prompt-engineer

What it does

Guides an agent to craft and refine prompts, system prompts, tool descriptions, and agent instructions for multiple LLM APIs (Claude, GPT, Gemini). Triggers when editing prompts, tool descriptions, or discussing prompt quality; also covers prompt optimization, tool-use accuracy, cross-provider compatibility, and prompt reviews.

How it works

Outlines a step-by-step process for editing prompts: read the existing prompt, identify necessary improvements, draft the minimal change, re-read for contradictions, and test on the target provider. For new prompts, start minimal (role + constraints + examples + output format), test, then iteratively add instructions only if failures are observed. Requires checking for contradictions and clarity before submission to ensure unambiguous follow-through by colleagues.

General best practices include: structure prompts with explicit roles and XML-like sections, include 3-5 diverse few-shot examples wrapped in <example> with <input> and <output>, define a clear output format with concrete examples, and use positive framing about actions to take. Emphasizes context engineering, phased task decomposition, and agentic system considerations (subagents, state management, and controlled autonomy). Tool descriptions should be detailed (3-4 sentences minimum for full models) and cover what the tool does, when to use it, when not to use it, parameters, return value, and caveats. Encourages patterns like decision-tree routing, explicit alternatives, and explicit safety constraints.

It provides guidance on how to structure and optimize cross-provider tool descriptions, and highlights response design principles to return high-signal information and concise results, with explicit absence notes and error handling.

When to use it

Triggered when prompts or tool descriptions are being written or edited, when agent instructions or function calling schemas are being designed, or when discussing prompt quality for any LLM API; also applicable for prompt optimization and cross-provider compatibility discussions.

What it can touch

Not applicable in this summary because the content focuses on procedural guidelines rather than specific tools or scripts in this excerpt. The skill emphasizes using standardized prompt structures, tool descriptions, and workflow for testing across Claude, GPT, and Gemini.

Caveats

Contains guidance and best practices but does not guarantee outcomes; results depend on provider behavior and prompt execution. No explicit licensing or risk disclosures beyond standard constraints for prompt engineering guidance.