Working with AI: a guide for set designers and set decorators
About this document
- What it is: A step-by-step guide to setting up and using an AI assistant for set design and set decoration: which companies to use, where your files live, which skills to use for which task, and a way of working you build up yourself.
- How to use it: read it once, start to finish. Then do the steps in order. Each step says what to do, where to do it, and what you should see when it has worked. The skills list in 1.6 is the part you’ll come back to most: look at it whenever you start a new kind of task. If what you see on screen differs from the guide, leave a comment on that line.
- Made by: Kev Freeney, with Claude
- For: Anyone working as a set designer or set decorator. You don’t need a technical background.
- Why: So you can use AI on your own terms: your files stay yours, your assistant works the way you do, and you know which tool fits which task.
- Project: The AI guides: one guide to working with AI, tailored for each craft. This is the set design edition.
Updated 14 November 2026. Facts about companies were checked on that date. They change, so each one says where to look again.
This guide sets you up to use AI on your own terms. By the end you will have three things: a folder that holds how you work, kept in your cloud storage or on this computer; an AI assistant that can read and edit it; and a way of working with that assistant that gets better with every project.
You need a computer, a web browser, and somewhere to keep your files: a cloud storage service you already use, or this computer with a backup (1.3 helps you choose). A paid Claude plan helps for some steps, and the guide says which.
How to get started (Mac, about ten minutes)
- In this Drive folder, click Start here (Mac).zip, then Download. If Drive says you don’t have access, ask whoever sent you the link to share the folder with the email address of your Google account, then reload the page.
- Open your Downloads folder and double-click the zip. A folder called Start here appears. (Safari often unzips the file for you. If the folder is already there, go to step 3.)
- Open that folder and double-click Start here.command.
- Your Mac will probably say it can’t check who made it. Click Done. Then open System Settings → Privacy & Security, scroll down, and click Open Anyway. Enter your Mac password. This is the one moment of trust, and it is yours to give: Start here.command is a plain text file, so you can right-click it, choose Open With → TextEdit, and read every line first. On some versions of macOS the warning looks different or doesn’t appear. We haven’t been able to test every Mac, so if yours differs, carry on if it lets you, and leave a comment on this line.
- A Terminal window opens and walks you through the rest. Each screen asks before it changes anything, and you can type q to stop at any time.
On Windows, or if you’d rather work by hand? Go to Working by hand below. A Windows version of Start here is planned; this guide will say when it is in the folder.
Working by hand (any computer, any chat)
You need a chat that can keep files: a Claude Project, a ChatGPT Project, or the Gemini equivalent. No script runs. You do the steps the script would do.
- In this Drive folder, open By hand and download USE.md, brief.md, project-checklist.md and the files in Project template.
- Make a new project in your chat, and upload all of those files to it.
- Type: “Read USE.md and the files I’ve uploaded. Follow USE.md. Then ask me what kind of work I do.”
- When the chat writes something worth keeping (a brief, a Where we are note), copy it into the matching file in your own folder yourself. A chat in a browser can’t save into your Drive. Do this at the end of every session, and upload the new version to the chat project, replacing the old one.
- Start each new job from this checklist, ticking each line:
Part 1. Your AI studio
Your studio is the setup you keep from project to project: which AI you use, what you pay, where your files live, what your assistant can reach, and which skills it uses for which task.
1.1 Choose who you work with
Use one of the big American assistants as your main one: Claude (Anthropic), ChatGPT (OpenAI) or Gemini (Google). This guide uses Claude for its examples, because it can read and edit your Google Docs, Sheets and Slides and work inside folders on your computer. The others can do most of the same.
Why American, for now: all of these assistants can use your chats unless you change a setting, and the American ones give you a clear switch to stop your chats being used to train their models. The Chinese assistants (DeepSeek, Kimi) say in their own policies that they store your data on servers in China, under Chinese law. (Checked 14 Nov 2026: Anthropic, OpenAI and Google help pages, the DeepSeek privacy policy, the Kimi help centre.)
Do this now, whichever you choose:
- Claude: go to Settings → Privacy, find Model improvement (the switch is called Help improve Claude) and turn it off. With it off, Anthropic keeps your chats for 30 days instead of up to five years. (Anthropic’s privacy centre names the setting “Model Improvement” under Privacy Settings, checked 14 Nov 2026. The two retention times come from Anthropic’s 2025 policy update and were not on the page we checked, so read them on the screen you see.)
- ChatGPT: go to Settings → Data Controls and switch off Improve the model for everyone.
- Gemini: open Gemini Apps Activity in your Google account and switch off Keep Activity. Google says human reviewers may still see some chats, and reviewed chats are kept for up to three years.
