The AI Learning Hub Journal

The Task as an Outcome

Name the result you want, not the activity"analyse this" could mean twenty things — an outcome is something you can hold the answer againstTHE ACTIVITY — IT PICKS A MEANING"Analyse this customer feedback"twenty possible readings — the model picksone, and you cannot tell whether theanswer did the jobthere is no way to check itTHE OUTCOME — YOU CAN CHECK IT"Find the three issues most likely to bedriving cancellations, with mention countsand one representative quote for each"the answer can be held against the askcannot describe the outcome? that is useful information — you have not decided what you want yetSAY WHAT HAPPENS NEXT — THE DESTINATION SHAPES THE ANSWERa slide — six short bulletsa report — full sentencesread aloud — spoken rhythma rough draft — coverage over polishone sentence about the destination often beats a whole paragraph about styleONE TASK, NOT FIVE — CHAIN INSTEADget the summary rightfeed it into the analysisthen ask for the recommendationsmore than three deliverables in one message? split it — compound prompts produce compound mediocrityAN OUTCOME TELLS IT WHAT TO MAKE — AND TELLS YOU HOW TO CHECK ITstate the result, name the destination, and keep it to one job per message
State the outcome you want and where it goes next — an outcome can be checked, an activity can only be attempted.

Describe the Result, Not the Activity

There is a real difference between naming an activity and naming an outcome. "Analyse this customer feedback" is an activity — it could mean twenty things and the model will pick one. "Read these forty pieces of customer feedback and tell me the three issues most likely to be causing cancellations, with the number of mentions for each and one representative quote" is an outcome. You can tell whether you got it. Outcome-shaped prompts do two things at once: they tell the model what to produce, and they tell you how to check it. If you find yourself unable to describe the outcome, that is genuinely useful information — it usually means you have not decided what you want yet, and no amount of prompting technique substitutes for that decision.

  • Weak: "Analyse this customer feedback"
  • Strong: "Find the three issues most likely to be driving cancellations, with mention counts and one quote each"
  • An outcome is something you can hold the answer against; an activity is not
  • If you cannot state the outcome, the prompt is not the problem — the thinking is

Say What Happens Next

A small addition with a big effect: tell the model what the output is for downstream. "This will be pasted into a slide, so it needs to fit six bullets of under ten words each." "This goes into a formal report, so full sentences and no contractions." "I am going to read this aloud in a meeting, so write it to be spoken." "This is a first draft I will heavily edit, so prioritise coverage over polish." Each of these changes the shape of a good answer substantially, and none of them is obvious from the task alone. Downstream use is one of those facts that lives entirely in your head and costs one sentence to share.

  • "This goes on a slide" produces very different text from "this goes in a report"
  • "I will edit this heavily" tells the model to favour coverage over polish
  • "I will read this aloud" changes sentence length and rhythm noticeably
  • Naming the destination is often more effective than describing the style you want

One Task, Not Five

Prompts that ask for a summary, an analysis, three recommendations, a risk assessment and a draft email all in one message tend to produce five mediocre things instead of one good one. Attention gets spread and the later items are usually thinner than the earlier ones. Splitting is nearly always better: get the summary right, then use it as input to the analysis, then ask for the recommendations. This is slower in wall-clock terms and much faster in total, because you are not rewriting a compound answer where only part of it was wrong. A reasonable rule: if your prompt contains more than about three distinct deliverables, split it.

  • Compound prompts produce compound mediocrity — the later asks get the least effort
  • Chain instead: output of step one becomes the input to step two
  • Splitting makes it obvious which step failed when something is wrong
  • Rule of thumb: more than three deliverables in one message, break it up

Prefer slides, quizzes, and saved progress? Read this lesson in the library — free, no sign-up.