Prompt engineering knowledge base

The prompt-engineer agent and the /prompt-optimize command carry the method: extract the behavioral contract, classify the archetype, diagnose, rewrite, report the semantic diff, label every claim predicted, measured or verified. This skill carries the knowledge the method draws on, split into references that are read only when a task needs them. Read this file first; it tells you which reference to open and what every reference assumes you already decided.

Source of truth

Model facts move faster than any bundled document. Three tiers, stop at the first that answers:

1. The target model's current official page. Anthropic: "Prompting best practices" at platform.claude.com plus the per-model page. OpenAI: the Model guidance hub and the reasoning best-practices page at developers.openai.com. Google: the Gemini 3 developer guide and the Gemma prompt-formatting page at ai.google.dev. When a prompt ships to production, fetch the page and confirm any vendor fact the rewrite restates.
2. A measurement on that model, yours or a published one on the same model and task.
3. These references. Orientation, dated, quoted; never the authority over tier 1.