to paste one straight into an image model today. Upload your reference photos in the same message.

1. White background, for product pages, marketplaces and anywhere a listing needs a clean cutout.

๐Ÿ“ธ white background listing shot
Using the attached reference photos as the exact source of truth for the product, place this same product on a seamless pure white studio background with no visible horizon line. Keep the product identical to the references: same shape, same proportions, same material and finish, same colours, and keep every piece of text and every logo exactly as it appears, including the label reading "[PASTE THE EXACT LABEL TEXT]". Do not redesign, restyle or clean up the packaging.

Light it like a professional ecommerce studio: a large soft key light from the upper left, a softer fill from the right so the shadow side still shows detail, and a subtle contact shadow directly under the product so it sits on the surface instead of floating. Add one soft specular highlight down the edge to show the material finish, with no blown-out hotspots.

Camera: straight-on at product height, roughly a 50mm to 85mm lens look, everything in sharp focus front to back. Frame the product centred with even margins and generous white space around it, square 1:1 aspect ratio. Output a clean, unedited-looking photograph with no text, no graphics, no watermarks and no props.

2. Lifestyle scene, for ads, social and your website hero, where the product needs a world around it.

๐ŸŒฑ lifestyle scene shot
Using the attached reference photos as the exact source of truth for the product, place this same product into a real lifestyle scene: [DESCRIBE THE SETTING, for example "a warm morning kitchen bench in soft natural light, with a linen cloth and a small ceramic dish just out of focus behind it"]. The product must stay exactly as it appears in the references, including its shape, material, finish, colours, and every piece of label text and logo, reading "[PASTE THE EXACT LABEL TEXT]". Change only the environment, the lighting and the camera.

Light the scene with directional window light coming from the left, soft shadows falling to the right, and a warm colour temperature around golden hour. Nothing in the scene should compete with the product for attention: props stay muted, partly cropped, and behind the plane of the product.

Camera: a three-quarter angle slightly above product height, roughly a 35mm to 50mm lens look, shallow depth of field so the background softens while the product and its label stay crisp. Frame it 4:5 vertical with clean negative space in the upper third for a headline. The result should look like a photograph taken on a real set, with natural imperfections in the surface and fabric, not a rendered image.

3. Editorial flat lay, for launches, press kits and gifting content, shot from directly above.

๐Ÿ“ editorial flat lay
Using the attached reference photos as the exact source of truth for the product, build a top-down editorial flat lay with this product as the hero. Keep the product exactly as shown in the references: same shape, proportions, material, finish and colour, and preserve every piece of text and every logo, including "[PASTE THE EXACT LABEL TEXT]". Do not restyle the packaging or add branding that is not in the photos.

Arrange it on [DESCRIBE THE SURFACE, for example "a textured cream plaster surface"] with two or three restrained supporting elements placed on a loose diagonal: [LIST YOUR PROPS, for example "a folded linen swatch, a sprig of dried eucalyptus, and one raw ingredient"]. Leave real breathing room between objects. This is an editorial page, not a crowded market stall.

Light it with a single large soft source from the top left, producing gentle directional shadows that all fall the same way, with no double shadows and no harsh edges. Camera: perfectly perpendicular from directly overhead, roughly a 50mm lens look, everything evenly sharp. Frame it 4:5 vertical with the hero product slightly off-centre and clear negative space in one corner for a headline. Muted, cohesive colour palette pulled from the product itself.

where it renders

Where the picture actually gets made

Claude writes the brief. You need one image model that accepts reference images, so it edits your product instead of dreaming up a new one. Three honest options, cheapest first.

  1. ๐Ÿ†“ Google AI Studio, free to try. Open aistudio.google.com with a Google account, upload your photos, paste the prompt, and generate with the Nano Banana 2 image model (available there in preview, per Google's own announcement). The free tier is generous and costs nothing to test. The catch is data: Google's terms for the free tier say human reviewers "may read, annotate, and process your API input and output" and that submitted content is used to "provide, improve, and develop Google products and services", and they explicitly say not to submit confidential information (terms). Fine for your own candle photos, not for an unreleased client launch.
  2. ๐Ÿ’ณ fal.ai, pay per image. Open the Nano Banana 2 edit model on fal, upload up to 14 reference photos, paste your prompt, and run it in the browser playground. It is $0.08 per image at 1K (2K costs 1.5x and 4K costs 2x), pay as you go with no subscription. This is the one I would use for anything going on a real store, and it is also the route the Claude Code version below uses.
  3. ๐ŸŽจ Reve, if you want a studio interface. Reve is a 4K image model with a proper editing interface and a "reference anything" feature, so you can push a shot around visually after the first render instead of rewriting the prompt. Check its current plans on the site before you commit to a month.
Whatever you pick, upload the same reference photos every time you regenerate. Dropping the references and reusing the text prompt alone is exactly how you end up with a product that looks nearly, but not quite, like yours.

the power version

Make the skill render it for you

This is optional and it is the version from the reel where you drop photos in one end and get finished images out the other. It needs Claude Code and a fal account with a few dollars on it. You still write no code, Claude does.

