hwip
App Store · Google PlayCoding agents stop when they need you. hwip sends the question to your phone and types your answer back.
I’m Hyungtak Jin, a software engineer of eleven years. I have led a frontend team for eight of them, building web, mobile and desktop apps, and now I ship whole products, from the app to the AWS backend, with an AI agent workflow I designed.
claude
Claude Code
~/hwip
/orch let people answer agent questions from an Apple Watch
orch
add · M · done: a wrist reply reaches the terminal
/01-plan
3 slices · 4-row acceptance contract
/02-implement
s01 red → green · tests pass
/03-review
fresh-context finder · 0 critical · PASS → commit
/ship
verify-check 5/5 green
draft PR opened
Coding agents stop when they need you. hwip sends the question to your phone and types your answer back.
One outfit every morning from the forecast, on the lock screen, widgets and Apple Watch.
Commit messages for Git and Jujutsu from Ollama, Claude, Gemini, ChatGPT and more, side by side.
brew install aicommit2
npm install -g aicommit2
View on GitHub
Fig-style autocomplete for zsh on macOS: subcommands, flags, branches and scripts pop up as you type. No login, no AI, no telemetry.
brew install nerv-sh/tap/nerv
View on GitHub
Local hybrid search for Korean Markdown notes and Obsidian vaults: BM25, Hangul bigrams, embeddings, rerank. CLI, MCP and SDK.
npm install -g ko-qmd
qmd query "회의록 결정 사항"
View on GitHub
Local decision models on Apple Silicon, benchmarked side by side.
Building software since 2015: medical-device software first, then eight years leading a frontend team. Today the work runs on an agent setup I built on Claude Code, opencode and pi. Planning, building, review and QA each have their own gates and hooks, so agent-written code is checked like human-written code before it ships. That is how one engineer ships at team speed.
Claude Code · opencode · pi · MCP · custom skills, hooks and review gates
Each task starts with its pass conditions: tests, type checks, lint. The agent checks its own result against them. “Looks done” is not done.
Review runs in a separate agent that did not write the code, and it never sees the rule that turns its findings into pass or fail, so it cannot bend findings to meet it.
Apart from conventions, every change is checked against the spec: anything missing, wrong or extra.
Design and scope have no machine-checkable answer, so they go to a person instead of a loop.