I build AI pipelines that refuse to ship bad output — every skill has a hard gate, and every gate has teeth.
- Building Claude Code skills that turn one-line intent into production-grade output
- Obsessed with hard quality gates — deterministic validators, graders, and refuse-to-ship checks
- Orchestrating codex worker fleets for mass production: images, decks, cards, corpora
- Grounding style in measured corpora, not vibes — if it isn't verified, it doesn't merge
- Writing in Korean, shipping for everyone
|
Compiles vague asks into finished gpt-image-2 prompts — tiered negative policy, result-based description, HEX-pinned palettes, C1–C12 category playbooks, and a validator that fails loud.
|
Rewrites AI-drafted Korean into one person's real voice — corpus-measured genre packs, a 3-axis register system, and deterministic style gates that a judge agent grades A to D.
|
|
McKinsey/Bain consulting-style HTML charts from a Claude Code skill — action titles, waterfall/marimekko layouts, and Chrome-rendered export.
|
One line of intent → a presentation-grade, dark-editorial HTML deck. A 90-point grader with seven hard-fail checks quietly quarantines anything below the bar.
|
|
Spawns Codex workers like a fleet — background
|
One topic → a gate-checked Instagram carousel plus a 9:16 Reels cut, fully automated — twenty physical gates between the idea and the upload button.
|
Compiling Vague asks into production image prompts (gongnyang-prompt-kit)
Rewriting AI-drafted Korean into a human voice (gn-voice)
Shipping Gate-checked decks, cards, and charts (deck-factory · cardprinter · chart-skill)
Orchestrating Codex worker fleets in parallel (codex-fleet)
Hard gates. Soft paws. Real output.