Select a result to preview

Prompting

Intro

Prompt engineering is the practice of turning a vague user intention into a precise model task. It matters because LLMs are probabilistic generators: small wording and setting changes can shift correctness, style, and reliability. In production, a prompt is part of your system interface, so you should treat it like code: explicit, testable, and versioned. This hub covers the foundations, while child pages in this folder go deeper into in-context learning, reasoning, prompt composition, and automated optimization.

Prompt Anatomy

Most effective prompts combine four elements:

Example with all four elements:

Instruction: Extract security risks from the incident note.
Context: You are helping a SOC analyst. Keep findings actionable and concise.
Input data: "API keys were stored in plain text logs for 3 days in staging."
Output indicator: Return JSON with fields risk, impact, mitigation.

Mechanically, each element removes uncertainty: instruction narrows behavior, context biases interpretation, input anchors the specific case, and output indicator constrains format.

LLM Settings

Prompt text controls intent, while generation settings control sampling behavior and output boundaries.

Starting ranges below are heuristics, not universal defaults. Validate them with task-specific evals for your chosen model before using them in production.

Task type Temperature Top-p Max tokens Stop sequences
Creative writing 0.8-1.0 0.9-1.0 600-1200 Optional section markers
Classification 0.0-0.2 0.1-0.4 20-80 Label boundary, newline
Code generation 0.1-0.3 0.8-1.0 200-800 ``` or custom delimiter

Practical rule: tune temperature first, keep top-p near default unless you have a measured reason to change both.

Instruction Prompting

Instruction prompting is direct natural-language control: tell the model exactly what to do, how to do it, and how to format the result. It works best when instructions are specific, observable, and testable.

Good instruction pattern:

Example 1 (name normalization):

Convert the person name to this format: <Last name>, <First name>.
If suffix exists, keep it after first name.
Input: "Nikita Reshetnik"
Output:

Example 2 (PII redaction):

Redact all personal data from the email.
Replace names with [NAME], phones with [PHONE], and emails with [EMAIL].
Return only redacted text.
Input: "Hi John, call me at 410-805-2345."

If outputs drift, tighten the output indicator before adding complexity.

Role Prompting

Role prompting assigns a perspective that shapes style, depth, and framing. It does not replace task instructions; it modifies how the model executes them.

Illustrative contrast:

Standard: Write a review of this pizza place.
Role-based: You are a food critic writing for a city newspaper. Write a review of this pizza place in 120-150 words, focusing on crust texture, sauce balance, and service.

The role-based version typically yields richer domain vocabulary and better evaluative structure because the model has a clearer perspective prior.

Choosing a Technique

Use this quick decision flow for first-pass prompt design:

flowchart TD
    A[Start with task goal] --> B{Simple direct task}
    B -->|Yes| C[Use instruction or zero shot]
    B -->|No| D{Need strict output shape}
    D -->|Yes| E[Use few shot examples]
    D -->|No| F{Needs deeper reasoning}
    F -->|Yes| G[Use reasoning scaffolding plus verification]
    F -->|No| H{Multiple dependent steps}
    H -->|Yes| I[Use prompt chaining]
    H -->|No| J[Use role plus instruction]
    C --> K[Iterate with meta prompting]
    E --> K
    G --> K
    I --> K
    J --> K

For deeper implementation patterns, use targeted follow-ups such as In-Context Learning when format consistency is weak and Prompt Composition when one prompt is not enough. Prefer verifiable outputs over eliciting hidden reasoning traces.

Pitfalls

Questions

References

Whats next

Parent
LLM

Pages