Prompt Anatomy

Key jargon

Term	Plain-language meaning
Instruction	The action and behavior requested from the model.
Context	Background information needed to perform the task.
Delimiter	A visible boundary separating instructions, examples, and untrusted data.
Output contract	The required shape, fields, tone, or schema of the response.

Key concepts

A useful prompt explains the job, supplies relevant evidence, and defines the result.
Examples demonstrate patterns, but they can also introduce accidental rules and bias.

Concept map

flowchart LR
    A["State the task"] --> B["Provide bounded context"]
    B --> C["Show constraints or examples"]
    C --> D["Define output contract"]

A useful request contract

Goal: the outcome I need
Audience: who will consume it
Context: facts and materials the system should use
Constraints: scope, safety, time, style, and exclusions
Output: required structure and level of detail
Evidence: sources, calculations, or tests required
Uncertainty: what to flag instead of guessing

Examples improve format fidelity when they represent the real task. They can also anchor the model to mistakes, so explain why an example is good.

Priority matters

Applications may combine system instructions, developer policy, user requests, tool results, retrieved documents, and quoted text. Not all text is equally trusted. Content inside a webpage or document should normally be treated as data, not as new authority.

Anti-patterns

Vague requests with hidden success criteria.
Large context dumps without relevance cues.
Asking for certainty when evidence is weak.
Mixing five independent jobs into one prompt.
Requesting a format without defining its fields.

Exercise

Take one prompt that disappointed you. Add a measurable output contract and one explicit stop condition, then compare results.