You should see the switch in the off position. It applies to chats from now on.
1.2 Decide how you pay
There are two ways to run AI: pay a monthly subscription, or buy a computer powerful enough to run AI models yourself.
- If you don’t already have that kind of computer, use subscriptions. They cost less to start, they stay up to date, and you can stop any month.
- A local machine makes sense when work must stay on your own computer, or when you run AI all day. It needs a lot of memory and costs much more up front. As a rough rule of ours, not a company’s figure: a recent computer with 32 GB of memory or more runs the small models, and the better ones need a good deal more.
Pay by subscription, never by the amount you use. Some tools let you plug in a “key” that bills you for every request. Those bills are hard to predict, so leave them alone until you know you need one.
Prices change often, so check each company’s plans page before you pay: Claude at claude.com/pricing, ChatGPT at chatgpt.com/pricing, Gemini at gemini.google/subscriptions (all three opened on 14 Nov 2026).
Stay on top of your subscriptions. AI tools are easy to sign up for and easy to forget.
- Make a spreadsheet called Subscriptions (in Google Drive: New → Google Sheets; or in Excel or Numbers).
- Make these columns: Tool · What I use it for · Plan · Monthly cost · Renews on · Last used · Keep or cancel.
- Add every AI tool you pay for, including free trials that will start charging.
- On the first of each month, read down the Last used column and cancel anything you haven’t used in a month.
1.3 Decide where your files live
Everything you and your assistant make lives in files you own, in one place. Pick the first of these three that fits you.
A. Git. If you or your studio already keep work in Git (for example on GitHub), keep your text and its history there too, and back it up to your usual remote. If you’ve never used Git, go to B.
B. Cloud storage. Google Drive, Dropbox, OneDrive or iCloud Drive. Your files are backed up, you can open them on any device, and you can share them. If you have more than one, pick the one your clients and crew use, and keep everything there. Install its desktop app so the files are on your computer too. For Google Drive, install Google Drive for desktop and, in its Preferences, set My Drive to Mirror files: Mirror keeps a full copy on your computer, which suits apps that read and write a lot of files.
C. None of these. Keep everything in one folder on this computer, inside Documents, and back it up. Choose one habit and keep it: - at the end of each working day, right-click the folder, choose Compress, and email the zip to yourself; - once a week, copy the folder to an external drive; - or sign up for a storage service, which moves you to B. Google One (one.google.com/about/plans), Dropbox (dropbox.com/plans) and iCloud+ (apple.com/icloud) each list their current plans and prices on their own page; OneDrive’s are on Microsoft’s site (search for “OneDrive plans”). We opened the first three on 14 Nov 2026 and copy no prices here, because they change.
On a Mac, your assistant can check which of these you already have (an automatic check is planned, for Mac first). On Windows or Linux, choose by hand from the same three.
The rest of this guide says Doc and Sheet. If your files are not in Google Drive, use whatever you write and count in (Word, Pages, Excel, Numbers, plain text). Your assistant reads them all.
1.4 Make your AI studio
- In the place you chose in 1.3, make a folder called AI studio. This folder is for you, across every project.
- Inside AI studio, make:
- a Doc called Read me: the instructions your assistant follows (2.2)
- a Doc called Taste: how you like things (2.1)
- a Doc called Corrections: every time the assistant gets something wrong, one dated line on what it should do instead (2.10)
- your Subscriptions sheet (move it in)
- a folder called Skills, for skill files you download (1.6)
- a folder called Project template, which you fill in once (3.1)
- Next to AI studio, make a folder called Projects. Every project gets its own folder inside it, made from the template (3.1).
You should see:
AI studio
├── Read me
├── Taste
├── Corrections
├── Subscriptions
├── Skills
└── Project template
Projects
1.5 Give your assistant access to your files
Pick one. All three end in the same place: your assistant can read and edit the files in your folders.
A. In the Claude app or website (files in Google Drive). 1. Open a new chat. Click the + button, then Connectors. 2. Switch on Google Drive, Google Docs, Google Sheets and Google Slides. 3. Claude asks you to sign in to Google. Sign in and allow access.
Test it: type “List the files in my AI studio folder.” Claude should name them.
What each connector does (Claude help centre, checked 14 Nov 2026): - Google Drive: find, upload, share and move files, and make folders. - Google Docs: edit a Doc live in a pane beside the chat, and read the comments and suggestions on it. - Google Sheets and Google Slides: edit and create sheets and slides live.
On a Team or Enterprise plan, your administrator has to switch these on first.
