Every exchange with a generative AI system begins with a transmission you write yourself. This handbook covers how to send a clear signal, what happens behind the receiver, and how to spot someone jamming the frequency.
A prompt is the plain-language instruction handed to an AI system — the closest thing it has to a control panel. How well that instruction works depends on two different kinds of work, and most operators only ever touch one of them.
The craft of wording an instruction so the model understands exactly what is being asked of it.
The technical work that happens behind the interface, usually invisible to whoever's typing.
Every prompt is built from the same four parts, in different proportions. Learn to spot each one and you can debug a weak prompt by asking which piece is missing.
The action that should be performed.
Additional information the model might not be aware of — examples, demonstrations, or the role and behavior it should adopt.
The data the task should be performed on.
The desired shape or style of the response, and how long it should run.
Wording alone can change how a model reasons through a problem. These techniques control how much guidance you give and how visible the model's thinking becomes.
How many worked examples you include before asking for the real thing. More examples usually means a more predictable format — though, oddly, whether those examples are factually correct rarely changes the result. The model copies the pattern, not the facts.
Ask the model to narrate its reasoning before landing on an answer. The output becomes easier to follow and easier to check — you can see exactly where the logic might have gone wrong.
Run the identical prompt several times and compare the answers instead of trusting the first one. The response that turns up most often is treated as the most likely to be correct — strength in numbers, not in any single attempt.
Push chain-of-thought one step further: instead of one line of reasoning, the model develops several independent perspectives in parallel — like a panel discussion — then merges them into one recommendation.
Split a large task into a sequence of smaller prompts, where each response becomes the input for the next one. Better suited to long or multi-stage work than a single oversized instruction.
Flip the switches to add or remove parts of a request and watch the transmission take shape. Task stays live — everything else is what turns it from noise into signal.
Small adjustments that consistently improve results. Expand any tip for the reasoning behind it.
Because language models respond to whatever text they're given, that openness can be turned against them. Prompt injection is the umbrella term for any attempt to steer a system outside its intended behavior using crafted input.
Run any draft prompt through this list before sending it. Most well-formed prompts touch at least five of these.