Prompt Engineering Patterns

Prompts inside application code are code: versioned, tested, evaluated, monitored. A prompt is not a string literal you tweak in place until "it works" — it is a contract with a probabilistic runtime.

Activation triggers

• Writing a new LLM call from application code (Python, TS, or otherwise).
• Refactoring a hand-crafted prompt that has grown into a wall of prose.
• Structured outputs fail intermittently; JSON parsing errors at 3AM.
• Few-shot examples don't land — model ignores format, style, or edge case.
• A long RAG prompt drifts as context grows (lost-in-the-middle).
• The LLM bill is dominated by uncached prefix tokens.

Not for: agent-body system prompts (use agents-claude-creator / opencode-agent-creator / pi-extension-creator), Agent Skill bodies (use skill-creator), tool signatures (use tool-schema-design), or model comparison / benchmarking work.

The core patterns (stack them)

Production LLM calls typically compose 3–4 of these together. Each one is worth using alone; combined they behave like typed function calls: inputs in, structured outputs out, no surprises.