Notes
Numbers I measured while building tools, the options I passed on and why, and the attempts that failed. Every note comes from real work logs.
I compared AV1 and H.264 while shrinking the 36.9MB video on the PromptFilm page. AV1 scored higher on quality, but 284 of 439 faint stars disappeared.
In a gallery with 400 animated elements on one page, the reels view and the prompt dialog stuttered. How I found the causes with a Chrome performance trace and cut 283ms to 33–50ms.
The UX/UI intro page requested 200 small JPGs one by one on every visit. Bundling them into 4 files of 50 images each cut more than 200 requests to this site.
With Lotto Lab's draw engine, I ran a 1,000-game experiment 20,000 times, 20 million games in all. The jackpot came up twice, and every ₩1 million spent returned ₩180,000 on average in 4th and 5th prizes.
Plain messages sent while Claude Code is working slip in between tool calls. Slash commands wait until the turn ends. I used that difference to build a one-line skill and verified it with real sessions.
I read Codex session logs directly to measure where the context goes. About 33,000 tokens before you type anything, tool output over half during work, and the place to trim was not where I expected.
Codex gives the skill list only 2% of the context. With many skills, the end of each description gets cut, and the model never reads when to use a skill. What I found with Astra Xray and how I cleaned it up.
DELETE-DENIED is a Codex PreToolUse hook that checks terminal delete commands right before they run. Why I chose this approach, and what it does and does not block.
PromptFilm goes through research, storyboard, production and review to make a 3D film that plays in the browser. Using three films about space, a GPU and a movie scene, here are the safeguards I built into the film code so the numbers on screen stay right.
AI Chat Exporter builds a handoff document by picking original sentences from the chat with rules alone, no LLM. I compared methods on 27 real chats and raised the accuracy of an AI given only the document from 78.8% to 97.6%.