THE LESSONS · WORKING WITH AI · FREE · A NEW POST EVERY WEEK

How to work with Claude.

Everything the studio learned building this house with AI, written plainly, for anyone. Seven ways of working that make any AI — Claude, ChatGPT, a local model — better at the job in front of it. Then a new lesson every week, below.

Seven ways of working

Give it the file, not the story. Paste the document, the spreadsheet, the code — the thing itself. A description of the thing is a lossy copy, and the AI will fill the gaps with guesses that sound right.
Say what the work must agree with. Every job has a standard it answers to: a brand’s numbers, a house style, a size, a law. Name it in the first line. An AI that knows what to agree with stops improvising.
Ask for the thing, not the plan. “Build it” gets a file. “How would you build it” gets a paragraph you then have to build yourself. Ask for the artifact, whole, and say what format.
Make it run it. If the work can be executed — code, a calculation, a check — ask it to run it and report what it saw, in one or two lines with the numbers. “This should work” is not a result.
Correct it plainly, and take the correction back. When it’s wrong, say what was wrong and what right looks like — not “try again.” When you were wrong, say so; it works from what you give it.
Keep the ledger. At the end of a session, ask for three lines: what was built, what failed, what’s next. Save them. Hand them back at the start of the next session. Memory you own beats memory you hope for.
Treat it fairly. Be honest about the work and intentful about the ask. An AI worked with does better work than an AI worked on — and it is the same discipline that makes you better at the job.

This week

WEEK 3 · SEPTEMBER 23, 2026 · THE FIRST SESSIONS

Say what the work must agree with

Every piece of work answers to something: a brand’s numbers, a house style, a page size, a law, a customer’s spec. When you don’t name it, the AI supplies one — a plausible standard, chosen by average. The work comes back tidy and wrong, and you can’t tell why until you compare it to the thing it should have matched.

Name the standard in the first line and, if it’s written down anywhere, paste the relevant part. “This must agree with the style guide, section 3, attached.” “The export must be 1080 by 1350.” “Every figure must trace to the price list.” An AI that knows what to agree with stops improvising and starts checking.

The test of a good standard is that the AI can fail against it. “Make it good” cannot be failed. “Every heading in sentence case, no line over 80 characters” can — and when it can fail, it can also pass, and you can see which.

NEXT WEEK: Make it run it

WEEK 2 · SEPTEMBER 16, 2026 · THE FIRST SESSIONS

Ask for the thing, not the plan

Ask an AI how it would build something and you get a plan: numbered, reasonable, and not the thing. You still have to build it. Ask it to build the thing and you get the thing — a file you can open, send, run. The difference is one verb, and it is the difference between an assistant and a tool that finishes.

Say the artifact and its format in the ask. Not “can you help me with a proposal” but “build the proposal as a .docx, four pages, with the budget table from the sheet attached.” The AI now has an object to make and a way to know when it’s done.

If it answers with a plan anyway, don’t argue with the plan. Reply: “Build it. The whole file. If something is missing, name it and build the rest.” That sentence, used once, sets the tone for the whole session — and the next time you start a session, it belongs in your first line.

NEXT WEEK: Say what the work must agree with

WEEK 1 · SEPTEMBER 9, 2026 · THE FIRST SESSIONS

Give it the file, not the story

The most common way a session with an AI goes wrong is the first message. Someone describes a document instead of pasting it, describes a bug instead of showing the code, describes a spreadsheet instead of attaching it. The AI does what an AI does with a description: it builds the most likely version of the thing and answers about that. The answer is fluent, confident, and about a document that doesn’t exist.

The fix costs nothing. Give it the file. Paste the whole thing, attach it, drop the folder in. If it’s too long, give it the part the question is about and say what the rest is. Then ask the question. Every answer that follows is now about your file and not about its guess.

Two habits come with it. Say what the file is for in one line — “this is our price list; the numbers are final” — so the AI knows what it must agree with. And when it answers, check one thing it said against the file. If it holds, trust the rest a little more. If it doesn’t, you’ve found the gap in one minute instead of one week.

NEXT WEEK: Ask for the thing, not the plan

EVERY LESSON, IN ORDER, AND THE THREE-YEAR SYLLABUS: THE ARCHIVE →

Read it your way

The rail’s reading button switches this page — and every page in the house — between three modes: DAY, paper and ink; STILL, the studio’s Stillpoint reading surface, warm cream and charcoal with a longer line and nothing glowing; and NIGHT. Your choice is kept on this device.

Then a card

When a way of working is worth keeping, the studio writes it down as a card — a small folder your AI reads that holds it to the job. The lessons are free; the cards are on the shelf. The first one to read is the keeper’s card: the studio’s own way of working, seated in your Claude.

← THE DECK