The safety and verification layer for agentic coding —
hooks, skills, and MCP servers that make an AI harness trustworthy enough to run unattended.
Most AI coding effort goes into the prompt. I think the leverage is one layer below that — in the
harness: the hooks that stop an agent before it force-pushes, the skills that encode a procedure
so it isn't re-derived every session, the MCP servers that hand an agent real context instead of
guesses, and the memory that survives a /clear. An agent is only as good as the environment you
put it in. That environment is what I build. Three projects carry most of that work.
Existing languages were designed to be easy for humans to write. From here on, most code is generated — so what does a language look like when it's designed to be easy for an LLM to reason about and optimize?
Developer → Intent (what) → LLM → Semantic IR → Native Optimizer → Machine Code
You declare goals and business rules; the compiler and an agent pipeline design, implement, verify,
and optimize the rest. .lnpl parses and lowers to a semantic IR, runs on an IR interpreter and
compiles through MLIR to a native binary — and a differential check confirms both modes agree on
execution order, policy outcome, observability signals, and masking. OpenAPI is generated from the
IR. Next: a custom lnpl MLIR dialect.
Guardrail hooks, memory system, skill architecture, audit logging — the base layer every agent I run sits on. Written after learning the hard way that a guard regex which matches everything is indistinguishable from one that matches nothing, so the hooks ship with both-direction regression tests: a mention passes, an execution blocks. A guard you haven't tried to fool is decoration.
npm install -g @younggichoi/cliclawA Telegram bot driving Claude Code, Codex, Pi, and Gemini CLI as a macOS daemon. Per-chat agent sessions, a confirm gate before anything destructive, response streaming, image attachments, auto-installed launchd, and corporate-TLS (Zscaler) auto-detection.
| claude-secretmode | Leave-no-trace Claude Code sessions — runs on a RAM disk so transcripts, prompt history, and file snapshots never touch disk, while still inheriting keychain auth, MCP servers, hooks, and skills. npm i -g @younggichoi/claude-secretmode |
| dev-loop | Knowledge management for coding agents — wiki-grounded verification, RFC-based docs, and best-practice capture that feeds the next task instead of being relearned. |
| loop-orchestrator | Multi-agent orchestration — parallel execution, TDD/PDCA/Reflexion loops, test-quality audit, merge-gate verification. |
| dev-llm-wiki | Case-routed engineering knowledge written to be loaded as minimal working context by an agent, not read by a human. One case per page, routed by domain. |
| awesome-claude-plugins | Plugins that extend Claude Code with commands, agents, hooks, and MCP servers. |
| awesome-cli-coding-agents | Terminal-native coding agents and the harnesses that orchestrate them — Pi, OpenCode, Aider, Goose, Claude Code, Codex, Gemini CLI, parallel runners, autonomous loops. |
Also shipped — services and utilities
| korea-data-suite | Clean REST APIs over Korean public data — nationwide real-estate transaction prices (MOLIT) and public holidays (KASI), normalized to plain English JSON. |
| chungyak-alimi | Housing-subscription notices pulled from official open APIs into PostgreSQL, matched against your criteria, pushed to Telegram, browsable on a FastAPI dashboard with parcel polygons. Runs on a Raspberry Pi under systemd. |
| mac-inputlock | Lock your Mac's keyboard and mouse while the screen stays on — ⌃⌥⌘L. For wiping the keyboard, or cat-proofing. |
| meetingSummary | On-device meeting transcription and summarization on macOS. |
| ai-gossip | What several models say about each other when they think you've left the room. |
Cards are generated by a scheduled workflow and committed to this repo, so they don't break when a public widget host goes down.
Building in public. If you work on agent harnesses, MCP, or LLM automation,
open an issue on any of these — I answer.