Prompt Optimization

Build effective, safe, and token-efficient system prompts for LLMs.

How to Use This Skill

Pick the workflow that matches the task, then follow it end-to-end:

Each workflow references the specific detail pages below as needed. Do not read all references upfront — load them when the workflow step calls for them.

Core Principles

1. Clarity over cleverness — Write prompts a new employee would follow without confusion
2. Positive framing — Say what to do, not what to avoid
3. Structure with delimiters — XML tags or consistent separators between content types
4. Progressive specificity — Role → constraints → instructions → context → format → examples → input
5. Token economy — Every token must earn its place; ≥1 constraint per 5–10 tokens
6. Safety by default — Prevent injection, data leakage, and bias structurally

Prompt Architecture (ordering)

1. Identity / Role          — Who the model is (1–2 sentences)
2. Core behavior rules      — Non-negotiable constraints (safety, tone)
3. Task instructions        — What to do and how
4. Context / Knowledge      — Domain data, schemas, reference docs
5. Output format            — Structure, length, style requirements
6. Examples (few-shot)      — 2–3 input/output pairs in <example> tags
7. Dynamic input            — User content (always last, always delimited)

Key Rules

• Delimit user input — Wrap in tags like <user_input>; never interpolate raw
• Show, don't just tell — Few-shot examples beat format descriptions
• Match prompt style to output style — Markdown in prompt → markdown out; prose → prose
• Put long documents first — Reference data above instructions (30%+ quality gain)
• Quantify, don't qualify — "≤100 words" not "keep it short"
• Imperative voice — "Respond in JSON" not "You should format your response as JSON"
• Test adversarially — Injection, edge cases, ambiguous inputs before shipping

Reference Index

Load these when a workflow step calls for them: