Prompt

What is a Prompt?

A prompt is the input sent to an AI model at inference time. In the simplest case it is a question or instruction. In production AI workflows, it is a carefully structured document that frames the model's role, supplies relevant context, states the task precisely, specifies the output format, and often includes examples of correct behavior.

The quality of the prompt directly determines the reliability of the output. The same underlying model will produce dramatically different results depending on how the prompt is written. This is not a limitation to work around — it is a lever to use deliberately.

Anatomy of a Prompt

A production-grade prompt typically has five components:

• Role: Tells the model who it is and what expertise it should apply. "You are a document processing agent for a wholesale distributor."

• Context: Supplies the information the model needs to do the task — document content, relevant ERP data, previous step outputs.

• Instruction: States precisely what the model should do. "Extract the following fields from the invoice below. If a field is missing, return null."

• Format: Specifies how the output should be structured. "Return a JSON object with keys: invoice_number, supplier_name, date, total_amount, line_items."

• Examples (few-shot): One or more input/output pairs demonstrating expected behavior. Especially valuable for complex extraction or edge cases.

Prompts in Operations

In an automated operations workflow, prompts are code — they should be version-controlled, tested against edge cases, and updated when outputs degrade. An invoice extraction prompt that worked reliably for six months may start producing errors when a supplier changes their template. The fix is usually a prompt update, not a model change. Treating prompts as throwaway instructions rather than maintained assets is one of the most common reasons AI workflows become unreliable over time.