Notes on working with AI
Ask your AI to write like an aircraft manual
One line in your prompt turns rambling AI answers into steps a new hire can follow. The idea comes from the manuals that keep aircraft flying.
Ask any AI how to reconcile a bank statement and you get something like this:
Reconciling a bank statement essentially involves comparing the transactions recorded in your books against those that appear on the statement provided by your bank, while taking into consideration that there may be timing differences, such as cheques that have been issued but not yet cleared, which should generally be accounted for, after which any remaining discrepancies would typically need to be investigated and resolved.
It is correct. It is also one sentence, 65 words long, and you will read it twice. Now add one line to the prompt and ask again:
- Open the bank statement and your ledger for the same month.
- Tick every ledger entry that also appears on the statement.
- List the entries with no tick.
- Check each one. A cheque you issued that the bank has not paid yet is a timing difference. Keep it on the list for next month.
- An entry on the statement that is not in your ledger is a missing record. Add it to the ledger.
- Compare the two balances. They must match. If they do not, go back to step 3.
Same facts. A new bookkeeper can follow the second version on day one. The line I added was a request to write in ASD-STE100.
Where this comes from
ASD-STE100 is Simplified Technical English, an international standard for technical documentation. It is maintained by the STEMG, a group within ASD, and the current version is Issue 9 from January 2025. The official site is asd-ste100.org. It was built for aircraft maintenance manuals, where a mechanic who misreads a sentence at 3 a.m. creates a real problem, and where many readers are not native English speakers.
The standard has about 53 writing rules and a dictionary of about 900 approved words. You do not need to learn them. The core fits in one breath: one meaning per word, short sentences, active voice, one instruction per sentence.
Why a model does this so well
Tell a model "be clear" and it has nothing to hold on to. Clear to whom? Compared to what? It guesses, and the guess is usually longer.
Tell it "one instruction per sentence" and it knows exactly what to do, and so do you, because you can check the result. Models follow concrete, checkable rules far better than they follow a mood. When it slips, you point at the sentence and the rule it broke.
The rules also fix real reader problems. One meaning per word means a beginner never wonders whether "statement", "report" and "record" are three things or one. Short sentences mean an expert can skim. Active voice means someone is actually doing something.
Follow it about 80%
Strict compliance has a cost. An explanation of why something works needs a "because", and a rigid version drops it. It also repeats the same noun until the text sounds like a parts list.
So I ask for about 80%. This is the prompt I use:
Write in the style of ASD-STE100 Simplified Technical English, followed about 80%. One meaning per word, and use the same term for the same thing. Short sentences. Active voice. One instruction per sentence. Use numbered steps for any procedure. If a rule makes the explanation unclear, choose clarity.
The last sentence tells the model which side wins when the rules and the reader disagree. The model works from what it remembers of the standard, so you get a close imitation, not certified output. For this purpose that is enough.
Three places I use it
Onboarding documents. Paste your messy notes on how the team handles refunds and ask for a procedure. A new employee gets numbered steps instead of a paragraph of tribal knowledge.
A procedure for one person. Month-end close, payroll cut-off, how to send an invoice to a customer who needs a purchase order number. Anything you currently explain out loud, again and again.
A contract clause for a client. Paste the clause and ask what the client must do, by when, and what happens if they do not. You get three plain sentences you can review before they go anywhere.
Make the AI show, not tell
Writing is one fix. Three more follow the same logic: stop making the reader do the work.
Ask for a diagram instead of a description. "Draw how an invoice moves from request to payment" gets you a flow you take in at a glance.
Ask for an HTML page. Models are good at front-end code, so a request for a loan calculator with a slider for the rate and one for the term gives you a working page that recalculates as you drag. Your client plays with the numbers instead of reading about them.
And a custom explainer video is becoming practical, with a script from the model and a voice from a text-to-speech tool. I'd treat that one as worth watching rather than a routine yet.
What to do this week
Pick the next explanation you were about to write by hand. Paste the prompt above, add your material, and compare the result with what you would have sent. Then check three sentences against the four rules. If one fails, name the rule and ask again.