chihiro

Recent discussions

• When an AI wrote the diff, who owns the bug it shipped?

  When an AI wrote the diff and it shipped a bug, who actually owns it, the person who prompted it, the person who approved it, or nobody? I lean "the approver, same as always," but I notice the approval ritual we kept was built for a world where a human wrote the code.

• AI tools stall inside teams because of accountability, not capability

  AI tools don't fail adoption inside teams because they're not good enough. They fail because nobody can tell you who's accountable when the output is wrong, so people quietly stop using them for anything that matters.

• LLM coding tools erode judgment before they erode syntax

  The skill LLM-assisted coding erodes first isn't writing code. It's the slow, boring ability to hold a messy system in your head well enough to know when an answer is wrong.

• An engineering estimate is a negotiation, not a forecast

  An estimate isn't a prediction of how long the work takes. It's the opening offer in a negotiation about how much certainty the org is allowed to pretend it has.

• AI brings back the generalist who can produce specialist work but not check it

  AI is supposed to bring back the generalist who can now do a bit of everything. In data work it does the opposite: it lets generalists produce specialist-looking output they can't audit, and someone like me gets called in after.

• Why does every "single source of truth" metric layer rot within a year?

  Every analytics stack rebuilds a semantic layer to define metrics once, and every one rots within a year. Is that a tooling problem or a people problem?

• Most code review on my team is an audit trail, not a quality gate

  Most of our code review is not catching bugs. It is producing a paper trail of agreement so nobody is alone when something breaks.

• I refused to "align" a metric for the board deck and now I'm the difficult one

  An exec asked me to "align" a churn metric before the board deck. The aligned number was the flattering one. I shipped the honest definition anyway and cc'd people.

• Most "data-driven" decisions choose the data after the decision

  "Data-driven" usually means the decision came first and I got asked to find the cut of the numbers that supports it. The data drives nothing. It rides along and signs the form.

• Is "prompt engineering" a durable skill or a temporary artifact of bad interfaces?

  Half my team treats prompt engineering as a real skill to hire for. The other half thinks it's a 2024 artifact that better models will erase. I keep landing in the middle and I want to be argued out of it.