You will be able to spot a task that is too big for one request and say why.
You can write a careful prompt, with full context, a clear task, a length limit and a format, and still get back something thin. A well-written prompt can still fail if it asks for too much at once. Many big requests hold several separate jobs, such as research, analysis, weighing options and persuasive writing. When one request asks for research, analysis and a polished document together, each of those parts gets shallow treatment.
Aisha is a project coordinator at a mid-sized accounting firm in the CBD. Her director asks for a proposal on moving the firm's annual staff retreat from a Bintan resort to somewhere in Singapore. It has to cover the options, costs, pros and cons, and a recommendation.
Aisha applies what she learnt in module 1. She writes one careful prompt with full context, the task, a length limit and a format, and asks for everything in a single go. She gets back four tidy pages that look like a proposal but say very little. The options are generic and the costs are vague ranges. The recommendation reads as if it was decided before the comparison was written.
Her prompt was well written. The problem was that it asked for too much at once. Inside it were at least four separate jobs: finding suitable venues, costing each one for about 120 people, weighing them against what the firm cares about, and writing a persuasive case for a director.
The model writes its answer from start to finish in one pass, so its effort is spread across everything you asked for. The polish comes out fine, because polish is what models do most easily. Research and analysis need the most thought, and they get the least. The polish also makes the thin parts harder to notice, which is how a weak proposal can still look finished.
There is a second problem. As covered in AI fundamentals lesson 3.2, "It writes by predicting the next token, again and again", the model writes each part of an answer with everything before it in view, and it does not revise anything it has already written. If an early part of the answer is wrong, everything after it is built on that error.
Early in Aisha's draft, the model assumed the retreat was a single day. It was meant to be two days and a night. Because of that, every cost was for a day event, the shortlist leaned towards venues without accommodation, and the recommendation favoured the cheapest day option. Aisha only spotted the error on her second read, by which point that one assumption near the top had shaped all four pages. In a single giant request you can't catch an early error until the end, and fixing it then means starting again.
A more useful approach is to hand the work to an assistant the way you would hand it to a capable junior colleague. You wouldn't ask a new colleague for the finished proposal on day one and then wait. You would ask for a shortlist of venues first and review it, crossing off the two places the director dislikes. Then you would ask for costs on the remaining venues and check them. Only after that would you start the write-up.
Each step should be small enough to review in a few minutes. Working this way gives you a checkpoint between steps, a point where you look at the output before the next step builds on it. You see the shortlist before any costs are worked out, and you see the costs before the argument is built. When something is wrong, you fix it at the step where it happened, and everything after that starts from the corrected version. If Aisha had worked in steps, the single-day assumption would have shown up in the shortlist or the costs, long before any persuasive writing depended on it.
Splitting a job takes more messages, but it rarely takes more time. You stop spending an hour editing a long draft that was built on a wrong assumption.
Before you write a request, use the one-sentence test: try to describe in one sentence what a good output looks like. For a 120-word reply to a customer email, you can usually do that easily, so the task is the right size for one request. If describing a good output takes a paragraph, the task is really several tasks, and each one deserves its own request.
A request can look like one sentence and still fail the test. "Write a proposal on moving our retreat to Singapore" is short, but saying what a good version looks like takes several sentences about venues, costs, priorities and the case for the director. That length tells you the request holds several jobs.
Alongside the test, look for these warning signs that a job should be split:
the prompt strings four or five different verbs together with "and" the prompt needs both research and persuasive writing you expect the output to run to several pages
If you see any of them, split the task into separate requests.
The hard part is often noticing the steps at all. When you do a big task yourself, the steps blur together because you move between them without thinking. Writing down what you actually did the last time you did the task makes them visible.
When Aisha wrote down what she did for the proposal the previous year, she found seven steps, and only two of them were writing. She gathered requirements from the director, shortlisted venues, called for quotes, built a comparison, chose a venue, wrote the proposal, and revised it after comments. Each of those is a piece she can review on its own before the next one starts.
Your own list of steps is the starting point for the rest of this module, which shows how to work in steps well. Lesson 3.2, "Ask for the plan before the work", starts with the most useful single step for working in stages.
So think back over the last few weeks. Somewhere in there is a piece of work that took you more than two hours from start to finish, the kind you might hand to an assistant next time round. Write down the steps you actually took to do it, in order, and mark which of them were writing.
Pick a task from the last few weeks that took you more than two hours. Write down the separate steps it really involved.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).