One Hundred and Forty-Seven

That is not a round number chosen for effect. That is the actual count , a developer who got frustrated enough to document every AI prompt that did not work, with notes on why each one failed.

The project started as venting. It turned into something more useful. After enough failures, a pattern emerged that had nothing to do with prompt length, tone, or topic specificity.

It was about structure. Where in the prompt you put things.

What Most People Do Wrong

The default way most people structure a prompt: context first, instruction last.

"I'm working on a marketing strategy for a B2B SaaS company targeting mid-market financial services firms. We have limited budget and need to prioritise channels. The company has been operating for three years and has 45 enterprise clients. What should our strategy be?"

The AI reads all of that. But it weights the end. The most recent tokens have more pull on the output than the earlier ones. So the prompt above produces a generic strategy answer, lightly flavoured by the context you buried at the front.

Flip it. The instruction goes first. The context comes second.

"Give me three non-obvious channel recommendations for a B2B SaaS company. Context: mid-market financial services, three years old, 45 enterprise clients, limited budget."

Same information. Different order. Dramatically different output.

The Five Structures That Work

Structure 1: Instruction → Context. What you want first. Everything else second. Works for research, recommendations, analysis, and summarisation.

Structure 2: Constraint → Task. Name the limits before you name the job. "In plain English, no jargon, under 100 words: explain why [X]." The constraint shapes how the AI interprets the task , it does not just clip the output at the end.

Structure 3: Role → Instruction. "You are a sceptical investor reviewing this pitch deck. What are the three weakest assumptions?" Role before task. The role is not decoration , it changes what the model considers a "good" response.

Structure 4: Failure Case → Request. "The last three times I asked AI to do this, I got [X problem]. Avoid that. Now: [task]." Pre-empt the failure mode before it happens. Most AI outputs are predictable , use that predictability against the problem.

Structure 5: Output Format → Content. "A numbered list, each item one sentence, no preamble: [task]." If you need a specific format, specify it before the content request. Specifying it after produces a well-formatted response to the wrong question about half the time.

The Principle Underneath All Five

Every one of these structures does the same thing: it shapes the model's frame before giving it the task. Rather than asking the AI to do something and hoping it interprets the request the way you meant it, you construct the interpretive frame first.

Think of it like giving directions. You do not say "go to the coffee shop on Main Street , oh, and it's 6 PM so avoid the bridge because traffic is bad and you'll need to turn left not right." You establish the constraint , avoid the bridge , before you give the route. Same principle.

One hundred and forty-seven failures. One insight. It is almost annoyingly simple once you see it.