9 docs tagged with "ai"

Running a coding agent in a container solves the trust problem by giving up the machine. If you want to keep the machine, the permission allowlist is where you make the tradeoff instead.

Told that a repeated timestamp proves the polling interval is shorter than the source's refresh interval, the agent concluded the gaps it had already collected were therefore the refresh interval. They were not — gaps measured while undersampling can be several refreshes collapsed into one, and no later evidence un-collapses them.

A schema field named defaultToReminders was read as proof that YouTube and TikTok require manual publishing. Its value was false on both channels. Ten posts were configured to send phone notifications instead of publishing.

Asked to capture design notes for a scanner, the agent filed them under tools/. That directory is only for portfolio-site automation — a convention that existed nowhere in the repo, because the README meant to state it was empty.

Asked to bring internal docs up to date, the agent read a 15-month-old planning doc and declared the feature unbuilt. It had shipped months earlier — including the exact session-tracking code the doc sketched.

How local LLMs make 'just add AI' no longer a joke, with a one-minute Ollama setup and reflections on environment mastery.

An article generated with an LLM recommended a command as the canonical way to list kitty keybindings. The command returns nothing, and always did. It sat published for months on one of my better-performing pages.

Two lines of real terminal output, then a third that was never printed by anything. It was heading into a video whose entire premise was that the output is real.

A running case file of assumptions a coding agent made that didn't hold — recorded in a fixed format so the collection can eventually answer when a human actually needs to be in the loop.