Chapters: 
AI

Core definition (AI agent context)

A prompt is the structured input given to an AI system that defines what it should do, how it should behave, and what context it should use.

Think of it less like a question and more like a mission envelope.

In an AI agent, a prompt is not just text

It’s typically a composite payload with layers:

  1. Instruction layer (the “charter”)
    • What role the agent plays
    • Goals, constraints, tone
    • Example: “You are a security analyst. Evaluate access requests…”
  2. Context layer (the “world state”)
    • Data, documents, memory, prior steps
    • Retrieved knowledge (RAG), tool outputs, state variables
  3. Task/input layer (the “current request”)
    • The actual thing to act on
    • Example: “Analyze this log entry…”
  4. Output expectations (the “contract”)
    • Format, schema, level of detail
    • Example: JSON, bullet points, decision + rationale

In agent systems specifically

A prompt becomes more like a control loop input than a one-off request:

  • It is reconstructed every step
  • It may include:
    • Previous actions
    • Tool results
    • Memory state
  • It evolves as the agent “thinks”

So in agent terms:

A prompt is the current assembled state of instructions + context + task that drives the next action.

Short, operational definition

If you want the version you’d put in a design doc:

Prompt = the full input context that conditions an AI agent’s next decision or response.

Why this matters (practically)

Because changing the prompt changes everything:

  • Same model, different prompt → different behavior
  • Weak prompt → hallucinations, drift
  • Strong prompt → predictable, testable agent

In your world (pipelines, agents, structured flows), the prompt is basically:

the API contract between your system and the model’s reasoning engine

Tiny example (agent-style)

Instead of: 

What is this error?

An agent prompt looks more like: 

ROLE: Linux SRE
GOAL: Diagnose failure cause
CONTEXT:
  - system: RHEL8
  - service: uWSGI
  - logs: <attached>
TASK:
  Identify root cause and next action
OUTPUT:
  { "cause": "...", "fix": "..." }

That entire block is the prompt.

If you want, we can go one level deeper and map this directly onto your current agent pipeline (ingestion → normalization → lookup → UI) and show exactly where prompts are being formed, whether you intended them or not.

prompt