AI can't read your mind

It's an interactive partner, not a vending machine. The quality of what comes back is mostly the quality of what you put in.

Discussion drills

Diagnose 1

A student prompts: "Explain quantum entanglement." Claude produces a 400-word explanation with Dirac notation and references to Bell's theorem. The student needed a three-sentence explanation for a Year 8 science class.

Name the description failure. Which lever or levers were missing or wrong? Could any amount of "be simpler" in a follow-up fix this — or was the root problem in the original spec?
Construct 2

Take this core idea: "Social media harms attention spans." Write three description prompts that would produce three genuinely different outputs — for (a) a parliamentary committee, (b) a 13-year-old who is sceptical, (c) a cognitive neuroscientist. Each prompt should differ only in how you pull the four levers, not in the core idea. State which levers you changed and why for each.
Predict 3

You specify audience and register precisely but leave format completely open. Predict what defaults Claude will apply to format. Then argue: are those defaults neutral? What assumptions are embedded in them?
Judge 4

"More description always produces a better output." Evaluate this claim. Name a case where adding more description produces a worse result. What is the condition that determines when more description backfires?
Compare 5

Specifying the audience in detail vs specifying the output format in detail — which lever produces the larger change in Claude's output? Under what task conditions does each lever dominate?
Falsify 6

"If Claude understands the task, it does not need a detailed spec — it will infer what you mean." Make the strongest attack on this claim using a concrete example where inference fails even when the task is unambiguous.

Apply this in the project

The tribunal brief you write today is description applied to an argument. Your framing determines how the case lands.