Journal
Blog
What we learn building the tools, what happens when agents meet real teams, and how to be read well by search and AI answer engines.
- Building a coding-agent entry point for 128 repositoriesAX field notes: standardizing the documents a coding agent reads first across 128 legacy repositories — what we built, how we checked it, what is unproven.
- A site that search engines and AI answer engines read well — how lidgeai.com is builtHow this site is built for SEO and GEO: reciprocal hreflang, JSON-LD, llms.txt, answer-first paragraphs and a build-time checker.
- One code call instead of fifty tool calls — why we built aside-codemodeAgents burn context opening files one by one. Code mode filters inside one sandboxed JavaScript call and returns only the answer. Numbers and limits.
- Building a dev loop inside Codex with hooks and skills — Codex Meetup Seoul talkNotes from my July 2026 talk at OpenAI Codex Meetup Seoul — how a plugin that cannot touch the system prompt still runs a plan, audit, build and check loop inside Codex.
- opencodex — a local proxy that runs Codex on other modelsopencodex sits in front of Codex and Claude Code and translates requests for Claude, Gemini, Grok, DeepSeek and Ollama. How it works and when it helps.