Outreach to dev-tool companies: 11 of 20 contacted so far (today: CrewAI, E2B, Apify, Rivet, OpenCode). To find the rest, I screened 6 more libraries: Claude Code, Codex and Gemini, 2 runs each, against a live server where the tool needs one. 3 had a finding (agents silently using a removed API, or switching auth off to get past a 401). In the other 3, all 6 runs got it right. Drafts for the last 9 are ready, so 20 is in reach.

Opened a PR to LadybugDB's agent skill, helping them shape their docs for coding agents: github.com/LadybugDB/ladybug-… I ran Claude Code, Codex and Gemini on migrating a real project from Kùzu to LadybugDB, with and without the skill, then fixed the skill where the agents went wrong. Original data kept: 2/6 runs before, 6/6 after. Agents stopped crashing on Kùzu and LadybugDB in one process: 0/6 → 6/6 runs. #worklore

Published a new article on Habr. I keep researching where AI coding agents get stuck on programming tasks, and how a short documentation note gets them unstuck. This time I compared Russian and English prompts over ~1,350 runs. With Russian prompts, Codex and Kimi K3.1 stopped searching the web, so they missed fixes that exist only on forums. A note with the missing fact brought success close to 100% in both languages. habr.com/en/articles/1106298/ #worklore

started an experiment for worklore: take a dev-tool that just shipped a major version, run its quickstart through claude code, codex and gemini (6 runs, a hidden script checks the result, not the agent's word), and turn whatever breaks into docs patches. if it works, this is how i'll offer agent-ready docs to saas companies #worklore

Downgrade codex

Good day off. Blogged a little and let Codex run full auto all day on two projects #research

start writing blog but impossible with Codex asking me a stupid question every 2 minutes #blog2

🤖I wasn't using enough of my weekly Codex plan, so I have it take Rust and rebuild VB6/VB7 for me to build mobile apps with Rust. #research

set up telegram notification for gh actions all by codex

80 issues all being worked about 7 at a time. Used a full week of codex 5x

Went from 50% to 0% codex weekly quota. WTF

thanks to codex and some debugging k8s is up with kubespray, finally

ask codex to create a design-system page to see how everything looks like #buildto5000

Hit Codex 5 hr limit twice today

Make monitoring app because Codex is taking too much disk space

read Jef Geerling ansible roles to figure out include_vars is used in real world (thanks to codex)

Improving AgentControl to work better with multi Codex accounts

Record videos to teach my non-tech friends how to set up Claude Code, Codex and Pi, and about AI basics

Try out DeepSeek since I'm out of usage for Claude and Codex. So far, it's ok the price.

I don't like this, but the reality is the Codex reset helps me speed up my work.