AI Guides › Step-by-step guides
By Nigel Guy · 6 min read
Most people plan a room redesign with a mood board: twenty beautiful rooms that are not your room, with your awkward window and your sofa wall nowhere in sight. It feels like progress, and then the paint arrives and the whole thing looks wrong. A mood board shows you someone else's finished room. It cannot show you yours.
The rule: change one thing per image, always start from your own photo, and treat every result as a sketch, never a spec.
Claude Code cannot draw pictures itself. What it can do is write a small script, or run a connected tool, that sends your photo and a short instruction to an image model, saves the result, and looks at it. This guide builds that, then gives you a prompt that keeps the edits disciplined.
You need:
room-rebuild).room.jpg in that folder. Stand in a corner or doorway so the main wall and the floor are both visible.gpt-image-2.5-sunburst (edit precision) and gpt-image-2.5-flare (everyday speed), and say you may need to complete API Organization Verification first. Prices vary by quality and size, so check the £ price on OpenAI's pricing page before you start.Budget a few pounds for a session of 30 to 60 images. Model names and prices change often; check the provider's page on the day.
Privacy: your photo is uploaded to that provider. Crop out family photos, documents and anything with your address on it before you start.
There are two routes. Take the first unless you already use MCP.
Route A: a local script (simplest, and you can read every line). In the folder, start Claude Code and paste:
Read the current official image-editing documentation for [PROVIDER: Gemini API or OpenAI Image API] and write a small command-line script, [LANGUAGE: Python], that takes an input image path, a text instruction and an output filename, sends them to [MODEL_NAME] and saves the returned image. Read the API key from the environment variable [KEY_NAME], never from the file. Do not run it yet. Show me the script and tell me how you will install anything it needs.
Fill in the provider, language, model name from the docs, and the variable name (for example GEMINI_API_KEY).
Route B: an MCP server. Claude Code adds servers with claude mcp add. For a local server the syntax is claude mcp add --env KEY=value --transport stdio <name> -- <command> [args]; all options go before the name, and -- separates them from the server command. The default scope is local to the current project; --scope user makes it available in every project. Check it with claude mcp list or /mcp. The official docs warn you to verify you trust each server: an image server you did not write sees your key and your photo. Only use one whose code you or someone you trust has read.
export GEMINI_API_KEY="..." (macOS or Linux), substituting your own variable name..gitignore.You are helping me preview a room redesign before I buy anything. The photo is [PHOTO_PATH]. The room is [ROOM_TYPE], roughly [APPROX_SIZE], and the fixed things that cannot change are [FIXED_ITEMS: windows, radiator, door, floor, built-ins].
Goal: produce a small set of realistic mock-ups where I change one thing at a time.
Steps:
1. Look at the photo and describe back to me, in five lines, what you see, including the light direction and anything you are unsure about. Wait for my correction.
2. Then, for each change in this list, create one image using the image script and save it as [OUTPUT_PREFIX]-01, -02 and so on: [CHANGE_LIST: for example wall colour to a muted green; a two-seat sofa in oatmeal fabric; rug moved under the coffee table].
3. In every instruction to the image model, say explicitly that the camera angle, room shape, windows, doors and lighting must stay the same, and that only the named item changes. Name the colour or material in plain words, for example "sage green, matte".
4. After each image, open it and check it against the original. Tell me in two lines what changed, and flag anything that moved, warped or disappeared that I did not ask for.
5. Finish with a table: image file, the one change, what looks convincing, what looks wrong.
Rules: do not invent dimensions or product names. If a measurement matters, ask me. If something is missing from my inputs, ask before generating. Do not change more than one thing per image. Before you answer, check each image really starts from my photo and not from the previous result, unless I asked for stacking.
Fill in the bracketed items. Once you like two or three single changes, ask for a combined image as a separate final step, so you can see which change did what.
room.jpg.