Hands-on prompt engineering and LLM evaluation. Built around a diagnose-then-ship loop: forensic transcript critique ▎ that names the exact failure mode, paired with a fixer that produces complete, deployment-ready rewrites — not ▎ vague advice.
This agent helps Python teams using the Anthropic or OpenAI SDK safely cut LLM spend without risking regressions. With zero code changes, it maps call sites and runs your test scenarios to rank token waste from missed prompt caching or heavy model tiers. It tests fixes in isolated git worktrees, enforcing a strict "do-no-harm" gate that reverts any patch failing to preserve behavior. It delivers a clear savings audit that converts into verified, behavior-preserving code patches upon opt-in.
I've reviewed hundreds of codebases and PRs across Python, JavaScript, C#, TypeScript, Go, Rust, and more. I've distilled that debugging and code review judgment into this skill to help you find bugs, spot security vulnerabilities, and improve code quality before merging.
Turn pasted AI evaluation cases into an evidence-bounded triage brief that separates expected and observed behavior, hypotheses, and missing checks without claiming a root cause or fix.
A senior-engineer-style code review assistant: severity-ranked findings, security smell detection, and constructive feedback across 8 language families.
Point it at your repo (or a specific change) and it scaffolds a unit-test suite for you — auto-detects your runner (Jest/Vitest, pytest, Go, XCTest, JUnit/Kotest), writes idiomatic AAA tests matching your conventions, then RUNS them and fixes test-side failures until green. It never edits your source — a real bug is reported, not papered over. Zero config.
From a local ISSUE file to a verified, committed deliverable — with a fresh-subagent acceptance gate that catches wrong screenshots and leaked logic.