llm-patterns

Use when a task requires autonomous multi-step reasoning — the LLM must observe, decide, act, and iterate until a goal is met or a termination condition is reached. Apply when a single prompt cannot solve the task, the number of steps is not known in advance, and the next step depends on the result of the previous one. Covers ReAct, Plan-and-Execute, state management, termination, and guardrails for autonomous agents.

Use when you cannot systematically measure whether your LLM feature is working correctly. Apply when testing is based on vibes rather than metrics, when you cannot detect regressions after prompt changes, or when production quality is unknown. Covers evaluation datasets, metrics, regression testing, LLM-as-judge, and production monitoring for non-deterministic systems.

Use when an LLM-powered feature must remain functional when the primary model is slow, down, over budget, or producing low-quality results. Apply when building any production AI feature that users depend on. Covers fallback chains, semantic routing, circuit breakers, cost management, and degradation levels.

Use when LLM inputs or outputs must be validated for safety, policy compliance, schema conformance, or content appropriateness before they reach users or downstream systems. Apply when LLM responses could contain harmful content, PII leakage, prompt injection, off-topic responses, or policy violations. Covers input validation, output validation, content filtering, and prompt injection defence.

INVOKE THIS FIRST before designing any LLM-powered feature. Use when integrating an LLM as a component in a software system — not as a chat interface, but as a decision-making, data-processing, or logic-executing building block. Maps the friction you feel to the pattern that removes it.