Prompt Pattern
Library

A four-part structure for making task instructions easier to evaluate.

Force the model to show its reasoning before answering.

Provide 2-4 examples targeting the specific ambiguity you want resolved.

Layer your system prompt: Identity, Persona, Constraints, Standing Context.

Thought, Action, Observation loop for tool-using agents.

Extract, validate, and retry with error context on failure.

Three-pass analysis: per-file, cross-file, then summary.

Use strict tool inputs on supported Claude models. Check stop reasons and validate meaning; schema compliance does not establish factual accuracy.

Break complex tasks into 3-5 focused sequential prompts.

Strip verbose tool output to structured facts before passing downstream.

Explicit rules for when the agent should hand off to a human.

"Do NOT do X" patterns that actually work in production.