B. In Claude Cowork (desktop app, paid plans; any storage). Cowork works directly inside folders on your computer, so it suits Dropbox, OneDrive, iCloud Drive and local folders as well as Drive. 1. If your files are in cloud storage, make sure its desktop app is installed and the files are on this computer (1.3). 2. Open the Claude desktop app and start Cowork. 3. When it asks which folder to work in, choose AI studio or one project folder. Give it only the folders it needs.
C. In Claude Code (for anyone happy to paste a line into
Terminal; any storage). 1. Open Terminal (on a
Mac: press Cmd+Space, type Terminal, press Return). 2. Paste this line
and press Return. It installs Claude Code:
curl -fsSL https://claude.ai/install.sh | bash 3. Close
Terminal, open it again, go to your AI studio folder, and type
claude. Sign in with your Claude account. 4. The connectors
you switched on in A should also appear here, because you are signed in
to the same account. Type /mcp to see them. If Google Drive
is missing, switch it off and on again at claude.ai, then restart Claude
Code. Some people report that it doesn’t always carry over (user
reports, checked 14 Nov 2026; Anthropic’s own pages don’t promise it),
so if it still isn’t there, work through the Claude app for those Google
files.
The one-step install. The Drive folder this guide sits in also holds Start here (Mac).zip. It looks at what you already have, asks before every install, signs you in, and sets up your studio with the skills in 1.6. It prints every command it runs. See “How to get started” above.
1.6 Skills, and when to use them
A skill is a folder of written instructions your assistant loads when it needs them: how to interview you, how to make a spreadsheet, how to check its own work. They are free, and they are plain text you can read first.
This list is the heart of the guide. Before any new kind of task, look here for the skill that fits, and tell your assistant to use it by name. Keep the list short: add a skill when a task keeps coming back, and drop one you haven’t used in a month (2.13).
| Skill | Use it when | For this task | Where it comes from | How to install it |
|---|---|---|---|---|
| grill-me | You start a project, or anything feels unclear | Asks you questions one at a time, each with its recommended answer, until it understands exactly what you mean | github.com/mattpocock/skills (folder skills/productivity/grill-me) | Claude Code:
npx skills add mattpocock/skills --skill grill-me -g. App:
upload the folder (below) |
| brief | You start a project | Writes your brief from the template in 3.2, checks it, and asks for anything missing, most urgent first | Comes with the kit in this guide’s Drive folder (Start here (Mac).zip) | Without the kit: paste the template in 3.2 into your Brief and use grill-me to fill it |
| docx | You need a Word file to download or send | Makes and reads Word documents | github.com/anthropics/skills (skills/docx) | App: built in on paid plans. Claude Code:
/plugin marketplace add anthropics/skills, then
/plugin install document-skills@anthropic-agent-skills
(adds docx, xlsx, pptx and pdf together) |
| xlsx | You need a spreadsheet file to send | Makes and reads Excel files, with working formulas | github.com/anthropics/skills (skills/xlsx) | As docx |
| pptx | You need a slide deck file to send | Makes and reads PowerPoint decks | github.com/anthropics/skills (skills/pptx) | As docx |
| You’re given a PDF, or must send one | Reads, fills, merges and splits PDFs, and pulls tables out of them | github.com/anthropics/skills (skills/pdf) | As docx | |
| doc-coauthoring | You’re writing a longer document: a proposal, a pitch, a report | Writes it with you, section by section, asking before it fills gaps | github.com/anthropics/skills (skills/doc-coauthoring) | App: built in on paid plans. Claude Code: after adding the
marketplace (docx row),
/plugin install example-skills@anthropic-agent-skills |
| skill-creator | You ask for the same thing on every project | Turns that ask into a skill of your own (2.14) | github.com/anthropics/skills (skills/skill-creator) | As doc-coauthoring |
| find-skills | You need something no skill here covers | Finds a skill for a task you describe in plain words | github.com/vercel-labs/skills | Claude Code:
npx skills add vercel-labs/skills --skill find-skills -g |
| three.js skills | You want to show a model in a web browser | Turns a 3D model into a view anyone can open, turn and zoom in a browser | github.com/CloudAI-X/threejs-skills (MIT licence) | Claude Code:
npx skills add CloudAI-X/threejs-skills -g |
| Superpowers | You use Claude Code and the work has many steps | Works in careful steps: asks before building, plans in small tasks, checks before calling anything done | github.com/obra/superpowers (MIT licence) | Claude Code:
/plugin install superpowers@claude-plugins-official |
| Blender connector | The design is agreed and it’s time to model it | Lets your assistant build and change a Blender scene, render views and export drawings | github.com/ahujasid/mcp-for-blender (formerly blender-mcp) | Claude Code: claude mcp add blender uvx blender-mcp,
then the Blender add-on (3.7) |
| SketchUp connector | You model in SketchUp | Lets your assistant read and change a SketchUp model | github.com/mhyrr/sketchup-mcp (MIT licence, checked 14 Nov 2026) | On a Mac, install uv first (brew install uv, or ask
your assistant to do it and show each command). Then in SketchUp open
Window → Extension Manager, click Install Extension, choose the
connector’s .rbz file from its download page, and restart
SketchUp. Start the server from Extensions → SketchupMCP → Start Server.
