10. Building the Setup You Keep
Three layers, four prompts worth saving, and the one habit that separates people who get value from this from people who try it once and drift away.
The WJS Desk
Sep 7, 2026 · 7 min read
Nine lessons of technique are worth very little if you have to remember them. This one is about building the setup that means you do not have to, and about the honest question of what all this is actually worth.
Where the course got you
Worth naming, because the shape matters. You can put context in a prompt so it stops guessing. You can separate your instructions from your material so your prompts become reusable. You can control the shape of the output instead of reformatting by hand. You can split a job when one prompt would flatten the thinking. You can work through a document you were never going to read. And you can tell a real improvement from noise.
What ties those together is not cleverness. Every one of them is a way of getting something out of your head and onto the page.
The three layers
Set these up once, in this order, because each is cheaper than the one before and covers more cases.
Layer one, standing instructions. Your settings, applying to every conversation. Roughly 150 words of concrete rules:
I run a 6 person design studio. Clients are mostly US
small businesses.
How to write: plain and direct, US spelling, no exclamation
marks, never open with "I hope this finds you well". Emails
under 150 words unless I say otherwise.
How to behave: if a question depends on facts you do not
have, ask me rather than assuming. Flag anything in your
answer I should verify.Those last two lines are worth more than the rest combined. They change the default from confident guessing to asking, which is the single biggest quality improvement available from a settings box.
Layer two, a project per recurring job. A container holding reference documents, three examples of your own output, your vocabulary, and instructions specific to that work. Set one up for anything you do more than twice a month. The payoff is that lesson 7 stops requiring a paste.
Layer three, the prompt file. A plain note. Every time something works, paste it in with a three word label. This costs nothing, needs no features, and is the layer we would keep if we had to drop the other two.
What actually goes in the file
After this course, four earn their place immediately:
DOCUMENT QUESTIONS
Answer only from the text below. Quote the sentence each
answer comes from and cite its line or section. Say "not
stated in this text" rather than inferring.
Then: what costs me money and when, what are my
obligations with deadlines, what is unusual, what is NOT
covered that I might assume is.
HARD THING TO WRITE
Do not write it yet. Tell me what is actually at stake,
what they need to hear, and what I should avoid saying.
Be blunt.
ADVERSARIAL READ
You are the person receiving this, looking for a reason to
say no. What would you push back on, what is unclear, and
what is missing?
CHECK IT
[fresh chat] Find the errors in the following. Show the
calculation behind any figure, and state anything that was
assumed rather than given.Four prompts, one note, and most routine work stops starting from an empty box.
The habit that decides whether any of this sticks
Here is the thing we would tell someone finishing this course if we only got one sentence.
The people who get value from this are not the ones with clever prompts. They are the ones who send a second message.
Almost every good result in this entire course came from a follow-up. The role in lesson 1 was a second attempt. The chained email in lesson 6 was the second half of a conversation. The error we caught in the Super Beginner course was found by handing the text back in a fresh chat.
People who try AI once and drift away almost all did the same thing: one prompt, read the output, judged the tool. It is a three round conversation and the first round is a draft.
When to start a new conversation
This is part of the setup even though it is not a file anywhere, and getting it wrong quietly undermines everything above.
| Keep going | Start fresh |
|---|---|
| Refining a draft | It has misunderstood twice |
| Follow-ups on the same document | You want an unbiased critique |
| Building on an answer you liked | Switching to an unrelated task |
| Anything where earlier context helps | The chat has run long and gone vague |
The second row is the one people get wrong most, and lesson 8 explains why: a wrong answer sitting in the context pulls everything after it. Correcting inside that conversation fights the pull. Pasting the same text into an empty chat does not.
The maintenance nobody mentions
A setup rots. Two rules keep it from turning into a liability.
Delete anything out of date. A project quietly feeding it last year's prices is worse than no project, because you will stop checking. Anything with a number or a date in it gets a glance every few months.
Keep the instructions short. Ten specific rules beat forty, and a forty rule block is usually ten real rules and thirty restatements. When a rule keeps getting ignored, the fix is usually to delete four other rules rather than to add emphasis to that one.
Being honest about what this is worth
A course like this has an obligation not to oversell, so here is our position after testing everything in it.
It is very good at language, structure and patience. It will take unlimited follow-up questions at 11pm without getting bored of you, and that alone beats every other resource for learning something.
It is unreliable about specifics, in a way that does not announce itself. We found a mortgage answer wrong by a factor of 12 while writing lesson 5 of the beginner course's predecessor, and the wrong sentence read exactly like the right ones.
And a fair amount of received prompting wisdom did not reproduce when we tested it. Step by step did not change two answers. Tags did not stop an injection that was never going to land. Position did not affect retrieval across 260 lines. We reported those as we found them, including the ones that made our own lessons less dramatic.
What survived testing was the unglamorous half: supply real context, control the output shape, split jobs that contain a decision, and check the numbers.
What to do in your first week
The setup above takes about twenty minutes and then does nothing unless you use it. A realistic first week:
- Day one. Write the standing instructions. Two blocks, how to write and how to behave. Fifteen minutes.
- Day two. Open a note called prompts and paste in the four above. Five minutes.
- Rest of the week. Every time something works, paste it into the note with a label. Every time something disappoints, send a second message instead of giving up.
That is the whole onboarding. The projects layer can wait until you notice yourself pasting the same document twice, which is the signal that it would have paid for itself.
A capstone worth doing
Pick one thing you do every week that involves writing, reading or explaining. Then build it properly, using the whole course:
- Run attempt zero and keep it. You need the baseline.
- Answer the five questions from lesson 5, especially the failure you can predict.
- Build the prompt: role, task, constraints with reasons, the failure named, material in tags.
- If it contains a decision, split it per lesson 6.
- Write down one scoreable criterion, run it five times, tally it per lesson 9.
- Save the result in your prompt file with a label.
That is perhaps forty minutes. At the end you have one job that is permanently easier, a template you will reuse, and the method to do it again for the next one.
The one thing to keep. If you remember nothing else from ten lessons: give it the actual context, and check the numbers. Everything else in this course is refinement on those two.
What we would not claim
This course will not make you an expert on AI and it does not need to. You cannot tell from any of this how the model works internally, whether it will be reliable on your particular niche, or what it will be able to do in a year.
What you should have is narrower and more durable: a working method for getting a specific job done, and a calibrated sense of when to trust the result. Those two survive model updates. Specific tricks do not, which is exactly why lesson 9 taught you to measure rather than to memorise.
Where to go from here
Stop reading tutorials, ours included. Pick the three things in your week that are writing, reading or explaining, and do those with it for a month. That will teach you more than another ten lessons.
When you find yourself wanting to know what a task costs, whether a prompt change really helped across a hundred runs rather than five, or how to connect this to your own tools, that is the Intermediate track. It picks up exactly here.
