The Ceiling of Prompting
Prompting Fixes Some Problems and Not Others
Prompting is the right first move for an enormous range of tasks, and it is genuinely surprising how far it goes. But it has a ceiling, and knowing where it sits saves a lot of wasted effort. Prompting is the answer when the gap is instruction: the model could do this if it understood what you wanted. It is not the answer when the gap is knowledge — the information simply is not available to it, and the fix is supplying that information rather than asking more elegantly. It is not the answer when the gap is capability — the task needs precise calculation, live data, or an action in a real system. And it is not the answer when the gap is judgement that should not be delegated at all.
- Instruction gap → prompting is exactly the right tool
- Knowledge gap → supply the material, or use a system that retrieves it
- Capability gap → use a real tool: a calculator, a database, a proper source
- Judgement gap → this one is yours, and no prompt changes that
Signals You Have Hit It
A few reliable signs. Your prompt is now longer than the output and still not working. You are handling exceptions in the prompt — "if it is a refund, do X, unless the customer is on the old plan, in which case Y" — which means you are writing software in prose. You need the same answer every single time and are not getting it. You are asking for something that depends on information nobody wrote down. Or three genuinely different approaches have all failed. Each of these points somewhere other than a better prompt: a different tool, a properly built process, a person who knows, or accepting that the task is yours to do.
- The prompt is longer than the output and still unreliable
- You are encoding branching rules in prose — that is a job for actual software
- You need identical output every time and cannot get it from wording alone
- Three different approaches failed, not three rewordings
The Skill Underneath
Here is the part that outlasts every tool. To write a good prompt you must decide what you want, who it is for, what a good result looks like, and what would make it wrong. That is specification, and it is the same skill behind a clear brief, a good delegation, and a well-defined problem. It is why the interfaces keep getting easier while the underlying ability keeps mattering — and why people who can specify clearly get more from every tool they touch, including the ones that do not exist yet. If you cannot describe the outcome you want, no amount of technique produces it. The prompting is the easy half; deciding what you actually want is the work.
- Prompting is applied specification — the same skill as a good brief or a clear delegation
- Interfaces will keep improving; knowing what you want will not become obsolete
- The habits transfer to briefing people, not just models
- If you cannot say what good looks like, that is your first task, not the model's
Prefer slides, quizzes, and saved progress? Read this lesson in the library — free, no sign-up.