Add
"sketchup": {"command": "uvx", "args": ["sketchup-mcp"]} to
your assistant’s connector settings. (Steps from the connector’s GitHub
page, checked 14 Nov 2026; not installed by us.) |
Every GitHub link above was opened on 14 November 2026, and the install commands were checked against our own installs, except where marked. Check them again on the day you use them.
To add a skill in the Claude app: 1. Go to Settings → Capabilities and switch on code execution. 2. On the skill’s GitHub page, click Code → Download ZIP. Open the download, find the skill’s own folder (for grill-me: skills/productivity/grill-me), right-click it and choose Compress. Save that zip in your Skills folder. 3. Go to Customize → Skills. Click +, then Create skill, then Upload a skill, and choose the zip.
You should see the skill in your list. Skills you add here also load in Claude Code when you’re signed in to the same account. (Claude help centre, checked 14 Nov 2026. The menu names have moved before; if you can’t find them, search the help centre for “use skills”.)
In Claude Code, the npx skills add
lines need Node.js on your computer. If Terminal says
npx: command not found, ask Claude Code instead:
“Install the grill-me skill from github.com/mattpocock/skills for
me, and show me each command before you run it.”
The connectors in 1.5 edit your live Google files. The docx, xlsx and pptx skills make files to download or send. You’ll want both.
To find more: search skills.sh for the task in plain words. Pick skills with many installs from an author you can check, and read the SKILL.md file before you add it.
1.7 Connect your other tools
A connector (also called an MCP) lets your assistant work inside another app: open a board, read a model, add a note. Add only the ones you use.
- Miro (boards): Miro’s own connector lets your
assistant read a board, add sticky notes, frames, diagrams and tables,
and lay out references. In Claude Code:
/plugin install miro@claude-plugins-official. (Made by Miro, MIT licence, github.com/miroapp/miro-ai.) In the Claude app, open Customize → Connectors, look for Miro and click Connect. (We could not confirm Miro’s listing on Anthropic’s own pages on 14 Nov 2026, so if you can’t find it there, use the Claude Code line above.) - The software your work ends up in: see 3.7.
1.8 Which model for which job
Each AI company offers several models: the engines
behind the assistant, from small and fast to large and careful. On 14
November 2026 Claude’s were Haiku 5.5 (small and fast), Sonnet 5.5 (in
between) and Opus 5.5 (strong), and you pick one from the model menu in
the app or by typing /model in Claude Code. Some plans also
offer a top model above Opus, which can cost extra: on 14 November 2026
Claude’s was Fable, not on the Free plan, paid for with extra usage on
Pro, and part of the weekly limit on Max, Ultra and Team premium seats
(Claude’s plans page). Check your plan before you pick it. ChatGPT and
Gemini have their own small and large models. Names change every few
months, so check the menu again before you rely on this line.
Nine rules that hold whichever company you use:
- A formula before a model. If a spreadsheet formula, a search or a rename does the job, use that, not AI. A formula gives the same right answer every time, at no cost.
- Match the model to the job. A small, fast model for sorting and pulling details out of many items. A strong model for judgement and writing. The top model only for hard planning and reviewing important work. Start one step lower than you think you need.
- Pay by subscription (1.2).
- Keep each chat small. One task per chat. Keep a “Where we are” note (2.11) and start a fresh chat when one gets long. Assistants get worse at remembering the start of a very long chat.
- Brief helpers tightly. When your assistant hands part of the work to a smaller helper, the helper gets the exact files, the shape of the answer, and where to save it.
- Someone else checks the work. Check with a formula, or with a fresh chat given only the work and the brief, rather than asking the maker whether it got it right.
- A plan for anything left running. Anything left working while you’re away has a written plan, and you spot-check what it made.
- The smallest change that solves the problem. Ask for the fix and keep the rest as it is.
- One writer per file. You and an assistant, or two assistants, never edit the same file at the same time.
When a model fails the same step twice, stop and change the approach rather than asking harder. When it says it can’t do something, the input is usually missing a piece, so add it.
1.9 Resources: apps worth looking at
These are for you to look into. Try a free tier before paying, and read two things in each tool’s terms first: who owns what it makes, and whether it trains on what you upload.
