Clarity: removing ambiguity

Clarity of phrasing

Most bad answers are precise answers to badly asked questions. The model doesn't ask for clarification: it picks the likeliest reading and proceeds confidently.

Four sources of ambiguity

Vague words. "Make it shorter" — by half? to a tweet? drop one paragraph?
Implicit audience. "Explain simply" — to a child, a colleague in another department, an investor?
Unclear scope. "Give me ideas" — three or thirty?
Hidden success criteria. You know what a good answer looks like, but you didn't say.

Technique: replace adjectives with the measurable

Vague:   "make the text more persuasive"
Precise: "add two concrete numbers and one real example,
          remove all adjectives like 'best' and 'unique'"

Rule of thumb: if a requirement can't be checked, the model can't deliberately satisfy it either.

Example: dissecting a request

Analyse our sales

When analysing sales it's important to consider trends, seasonality… (a lecture instead of an analysis)

Here's monthly sales data: [table]. Find: (1) the three months with the largest drop versus the previous month, (2) whether there's seasonality, (3) one conclusion about what to change. Conclusions only, no methodology explanation.

1) Drops: March (−22%), July (−18%)… 3) Conclusion: the dips coincide with promotions ending…

The second request supplies data, three specific questions and a ban on filler. The answer can't be padded.

The clarity test

Before sending, ask yourself: could someone with no context execute this unambiguously? If not, neither can the model.

Insight. Swap adjectives for checkable criteria. "More persuasive" → "two numbers and one example".

Common mistake. Assuming the model will ask about the unclear bits. It doesn't ask — it guesses and continues confidently.

Pro tip. For a complex task, start with: "ask me 3 clarifying questions before answering". It usually surfaces something you hadn't considered.

Cheat sheet

A bad answer = a precise answer to a bad question.
Adjectives → measurable criteria.
Test: could a stranger execute this unambiguously?
Complex work: start with clarifying questions.

1. What does the model do with an ambiguous request?

2. How do you improve "make it more persuasive"?

3. Useful move for a complex task?

Task — checked by AI

Take a real request of yours that produced a weak result. Find at least three sources of ambiguity in it and rewrite it with measurable criteria. Paste: the original request, your analysis of the ambiguities, the new request and a comparison of the answers.

← Back

🔒 Answer the question correctly to move on to the next lesson.