  1. ๐Ÿ”‘ Get a fal key. Sign up at fal.ai, add a small amount of credit, and create an API key in your dashboard. At 8 cents an image, ten dollars is over a hundred renders.
  2. ๐Ÿ— Store it as an environment variable, not in the file. In your terminal, run export FAL_KEY="your-key-here", or put that same line in a .env file in your project folder. Never paste a key into a chat message or into the SKILL.md itself.
  3. ๐Ÿ“‚ Make a working folder. Create a folder with a photos/ subfolder, drop your phone pics in there, and open that folder in Claude Code.
  4. ๐Ÿ›  Ask Claude Code to build the render step. Paste the prompt below. It writes a small script that sends your photos plus the generated prompt to the model and saves the results.
  5. ๐Ÿ” Run it on every new product. New photos into photos/, one sentence to Claude, three finished shots out. That is the part that turns this from a fun afternoon into an actual system.
๐Ÿ‘ฉโ€๐Ÿ’ป build the render step in Claude Code
I have a folder of phone photos of one product in ./photos and my product-photography skill installed. Build me a small render pipeline so I can go from photos to finished studio images in one command.

Read the photos first and run my product photography skill on them to produce the three render prompts (white background, lifestyle, editorial flat lay). Then write a Python script called render.py that uploads each photo in ./photos to fal storage, calls the fal-ai/nano-banana-2/edit model once per shot type with that shot's prompt and all of the uploaded photos as image_urls, and saves the returned images into ./output with clear filenames like white-background-01.png. Read the API key from the FAL_KEY environment variable and never hardcode it, and print the cost estimate before running so I can approve it.

Use the official fal Python client, install anything it needs, tell me exactly what you installed, and then do one test run on a single shot type so we can check the label text survived before spending money on the full batch.
Two guardrails. Keep your API key out of any file you would ever share or commit, environment variables only. And run one test image before a batch, because the difference between a good prompt and a great one is usually a detail you can only see in the first render.

before you post

The five second accuracy check

This is the step that separates shots you can put on a store from AI slop. Zoom to 100 percent and check, in this order:

  • The label text. Read every word out loud. Image models still drop and swap letters, and a misspelled brand name on a live product page is the worst possible version of this.
  • The logo shape. Compare it against your original photo side by side, not from memory.
  • Counts and hardware. Buttons, stripes, ridges, seams, the cap, the pump, the zip pull. Models quietly add or remove one.
  • The shadow. One light source means one shadow direction. Two shadows going different ways is the tell that reads as fake even to people who cannot say why.
  • Scale. If a hand or a prop is in frame, check the product is still the size it is in real life.
Two things worth knowing before these go anywhere commercial. Images from this model carry an invisible SynthID watermark, so they are detectable as AI generated by design. And if you are advertising a physical product, the image still has to represent what the buyer receives, so keep the product true and let the set and the lighting be the AI part.

all the links

Everything in one place

  • Claude (where the skill lives, works on Free): claude.ai
  • How to create custom skills (Anthropic's walkthrough): support.claude.com
  • Skills overview (what a SKILL.md can do): platform.claude.com
  • Google AI Studio (free to test, read the data terms): aistudio.google.com
  • Nano Banana 2 edit on fal ($0.08 per image at 1K, up to 14 reference photos): fal.ai
  • Reve (4K image studio with reference editing): reve.com
Want to build systems like this properly? Inside AI System Club I teach people to learn AI from scratch and build real systems, for work and for their own business. Doors open soon and founding pricing dies at close. ๐ŸŒฑ Join the waitlist (link in my bio too).
the waitlist is now open

AI System Club is launching soon

Build AI systems with me through recorded lessons you can follow at your own pace. Join the waitlist for first access and the founding price.

btw this list gets first dibs
the moment doors open :)