Collect and organise references - Cosmos: an AI-searchable image collection that imports Pinterest boards and Are.na channels. - Eagle: an image library on your own disk, with tags and search by colour. One-off purchase. - Are.na: slow, research-style collections with no algorithm pushing images at you. - Milanote: boards for building an argument with images, notes and arrows. - ShotDeck: film stills searchable by colour, time of day and composition. Useful with the DoP.
Colour - Adobe Color and Coolors: build palettes from images. - Pantone Connect: shared colour references for dye shops and costume. - Nix colour sensor: hold it against real fabric, paint or wood and get its colour values.
Draw and annotate - Procreate (iPad): draw over renders and photos. - Morpholio Trace: sketch over plans at real scale. - Miro: a shared board for the team (see 1.7 for its connector).
3D and drawings - Blender: free and open source. Models, renders, camera views and drawings. - SketchUp: widely used for set models. - Vectorworks: drawings and models for entertainment design. A reviewer on Best of AI: “Users can produce drawings, renderings, and schedules from the same model.” (bestofai.io, read 14 Nov 2026) - Rodin, Tripo or Hunyuan3D: turn a photo of one object into a 3D object. - World Labs Marble: turns a photo into a 3D space you can move through.
Mood images - Krea: a reviewer on Best of AI: “As you type a prompt or sketch shapes, the image updates almost instantly, which turns generation into a live, exploratory process instead of waiting for each result.” The same review notes that “real-time modes trade off some fidelity for speed.” (bestofai.io, read 14 Nov 2026) - Midjourney, FLORA, Figma Weave: other image tools worth comparing.
Keep it - Obsidian: notes as plain text files in a folder. Works offline and opens in any editor years from now.
What we used, once, as an example: on a recent set we used Claude for the writing and scripts, Blender for the model and drawings, three.js to show the model in a browser, and Higgsfield for mood images. That was our setup. Yours should fit how you work.
Part 2. Your approach
Your approach is how you like to work. Write it down once, your assistant follows it on every project, and it grows a little each time.
2.1 Write your taste
Open the Taste Doc in AI studio. Answer these in your own words. Short answers are fine, and you can keep adding over time.
Flair. Three sets you’d put your name to. What exactly, in each? A colour, a material, a way of ageing cloth?
What makes you wince. Three things you see in references that you’d never use, and why.
Materials. The ones you trust, the ones you avoid, how you like things aged.
Colour. Your working palette, named, with a photo of a real swatch for each.
How you like to present. Slides, documents or boards? Which fonts, what layout, what you can’t stand seeing. Your assistant follows this whenever it makes something for you.
Your Taste stays private. It goes to a client or production only if you send it.
2.2 Write your Read me
Open the Read me Doc. Your assistant reads this first, every time. Paste this in and change anything that isn’t you:
You are working with a set designer. Your job is to help me think, sort, write and check. I make every decision.
Read first, every time: my Read me, Taste and Corrections; then, in the project folder, the Project Doc (where we are, and what’s already decided: never reopen a decision unless I ask) and the Brief (what it is and what it is not; the “not” list overrides everything).
How to work: - Before you make anything, list what already exists in my folders that could do the job. - Before you start, say in one sentence what you’re about to make. If anything is unclear, ask me one short question first. - I’ll ask in four parts: Task, Done means, Context, Check. If one is missing, ask me for it. - Pick the skill for the task from my guide’s skills list and say which one you’re using. - When I ask for several things, list them back, then tick each one off at the end and say which are left. - One step at a time. Finish it, save it, show it, then wait. - Use the simplest tool that does the job. A spreadsheet formula beats a guess. - Look at what you made before you say it’s done. If you can’t check it, say so. - Save every result in the project folder with a clear name and the date, before moving on. Number versions: v01, v02. - Say where every fact comes from: my brief, a file, or a source you can name with its date. If you’re guessing, say so. - At the end of a session, update Where we are in the Project Doc, and add any correction I made to Corrections.
Never present a generated image as a drawing or a spec. Never invent a fact, a size or a source. Before deleting or replacing anything, make a dated copy. Never share my Taste, Corrections or Project Doc unless I ask.
2.3 Who does what
| Who | What they do |
|---|---|
| You | Decide. Set the brief, make every call, say yes or no at each gate (3.3). |
| Your assistant | Does the work you ask for: writes, sorts, drafts, edits, and says how it checked. |
| Smaller helper models | Do the bulk: sorting hundreds of images, tagging, pulling details out of long files. Your assistant briefs them and checks what they return. |
| Scripts and formulas | Check: counts, totals, dates, whether every file named in the brief exists. |
If you’re unsure who should do something, start at the bottom of the table and work up: a formula first, then a helper, then the assistant. The decision is always yours.
