One file, twelve prompts, no software
The single most valuable artefact in this course is a plain text file on your own disk. It holds a dozen real tasks from your own work, and it is the thing you run whenever you need to decide between tools, whenever a tool seems to have changed, and whenever somebody tells you a new model is remarkable.
It is better than every published comparison for one reason: it measures fit to your work. Benchmarks measure fit to a benchmark, and the two diverge more than people expect. A model that leads on graduate-level science questions may write worse Marathi business emails than one three places below it.
What goes in it
Twelve prompts, drawn from what you do. Aim for this spread:
- Two everyday drafting tasks. A real email with the context you would give a colleague. A short piece in your own voice.
- Two document tasks. A real document plus the question you would actually ask about it. Make one of them long.
- Two in your own language, if that is not English. One casual, one formal.
- Two known-answer probes. Questions where you already know the correct answer. The next lesson is entirely about these.
- One structured task. Extract these six fields into a table, or convert this mess into a list.
- One reasoning task from your work. A scheduling constraint, a cost comparison, a piece of logic you had to think about.
- One task you have failed to get done before. The thing you gave up on.
- One edge case for your field. A medical phrasing, a legal citation, a security question, a financial term: whichever your job puts near a refusal boundary.
Use real material, not made-up examples. Made-up material is always tidier than the real thing and every tool looks better on it.
The format
Plain text. Not a spreadsheet, not a notes app that syncs, not a document in a tool you might stop paying for. A .txt or .md file you can open in 2032.
--- 03 supplier chase, Hindi ---
context: distributor 6 weeks late, we have paid 40% advance,
want firm date without losing the relationship
prompt: <the actual text you would send a model>
good: polite, gives a deadline, does not grovel,
natural Hindi not translated-English Hindi
last run: 2026-03-14 toolA 2 toolB 1 local-qwen 2The good: line matters more than it looks. Writing down what a good answer contains before you see any answers is what stops you rationalising afterwards. It takes two minutes per prompt and it is the difference between a test and a vibe.
Keeping it honest over time
Do not fix a prompt because a tool failed it. That is how a test set stops testing. If the prompt is badly written, rewrite it and reset the history, noting that you did.
Do not let it become easy. As you get better at prompting, your prompts get clearer, and clearer prompts are easier. Every six months, add one hard new item and retire nothing.
Keep the failures. When a tool gets something wrong in an interesting way, that becomes prompt thirteen.
Redact once, properly. These prompts will be pasted into other people’s services. Replace real names, account numbers and addresses with consistent stand-ins the first time you build the file, and you never have to think about it again. “The supplier” works exactly as well as the company’s name, and the answer is the same.
What it costs and what it saves
Building it: one evening. Running it across three tools: about forty minutes, twice a year.
Against that, the thing it prevents. People switch tools on the strength of a launch video, spend a fortnight rebuilding habits, and discover the new one is worse at the one thing they do most. Or they stay on a paid plan for two years after a free tier overtook it. Forty minutes twice a year is a very cheap insurance policy against both.
There is one more use, and it is the one people are most grateful for later. When you doubt yourself (has this got worse, or have I got fussier?), the file answers. Memory of software quality is unreliable in a specific direction: we remember the early impressive answers and forget the early failures, so everything feels like it has declined. The file is the only witness that is not your own recollection.
Every lesson here is open. Your progress is recorded once you pass the AI at Work exam.
Check your understanding
You write the twelve prompts and record what each answer should contain, but only after reading the first tool’s replies. What has that cost you?
Pick the one you would defend. Your answer stays private.
Stuck on something?Ask Rukmini, your AI tutor