What is Prompting Techniques?

Why It Matters

Prompting is the cheapest lever in an AI system: no training run, no infrastructure, no deployment. That is also its limit. It shapes what the model is likely to produce, and it enforces nothing. A prompt that says “always cite a source” is a preference; a retrieval step that attaches the source is a control.

Knowing the difference is what keeps a prototype from being mistaken for a system. Prompting gets a workflow working. Structure is what keeps it working when the input is unusual and the stakes are real.

The Main Techniques

Zero-shot. The instruction alone, with no examples. Fastest to write and usually enough for tasks the model has seen a great deal of, such as summarising or rewriting.

Few-shot. Two to five worked examples of input and expected output. The reliable way to fix a format or a tone, because a demonstration communicates constraints that a description struggles to.

Chain of thought. Asking the model to reason before answering. It helps on multi-step problems and is not a guarantee of correctness, since a fluent chain can be as wrong as a direct answer.

Format constraints. Specifying the structure precisely, ideally as a schema the output must satisfy. This is the technique that most reduces downstream parsing work, and it pairs with validation rather than replacing it.

Role and context framing. Setting the persona, the audience, and the material in scope. Useful, and the point at which a prompt starts to resemble configuration and should be treated as such.

Where It Breaks

Prompting where structure is needed. The most expensive mistake in this list. An instruction cannot enforce a limit, check a source, or prevent an action. Anything that must hold regardless of input belongs outside the model.

Prompts that grow into programs. Fifty lines of instructions, three nested conditions, and a rule for every exception is a program written in the least testable language available.

No version control. A prompt is a production artefact. If a change to it is not versioned and evaluated, a behaviour change cannot be rolled back or explained, which is what prompt versioning exists to fix.

Examples that leak the test set. Few-shot examples drawn from the same cases used to evaluate the system make the evaluation meaningless. The examples and the test set have to be different.

Assuming a technique generalises. A pattern that works on one model version may fail on the next. Anything load-bearing belongs in the evaluation rather than in the folklore.

How Flytebit Handles It

We use prompting where it is strong, which is shaping output and fixing format, and we treat prompts as versioned artefacts with an evaluation gate rather than as text edited in place. Anything that has to hold under pressure, such as a limit, a source requirement, or an action boundary, is enforced by the control layers around the model instead. The build side is our generative AI development work, and the enforcement design is in AI architecture design.