You are not going to write code, and you still want to be genuinely good at this.
Reviewed
A couple of weeks of deliberate use
By the end you will have a small library of prompts you reuse, a clear sense of which tasks to trust a model with, and rules about what never goes near one.
The most capable AI users in most organisations are not engineers. They are the people who understood earliest which parts of their job are language problems, and who learned to brief a model as carefully as they would brief a new hire.
None of what follows requires code. All of it requires being specific, which is the harder skill.
A task is worth handing to a model when it is frequent, tedious, and tolerant of a first draft being imperfect. Frequency is what makes the setup pay back. Tedium is what makes you actually adopt it. Tolerance is the safety check — if a mistake is expensive and hard to spot, the task needs a person at the end regardless of how good the draft is.
Score your recurring work on those three and start at the top. Almost everyone's first genuinely useful automation is something dull they had stopped noticing.
The model knows the language, not your context. Say who the output is for, what it is for, what to include, what to leave out, how long it should be, and what a good one looks like — the same briefing you would give a competent freelancer on their first day, because that is exactly the situation.
Then show, do not only tell. One or two examples of the output you want will do more than three paragraphs describing it. If you are producing the same kind of thing repeatedly, your best previous output is your best example.
Models are least reliable exactly where they sound most authoritative: specific numbers, named sources, quotations, dates, legal and medical specifics, anything about a small or recent organisation. The output will read identically whether it is right or invented, so tone tells you nothing.
The practical rule: never accept a fact you would have had to look up. Ask for the source and check it, or paste in the document and ask it to answer only from that. Verification is a habit, not a step you do when something feels off — by design, nothing will feel off.
Every time a prompt works well, save it somewhere shared with a one-line note on what it is for. Within a month you will have a dozen; within a quarter your team will be using yours instead of writing their own badly. This is the highest-leverage thing a non-technical person can do with AI in an organisation, and almost nobody does it.
Review the library occasionally and delete what you no longer use. A library of forty prompts nobody trusts is worse than eight that everyone does.
Write down, before you need it, what you will not paste into a model: customer personal data, credentials, unreleased financials, anything covered by an agreement you have not read. Check whether the tools your team uses train on submitted content and what their retention period is — the answers differ sharply between consumer and business tiers of the same product.
This takes an afternoon and is the difference between AI adoption that survives its first security review and one that gets banned outright.