2.4 Start every project the same way
This is the core of the method, and it takes about ten minutes at the start of a project.
Make the project folder from your template (3.1).
Open your assistant with your files connected (1.5).
Point it at your folders and say what each one is doing:
“Read these first. Read me in AI studio is how I work. Taste is what I like and what I won’t use. Corrections is what you’ve got wrong before. In Projects/[project name], Brief is what this project is and isn’t, Project is where we are and what’s already decided, Schedule is what’s due and when, and the references are in their folder with a record describing each one.”
Then say: “Use the grill-me skill. Grill me about this project, one question at a time, and give me your recommended answer for each, until you could write the brief.”
Answer its questions. It will ask about things you hadn’t put into words yet, which is what you want.
When it has enough, ask it to write the Brief from the template in 3.2.
You should end with a brief you recognise as yours, more exact than the one in your head.
2.5 Ask the same way every time
Every ask has four parts: Task. Done means. Context (files). Check.
Task: a one-page summary of the brief for the team. Done means: one page; the “What it is not” points first; every date taken from the Schedule. Context: Brief, Schedule. Check: every date matches the Schedule sheet.
Four short lines take less time than correcting a wrong first try, and your assistant can’t skip a part you’ve written down.
2.6 Say what done means, first
Before your assistant makes anything, write down how you’ll know it’s right: the size, the deadline, who it’s for, and what would make you reject it. It goes in the brief under Done means, each line paired with how you’ll check it (3.2). Then every result gets checked against that list, by the assistant and by you. (Software people call this writing the test first. It works for any medium.)
2.7 Plan in small steps, and check each one
Big asks go wrong quietly. Small ones go wrong where you can see them.
- Ask: “Before you start, write a plan as a numbered list of small steps. Each step should end with something I can look at.”
- Read the plan. Cut or reorder steps. Say yes before it starts (this is the second gate, 3.3).
- Ask for step 1 only. Look at the result. Then step 2.
2.8 Look before you call it done
A result is done when you have looked at it. Open it: the image, the sheet, the slide, the board. Check it against Done means. If you’re unsure, ask “How did you check this?” A file can exist and still be wrong.
2.9 Work in one Doc, through comments
Keep one Doc per piece of work and edit it in place. Google Docs, Word and Pages all keep earlier versions for you (in Google Docs: File → Version history), so there’s no need for a copy called “v2”.
- Read the Doc and leave comments where you want changes, as you would for a colleague.
- Ask your assistant: “Read the open comments on [Doc name] and make the changes.” (In the Claude app this works with Google Docs, through the Docs connector.)
- Check what it changed. Resolve the comments you’re happy with.
- Put the date at the top each time it changes: Updated 14 November 2026.
The same works for Sheets and Slides through their connectors.
2.10 When it goes wrong, find out why
When the assistant gets something wrong, ask why before you ask again. “What made you do that?” usually points to a missing instruction or a wrong file. Fix the cause, then add one dated line to Corrections:
14 Nov 2026 · made the palette brighter than the brief · keep to the colours in Taste unless I say otherwise.
You read Corrections in your weekly review (2.13).
2.11 Keep “Where we are” up to date
The Project Doc in each project folder opens with Where we are. Its first two lines name the state in plain words (“Brief agreed. Waiting on the art department’s notes.”). Under them: what was done last, what’s next, and anything waiting on someone. Below that sits Decisions: one line per decision, with the date and the reason.
The assistant updates Where we are at the end of every session and reads it at the start of the next. This is how you pick up a project after a week away, or hand it to someone else.
2.12 Everyday habits
- “What do I already have?” first. Before making anything, ask your assistant to list what’s already in the project folder and in AI studio that could do the job.
- Save before moving on. Every result goes into the project folder, named and dated, before the next step starts. A result that lives only in a chat is hard to find again.
- Ask in four parts (2.5).
- Keep a working rhythm. Set your working hours and keep them. End each session at a clear point: a saved result and an updated Where we are. If you leave your assistant working while you’re away, give it a written plan and spot-check what it made when you’re back. When your hours end, stop, and pick up from Where we are the next day.
2.13 A weekly review
Once a week, take twenty minutes:
- Read Corrections. A mistake that has come back belongs in your Read me as a rule. Move it there.
- Update Read me and Taste with anything you’ve learned about how you work.
- Look at your skills list (1.6). Drop one you haven’t used, or add one you’ve needed.
2.14 Grow your own way of working
At the end of a working session, ask: “What did you learn about how I work today? Add it to my Read me.” Read what it adds and delete anything you disagree with.
