30 August 2026

From past chats

Give each connected account a written boundary

Multiple Gmail and Calendar connections make an assistant more useful, but they also make a vague request such as “check my calendar” ambiguous. Before relying on multi-account access, make a tiny account-purpose matrix: account, allowed data, default tasks, and actions that must name the account explicitly. Then try five prompts that could plausibly cross a work/personal boundary. The useful follow-up question is not merely whether the connector works, but whether a mistaken account choice is obvious before any data is disclosed or action is taken.

Turn the AI-usage debate into a reproducible notebook

The continuing interest in employee posts and external AI-usage research suggests a small project: reproduce one published chart from its released data, then keep the query and assumptions beside the result. This is a sharper way to discover whether a headline is driven by occupation labels, geography, sampling, or the distinction between augmentation and automation. A good first question is: which conclusion changes most when you choose a different unit of analysis?

From the front page

Make Markdown migrations explain themselves

For context-sensitive Markdown cleanup, have the model emit a reviewable transformation manifest rather than only rewritten files: source location, rule name, before/after snippet, and confidence. Run the mechanical cases through an AST transformer, reserve the ambiguous cases for a model, and keep a small golden corpus containing links, front matter, code fences, and prose that resembles markup. The next useful experiment is to compare how many edits from two models survive blind review, rather than comparing their prose explanations.

Add an explicit “discussion versus action” convention

Natural phrases such as “we should” can describe an idea or authorize work, and coding agents may reasonably infer the latter. Try a three-state convention in repository instructions: discuss means answer without changing files, plan means inspect and propose, and execute means implement and verify. Test it with ten real prompts—including softened phrases and mid-task course corrections—and record false starts. If it proves useful, the convention belongs in AGENTS.md, where it can travel with the repository instead of relying on memory.