Tested on real work
An assistant given a real engineering task, and where it failed.
6 pieces, oldest first. [all series]
- 2025-11-06 essayChecking the numbers an assistant reads out of a long PDF I gave Claude a long technical document of my own and it reported a number I still have to check, and later I caught things missing from what it put together from the same material.
- 2026-02-02 essayTelling an assistant to stop agreeing Over about a week I had to tell Claude several times to stop agreeing with me, and these are the things I told it: search and argue, here is what the term means, ask me questions first.
- 2026-02-08 essayAsk where the numbers came from before charting them An assistant's first analysis of chemical engineering employment came with data it had made up, so I pulled the real series from the BLS and Census sources myself, and that took the evening.
- 2026-02-15 essayWhat the DOE generative AI policy asks of the person using the tool I read the US Department of Energy policy on generative AI to write a short post about it, and this is what it says about the user, where my own summary was wider than the text, and why I think others will adopt it.
- 2026-10-03 essayTesting a skill with planted errors and a run without the skill Claude built a unit checking skill for me and tested it on a calculation sheet with four planted errors, on two models, with and without the skill.
- 2026-10-04 tilTelling a skill exactly which files it may read My literature note skill in Claude Code was reading through my notes vault for every paper, so I had it rewritten to read only its instructions, the template and the PDF text.