The discipline of designing LLM inputs so outputs are accurate, structured, and repeatable. Start with best-practices, then read the deep-dives by topic.
Start here
- best-practices: The seven strategies that converge across Palantir Foundry AIP, OpenAI, Microsoft Foundry, and Google Vertex AI. Read this first.
- glossary: Comprehensive prompt-engineering glossary in question form (PAA-eligible). Every term grouped by category and linked to the deep-dive.
Patterns
How to structure a prompt for production use.
- prompt-templates: Reusable template shapes with placeholders for instruction, primary content, examples, and constraints.
- output-constraints: Specify length, format, schema, and exclusions so the output is parseable.
- prompt-chaining: Multi-step prompt sequences when one prompt would overload the model.
- reasoning-model-prompting: When and how to prompt reasoning-tuned models (Claude thinking, o-series) differently from chat models.
Production hardening
What to do before a prompt ships to users.
- prompt-evals: Build an eval set, regression-test prompts on every change, ship only when the threshold passes.
- prompt-injection-defense: Detect and mitigate prompt injection from user input, retrieved documents, and tool output.
- prompt-caching-strategies: Structure prompts so cacheable prefixes are stable across requests, maximizing cache hit rate.
Deep-dives moved into this category
These were originally in ai-agents/ and now live here. Cross-category complements stay in ai-agents/.
- chain-of-thought: Chain-of-thought scaffolds for reasoning-sensitive tasks.
- prompt-design: The end-to-end prompt-design loop.
Related in ai-agents
- structured-output: JSON mode and tool-use prompts for parseable output.
- rag: Retrieval-augmented generation patterns.
- mcp-servers: Model Context Protocol for tool wiring.
Glossary anchors
Atomic definitions for the prompt-engineering vocabulary.
- structured-prompt: A prompt with explicit sections (instruction, context, examples, format).
- role-priming: Assigning a role to anchor tone and depth.
- system-message: The stable, role-and-policy prompt sent on every turn.
- few-shot-prompting: Including worked examples in the prompt.
- temperature: The decoding-randomness parameter.
- context-window: The token budget the model can attend to.
- prompt-injection: Adversarial input that overrides instructions.
- tool-use: Calling external functions from within a prompt.