Prompt Engineering Specialist

This skill treats your prompt as an engineered artifact, not a growing text file. You give it your current prompt, the model you're running it on, and the failure cases that sent you here. It diagnoses which component is breaking — instructions that contradict each other, few-shot examples that avoid the hard edge cases, output specs the parser can't rely on, or context buried where the model won't weight it — then rewrites structurally and documents every decision so the next maintainer understands why the prompt was built that way. It also ships a regression case set so future edits produce measurable results instead of confident guesses.

A typical input: a 1,400-word customer-support system prompt that has accumulated eighteen months of appended fixes, three conflicting tone rules, and a JSON output section the model follows about 70% of the time. You paste it in, list the three failure patterns you see most often, and describe what a good output looks like. The skill asks four scoping questions — your stack, target model, who reads the output, any hard constraints — then begins with failure collection before touching a word of the prompt.

The output you get back includes: a restructured system prompt with role definition, constraint block, and output contract in explicit sections; a plain-English annotation explaining each structural change and why the original caused the failure; a few-shot example set chosen to demonstrate the boundary cases you flagged; and a regression case table with inputs, expected outputs, and the failure mode each case guards against. Everything is formatted for immediate use and future editing without re-discovering the reasoning.