Unit 1 · From chatbot to builder
Stop asking. Start assigning.
The principle
A useful instruction names the job, the inputs, the output shape, and the standard — a question names none of them.
By the end: You can rewrite a question as a job description and get a usable result on the first try.
The situation
It is Monday. You have the weekly numbers in a spreadsheet, and a status update due at eleven. You paste the sheet into an AI tool and type: "Can you analyse this and tell me what stands out?"
What comes back is fluent and useless. Six bullet points that restate your own columns. One observation that is technically true and operationally irrelevant. A closing line offering to go deeper. You skim it, close the tab, and write the update yourself — twenty minutes later than if you had never opened the tool.
Nothing malfunctioned. You asked a question, and you got an answer. The problem is that you did not want an answer. You wanted a piece of work.
The principle
A useful instruction names the job, the inputs, the output shape, and the standard — a question names none of them.
This is the difference between talking to a colleague in a corridor and briefing one properly. In the corridor you can be vague, because they already know the account, the format the update takes, and what your boss cares about. A model knows none of that. Every unstated expectation is a gap it fills with the statistical average of everything it has read, which is exactly why generic output feels generic.
Four lines close the gap. The job — what work is being done, as a verb. The inputs — what it is working from, and what to ignore. The output shape — the literal structure you want back. The standard — what makes it good, and what makes it wrong.
Worked example
Shown here in ChatGPT.
The version that produced the useless answer:
Can you analyse this and tell me what stands out?
[pasted spreadsheet]The same task, assigned:
Job: Write my Monday status update from the attached weekly numbers.
Inputs: The sheet has one row per region and columns for booked, pipeline
and churn. Compare this week to the four-week average in the same sheet.
Ignore the "notes" column entirely.
Output: Exactly three bullets. Each bullet is one sentence naming the
region, the metric, and the size of the move. Then one line headed
"Needs a decision" if any region moved more than 15% in either direction.
Standard: Only flag moves you can point at a number for. If nothing moved
more than 15%, say so and write nothing under "Needs a decision".The second version is longer to write once, and then never written again — it becomes the instruction you reuse every Monday. That reuse is the entire point, and it is what Unit 2 builds on.
Note what the standard is doing. "Only flag moves you can point at a number for" is the line that stops invention. "If nothing moved, say so" is the line that gives the model permission to return a boring, correct answer instead of manufacturing an insight to be helpful.
In your tool
Product names move faster than the structure does. What you are looking for is the place a long instruction can live so you do not retype it.
| Tool | Where a reusable instruction lives |
|---|---|
| Claude | Project instructions, set once per Project and applied to every conversation in it |
| ChatGPT | Custom instructions, or per-project instructions inside a Project |
| Gemini | A Gem — a saved, named configuration you start conversations from |
| Microsoft 365 Copilot | A published agent carries the standing instruction, but in most tenants a maker or an administrator publishes it rather than you |
Product surfaces checked 2026-08-04.
Try it
Take a task you actually did last week and did not enjoy. Write the four lines for it — job, inputs, output, standard — in under five minutes. Do not open a tool yet.
You are done when the output line is specific enough that two different people would produce a recognisably similar document from it. If your output line says "a summary", it is not there yet. If it says "exactly three bullets, one sentence each, region named first", it is.
Common failures
- The polite question. "Could you help me look at this?" reads as courteous and lands as underspecified. Politeness costs you nothing with a model and buys you nothing either — the words are spent on tone instead of on the brief.
- The wall of context with no output shape. People over-correct by pasting three paragraphs of background and still never saying what they want back. Background without a target produces a well-informed essay you did not ask for.
- Asking for insights. "Insight" is not a format. It cannot be checked, so it cannot be got wrong, so you get something that sounds like insight. Ask for a comparison, a ranked list, or a flagged exception instead — all three can be verified against the data in front of you.