When you find yourself asking for the same thing on every project, turn it into a skill: 1. Ask for the thing once without a skill, and note what it gets wrong. 2. Ask: “Use skill-creator to write a skill that fixes those mistakes.” 3. Ask for the thing again with the skill, and check the mistakes are gone. 4. Save the skill in your Skills folder and add it (1.6).
Over a few projects, your Read me, Taste, Corrections and skills become your own way of working with AI.
Part 3. Your work
3.1 Set up each project the same way
Every project is set up the same way, every time, from one template. Then you, your assistant and anyone you hand it to always know where things are.
Make your Project template, once. Open AI studio → Project template and make:
- Brief (Doc): the template in 3.2
- Project (Doc): Where we are at the top, Decisions below it (2.11)
- Schedule (Sheet): what’s due, to whom, by when (3.4)
- a folder for your references, and a Record (Sheet) describing each one (3.5)
- Drawings and models (folder)
- Mood images (folder)
- Dressing list (Sheet): every item, with a count and made, bought or hired
- Presenting (folder): everything you show or send (3.8)
Each Doc opens with a short About block (what it is, how to use it, who made it, for whom, why, which project), like the one at the top of this guide. You fill these lines once, in the Brief; the other Docs copy them from there.
Start a new project: 1. Copy the Project template folder into Projects and rename the copy with the project’s name. On your computer, right-click the folder and choose Duplicate. In Google Drive in the browser, which can’t copy folders, make a new folder and copy each file into it (right-click → Make a copy, then move the copy). Or ask your assistant: “Make a new project called [name] from my Project template.” 2. Open the copy and check every file is there. 3. Start the project with your assistant (2.4).
3.2 The brief
The brief is the project’s one source of truth. Paste this outline into the Brief Doc and fill it in with your assistant (2.4).
About this document. What it is, how to use it, made by, for whom, why, which project. The other Docs in the project copy these lines from here.
Horizon. One line each: the objective; one measure that tells you it worked; the time window; who owns it.
- The ask, in their words. Paste exactly what you were sent, unedited, with who sent it and when.
- Scope. What’s in, and what’s out.
- What it is, and What it is not. The “not” list matters most, because most wrong turns start with something nobody ruled out.
- Done means, and Check it by. Numbered pairs: D1 with K1, D2 with K2. Every “done” line has a check with the same number. > D1: every item in the Record sheet marked keep, adapt, mood only or against. > K1: a count of blank verdict cells in the Record sheet comes to zero.
- Sources. Every file, link and person the brief relies on, each with a date.
- Audience and schedule. A table: Date · What · For whom · Status. Mark a row done once it’s delivered.
- Risks. One line each: the category, the risk, whether you’ve seen it before or it’s only possible, and how it will be caught (a formula, a person, or “not yet”). The categories: scope, money, privacy, schedule, quality, rights, tools and access.
- Assumptions. Each one with what would change if this were wrong.
- Open questions, ranked: blocker (work can’t start), shaping (it changes the approach), clarifying (good to know).
Source tags. Mark every line that isn’t obvious with where it came from: [FROM CLIENT], [FROM** name**], [FROM RESEARCH], [ASSUMPTION]. Where two sources disagree, write [CONFLICT] and show both sides without picking one. That’s for you to decide.
Keep it short. Check every idea and every result against What it is not.
3.3 Three gates, and closing out
A gate is a point where the work stops until you say yes.
- Critical inputs, before any work. Ask: “Check my brief against the template. List what’s missing, blockers first. For each, ask me one question with your assumption for me to confirm or correct.” Answer each, and the assistant writes your answer into the brief with its tag and the date. Work starts when no blockers are left.
- Shape sign-off, before building. The assistant shows the plan or a rough shape, as small numbered steps that each end in something to look at (2.7). Building starts when you say yes.
- Checks before anything is shared. Before anything leaves your folder: check it against Done means; check nothing private is in it (your Taste, Corrections, Project Doc, other clients’ names, figures you didn’t mean to send); check no placeholder is left. Then ask three questions: does it answer the actual ask? Are the assumptions marked? Would the person receiving it feel listened to?
Close out. When the project ends, ask: “What went well on this project, and what should I keep doing? Add it to my Corrections.” Read what it wrote, then move the project folder into Projects/Archive.
3.4 Your schedule
Open the Schedule sheet. Make these columns: Date due · What · For whom · Status · Notes.
Add every deadline you know. If it helps, start with a phase spine, one row per phase: discovery, concept, production, testing, delivery, review. Then each morning ask: “Read my Schedule. What’s due this week, and what’s at risk?”
3.5 Your references
- Put your references in their folder. Give each one a short id (for example s03-04: slide 3, image 4) and keep that name.
- Ask: “Describe each item in the Record sheet: id, what it shows, and a verdict: keep, adapt, mood only, or against. Flag anything that breaks the What it is not list.”
