Designing LLM inputs so outputs are accurate, structured, and repeatable. Start with prompt-design, then read by topic.
Start here
- prompt-design: Core rules for reliable prompts: specificity, system vs user content, tags, examples, ordering, and evals.
- prompt-migration: Checklist for moving prompts to a newer model: rejected parameters, outdated instructions, effort sweep.
Structure and output
- context-engineering: Assemble, order, and budget the whole context window, not just the prompt.
- prompt-templates: Template skeleton with delimiters, named placeholders, and versioned files.
- output-constraints: Format, length, escape hatches, and validate-then-retry for prose output.
- prompt-chaining: Split multi-step work into narrow prompts with typed hand-offs.
Reasoning models
- reasoning-model-prompting: Prompt reasoning models with goals, set effort and thinking levels, reserve output headroom.
- chain-of-thought: When prompted chain-of-thought helps, and when built-in thinking replaces it.
Production hardening
- prompt-evals: Eval sets, metrics, CI regression runs, and sign-test promotion gates.
- prompt-injection-defense: Prompt-level injection defenses: placement, encoding, screening, evals.
- prompt-caching-strategies: Layout and provider controls that raise prompt-cache hit rate.
Related in ai-agents
- system-prompts: What belongs in the system prompt, structure, and versioning.
- role-framing: Role, audience, and tone lines; calibration and anti-sycophancy rules.
- few-shot: Examples vs rules, writing examples, and rebalancing the mix.
- structured-output: Schema-enforced JSON, supported schema subsets, validation.
- prompt-injection-defense: System-level injection controls for tool-using agents.
Glossary anchors
- system-prompt: The stable role-and-policy prompt sent apart from the conversation; covers role priming.
- structured-output: Model output that matches a JSON Schema; covers structured prompts.
- 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.