Prompts that cannot work

Everything so far has been about writing better requests. This lesson is the counterweight: a category of request where better wording changes nothing, because the problem isn’t the prompt.

Recognising these is worth more than any technique in this module. It stops you burning an afternoon on something that was never going to work, and it stops you trusting an answer you shouldn’t have asked for.

Four kinds that cannot work

1. Information it has no access to. Your company’s Q3 numbers, what your manager thinks, whether your flight is delayed, anything after its training data ends. Unless the tool is genuinely connected to that source — and you have checked that it is — you will get invention shaped like an answer.

2. Anything requiring accountability. “Is this contract safe to sign?” “Should I take this medication with that one?” The failure here isn’t only accuracy. It’s that when it’s wrong, nobody is responsible, nothing is insured, and you cannot appeal. Professionals carry liability; that is a substantial part of what you’re paying them for.

3. Precise counting and arithmetic over long text. “How many times does this 40-page document say risk?” It predicts text; it doesn’t tally. You may get a confident, wrong number. Use search or a spreadsheet.

4. Claims about its own workings. “Why did you say that?” produces a plausible-sounding rationalisation, not an audit trail. It has no privileged view of its own processing. Treat such explanations as narration.

The tell

Ask yourself: what would this need to know, and could it possibly know it?

If the answer requires today’s price, your private file, someone’s intentions, or a fact nobody has published, then a fluent answer is a warning sign rather than a success. The confident tone from E1·L7 is fully available for things it cannot know.

Convert, don’t abandon

Most impossible requests contain a possible one. Move from the verdict to the structure around it:

Cannot work	Works
”Is this contract safe to sign?"	"List the clauses a solicitor would look at first, and what makes each risky."
"What are our Q3 numbers?"	"Here are our Q3 numbers — what patterns stand out?"
"Is this legal here?"	"What questions should I ask a local lawyer, and what facts will they need?"
"Count the risks in this report"	"Here are the 14 risk mentions I found — group them by theme.”

The pattern: keep the judgement, delegate the preparation. You end up better prepared for the professional, and you didn’t substitute for one.

Try it now (5 minutes)

Write down the last question you asked an assistant that left you uneasy. Run the tell over it: what would it have needed to know?

If it couldn’t have known, rewrite it using the conversion table. Notice the rewritten version is often more useful than the original.

Check your understanding

1. You ask an unconnected assistant for your company's Q3 revenue. You get a specific figure. That figure is: Probably right — it sounds specific Invention shaped like an answer; it has no access to your data Retrieved from a cached copy

2. The main problem with 'is this contract safe to sign?' is: It's too long a document Nobody is accountable when it's wrong, and you cannot appeal Contracts are always refused

3. The best conversion of an impossible request keeps: The judgement with you, delegating the preparation The judgement with the model, delegating the facts Both with the model, then verifying

Recap

Four categories no wording rescues — inaccessible information, accountable judgement, exact counting, and self-explanation — and one tell: could it possibly know this? Convert to structure and keep the decision. Next: making what works reusable.