Prompting Claude Models

Prompts are the interface to an LLM the way an API signature is the interface to a function — most "the model is wrong" bugs are really underspecified-prompt bugs. This skill covers the durable craft of writing prompts for Claude models in research and coding work, and points you at the live docs for the parts that change with each model release.

When to use this skill

• Writing or revising a system prompt or task prompt for a Claude-backed feature
• Building an LLM-as-judge / evaluation rubric, a RAG answer prompt, or an extraction/classification prompt
• Designing tool definitions or agent scaffolding (when to call tools, when to spawn subagents, how to report progress)
• Debugging behavior: ignored instructions, over/under-triggered tools, wrong verbosity, refusals, inconsistent judgments
• Calibrating effort / thinking depth for an agent or pipeline

When not to use it

• API/SDK mechanics (model IDs, pricing, sampling parameters, tool-use JSON schema, prompt caching, token counting, migration) → use the claude-api reference. This skill is about what the prompt says, not how the call is wired.

Ground model-specific behavior in the live docs

Behavior that depends on the model version — default verbosity, effort levels and defaults, thinking triggers, tool-use tendencies, subagent spawning, frontend house style — shifts with each release and goes stale fast. Do not recite it from memory or freeze it into a prompt. When a task turns on one of these, read the relevant page first and quote what it currently says.