Archives

Cách Undercroft giữ dữ liệu gốc, dựng lại dashboard và kiểm tra quyền qua mọi lối vào

How Undercroft preserves original data, rebuilds reports, and keeps every interface behind the same role gates

Dùng LLM cho phán đoán ngữ nghĩa, rồi xác minh kết quả bằng guardrail và hard gate tất định

Use LLMs for semantic judgment, then verify every result with deterministic guardrails and hard gates

Học từ cách Bun được port từ Zig sang Rust bằng 64 Claude instances song song, để thiết kế một orchestrator giữ Claude Code chạy liên tục không cần confirm giữa chừng.

Two opposing pincers that force an agent to break its own plan instead of rubber-stamping it from confirmation bias

Ymir lays down the project harness — rules, lint, CI, wiki — so the agent can own the code and you own the setup

How Kanna became my daily driver — one web UI for running Claude Code agents, subagents, and skills across projects

How Kanna's multi-window layout and push notifications let you ship while agents work, instead of watching output scroll

How Kanna funnels three wildly different coding-agent transports through a 22-line TypeScript interface

A Claude Code plugin that lays down rules, lint, CI, and an LLM-maintained wiki — never your application code

Compose processes with |, &&, ||, and & from Go — a small visitor over os/exec that returns a structured result tree.

Run parallel Claude Code agents on one repo — Docker isolation, git-as-lock coordination, and a commit guard that can't be bypassed.

A small zsh plugin that turns typing 'claude' into a project-aware --resume suggestion — pure zsh, no jq, no daemon.

A Go service layout that keeps domain logic transport-agnostic — testable with a mock repo, no database booted.