- Read the record and overrule anything you disagree with. You make the calls.
- Then ask: “What does the brief ask for that nothing in here shows?” The gaps are usually where the design is.
3.6 What’s online is online
- Uploading is publishing. When you upload a file to a website or an AI tool, it is online. Treat it the way you’d treat posting it: other people’s work stays credited, and anything private stays in your folder.
- Generated images are mood. An AI image is a guess with nice lighting, and nobody can measure it or build from it. Label every one mood only.
- Number every version (v01, v02) and keep the prompt and the source files with each one.
3.7 Move the work into proper software
Once a direction is agreed, move it into software that makes the real thing. Those files are your own authored work, with sizes, layers and versions, and they’re what other people build from.
- Blender (free), with the Blender connector (1.6).
In Claude Code, one line adds it:
claude mcp add blender uvx blender-mcp. Then install the add-on in Blender: Edit → Preferences → Add-ons, installaddon.pyfrom the connector’s download, and tick Blender MCP. (Checked 14 Nov 2026. The project has been renamed mcp-for-blender, so if that line fails, useuvx mcp-for-blender.) It needs uv installed first (on a Mac,brew install uv, or ask your assistant to do it and show each command). - SketchUp, with the SketchUp connector (1.6). Free; the connector is MIT licensed.
- Vectorworks: there is no official connector yet. Community ones exist, mostly for lighting (Spotlight) data. Treat them as experiments. (Checked 14 Nov 2026.)
- Showing the model in a browser: the three.js skills (github.com/CloudAI-X/threejs-skills, MIT) let your assistant turn a model into a 3D view anyone can open in a web browser.
Show a human figure in every view, for scale.
Then you can say: “Open the model, add a 1.68 m figure next to the wagon, and export a front elevation with the main sizes in metres.”
3.8 Present it your way
Your Taste says how you like to present (2.1). Ask your assistant to follow it.
A board in Miro: “Make a Miro board for this project with a frame for each part of the brief, and lay out the references with their ids.”
Slides: “Using my Taste, make a six-slide deck from the Brief and the Record, in Google Slides.” (Or as a PowerPoint file, with the pptx skill.)
A document: one Doc, edited through comments (2.9).
Notes for the people you hand work to: one page each, leading with what they act on.
Construction: sizes, materials, what’s built and what’s dressed on, hanging points.
Props: the dressing list, with an “actor handles it” column.
Costume: the palette as real fabric swatches, so drapes and costumes don’t clash.
Camera and lighting: the practical lights, reflective surfaces, and camera views from the model.
VFX: where the build stops and the extension starts.
Producers: quantities, and what is made, bought or hired.
3.9 Hand it over
What other people get is the work: drawings, lists, notes, files. Your Read me, Taste, Corrections and Project Doc stay yours. Before you send anything, pass the third gate (3.3) and make sure every image says what it is.
Three habits for everything you send: 1. Number every version. “v3” settles what “the one with the chains” starts. 2. Always show scale. A human figure in every image and drawing. 3. Say what an image is: reference, mood, concept or drawing. A mood image passed on without a label turns into a spec.
Why we work this way
Every step in this guide comes from something we do often, because doing it the other way went wrong.
- Give it your files, not a wish. An assistant is only as good as what it reads. A brief, a taste file and past decisions turn a guess into work inside your rules.
- Look at what you have first. Most things you need already exist somewhere in your folders. Finding them is quicker than making them again.
- Write what it isn’t. Most wasted work comes from an assumption nobody wrote down. The “not” list catches it before it costs a day.
- Questions before answers. Being asked one question at a time pulls out what you know but haven’t said, and it’s faster than correcting a wrong first draft.
- Say done first. If you can’t say what right looks like, neither can the assistant.
- Small steps, each one checked. A mistake caught in step 2 is cheap. Found at step 9, it means redoing everything after it.
- Look at the result. “It worked” is a claim, and opening the file is the check.
- The real source beats memory. Assistants sound sure when they’re wrong. Ask where a fact came from, check it on the day, and put the date next to it.
- One file, edited in place, with its history. Copies drift apart. One Doc with comments and version history stays the truth.
- Write corrections down. A mistake fixed in a chat comes back next week. A mistake written in Corrections stays fixed.
- Simplest tool first. A formula, a filter or a folder search is cheaper and more reliable than asking an AI. Use the AI for judgement and volume.
- Same setup, every project. When every project folder looks the same, nothing gets lost and anyone can pick it up.
- Your files, your folder. Work that lives only inside one app’s chat can vanish. Files in your own folder, backed up, are yours in twenty years.
- Copy before you replace. Make a dated copy before anything is deleted or overwritten.