A Claude skill that writes, reviews, rewrites, and debugs prompts for Claude: system prompts, custom instructions, agent and harness prompts, CLAUDE.md files, and one-off task prompts. It tunes each rewrite to the model that will run it, because prompting advice is measured per model and a fix for one model can hurt another.
SKILL.md walks through five steps:
- Identify the target model from what the user names, a model string in the material, or where the prompt will run.
- Identify the prompt type (chat system prompt, agent or harness prompt, task prompt, custom instructions) and whether the user controls API settings.
- Diagnose reported symptoms, lines that hurt the target model, missing essentials, and risky domains.
- Rewrite with the matching model guidance.
- Check and deliver the optimized prompt, what changed and why, settings, and test inputs.
prompt-engineering-expert/
├── SKILL.md # Skill instructions and workflow
├── CLAUDE.md # Copy of SKILL.md, so Claude Code uses the skill in this folder
└── references/
├── general.md # Baseline best practices, per-model notes, audit list, migration checklist
├── opus-5-5.md # Claude Opus 5.5 specifics
├── fable-5-1.md # Claude Fable 5.1 / Mythos 5.1 specifics
└── medical-guidance.md # Verbatim per-model medication dosing blocks
The model files layer on top of general.md and win where they disagree. Sonnet 5.5, Sonnet 5, Haiku 5.5, Opus 5, Opus 4.8, and Fable 5 are covered by general.md section 10, which points to each model's docs page.
The guidance was reviewed against Anthropic's prompt engineering docs on 2026-10-28 (general and Opus 5.5) and 2026-09-26 (Fable 5.1). The review dates and source links are in the Freshness section of SKILL.md and at the top of each reference file.
Copy SKILL.md and references/ into a folder named prompt-engineering-expert in your skills directory, or upload the folder as a skill in Claude.