Senior adversarial audit of an AI-agent-built and/or agentic/LLM backend, run after an agent finishes a feature. 18 lenses across security & access control, prompt & agent safety, data protection & leakage (PII/PHI, RAG/embedding/log/cross-tenant), reliability, observability/evals/governance, and AI-authorship hygiene. Names each issue, cites file:line, warns 'this may cause <impact>', rates severity 0-4, and ends with a Ship/Fix-then-ship/Block verdict.
1 skill
Build or debug an AI video dubbing/recap pipeline where the translated voiceover lands on the right moment on screen. Battle-tested on Chinese short-drama / manhua / donghua -> English. The AI (STT/translate/TTS) is the easy part; timing is the whole game. Encodes the winning architecture (burned-in subtitle OCR as ground-truth timing -> isometric syllable-budget translation -> timestamp-anchored TTS with a NO-SLOW fit policy -> mux), the ASR dual-pass (mix + demucs vocal stem) coverage fix, a catalogue of sync failure modes with numeric limits, and the validation methodology that never lies (extract the frame + transcribe the OUTPUT audio).
1 skill
Build 3Blue1Brown-style visual, intuition-first math explainer videos with Manim — a complete, battle-tested YouTube pipeline: topic confirmation gates, TTS narration with per-sentence sync (Sarvam Hinglish / Deepgram English), free pre-render QA (static layout frames, AST narration audit, independent editor-agent review), the render→frame-critique→fix loop, music mixing verified by subtraction, chapter markers, mascot rules, thumbnail recipe, and every Manim gotcha that cost real hours.
1 skill
Senior-level UI/UX audit of a complete website or web app: heuristic evaluation (Nielsen's 10), WCAG 2.2 accessibility, Core Web Vitals, visual hierarchy, responsive behavior, navigation/IA, forms and conversion friction — delivered as a severity-rated, scored, RICE-prioritized report.
1 skill
Write test cases that actually catch bugs — not tests that go green while the shipped feature is broken in production. Teaches the senior discipline from a real green-but-broken incident: test the real artifact (never a reimplementation), assert the behavior/content contract (not liveness), prove red-before-green, hunt silent failures, contract-test the boundaries you don't control, cover edges and limits, and verify the fix in the shipped build and production telemetry. Language- and framework-agnostic, with a smell catalogue, coverage checklist, 7-gate verification, case studies, and a test-plan template.
1 skill
Edit prose to remove common signs of AI-generated writing while preserving meaning, tone, and a distinct human voice. Catches puffery, filler, vague claims, canned transitions, forced lists, jargon, passive constructions, and formatting habits that make text feel machine-made.
1 skill