Give Claude Code A Visual Language
On this page
- Start with the outline, not the first commit
- The outline is one section, not the whole page
- A transcript is a log, not a map
- Park the plan where an agent can read it
- Cut it into steps that can each be reviewed
- The PR body carries the artifact link
- The link points both ways
- What else goes on the page
- The PR number is the load-bearing part
- Tag it when the last step lands
- What it costs, and when not to bother
- The general version
Related Articles
The ABCDE Of Agent-Driven Development
A five-skill todo lifecycle for Claude Code — Add, Brainstorm, Code, Deliver, Explore — that carries a one-line idea to a reviewed, merged PR without letting a fire-and-forget agent make silent choices. A full tour of every stage and every pillar underneath.
Give Claude Code A Voice
An agent that works for ten minutes has exactly two moments worth interrupting you for — it finished, or it is stuck waiting on you. Two hook events, two one-second voice notes, and about six lines of settings.json. Includes both sounds, playable here, and the details that will bite you if you wire this up carelessly.
What Is An Agent Harness? The Pattern Behind Every AI Coder
A deep dive into agent harnesses — the loop that wraps an LLM and turns it into something that can read files, run tests, and ship code. We break down the six parts, the turn loop, the design tensions, and build one in 30 lines of TypeScript.
Google Gemma 4 And TurboQuant Explained For Web Developers
A deep dive into Google's Gemma 4 open model family and TurboQuant KV cache compression — what they are, how they work, and why they matter for developers building AI-powered applications.