Nobody announces that a project is going wrong. What actually happens is quieter and slower: three weeks ago the update said “done Thursday.” Two weeks ago it said “on track.” Last week it said “should be fine.” This week it says “hopefully by end of sprint.” Nothing in that sequence is alarming. Every single update, read on its own, is a perfectly normal thing for a competent person to write. And that’s exactly the problem — the signal doesn’t live in any one update. It lives in the slope, and you only see a slope if something is holding all four updates side by side and measuring the delta between them.
No human does that. You read this week’s update against your memory of last week’s, which is already fuzzy, across six people, while doing your actual job. The drift is invisible not because it’s subtle but because comparing text to text over time is the specific thing human attention is worst at — and the specific thing an automated pass is best at.
This tutorial builds a monitor that scores every status update as it arrives, keeps a rolling history per person and per workstream, and has Claude analyze the trajectory to flag softening language weeks before it becomes a missed date. It also, importantly, shows you how to build this without turning it into surveillance — because the naive version of this idea is genuinely creepy, and the difference is in the design, not the intent.
Score the language of every update as it arrives, then measure each person’s trajectory against their own baseline — because the signal is in the drift, not the words.
Two halves matter there. Trajectory is why this is different from sentiment analysis. Their own baseline is why it works at all — and it’s the part almost everyone gets wrong on the first attempt.
The intuitive build: run sentiment analysis on every status update, alert when sentiment drops. Try it and you will discover, within about two weeks, that you have built a machine that identifies which of your teammates writes cautiously.
Some people hedge constantly and always deliver. Some people write “all good 👍” right up until the morning they miss the deadline by three weeks. An absolute confidence score tells you about someone’s prose style, not their project. Rank your team by hedging level and you’ll get a list of your most careful writers, which is useless — arguably worse than useless, because it looks like insight.
The fix is the whole trick: compare each person only to themselves. A chronic hedger who hedges is at baseline — no signal. A chronic hedger who hedges more than they normally do is a strong signal precisely because you know their normal. The person who writes in confident absolutes and suddenly adds a qualifier has just moved a long way from their baseline, and that tiny change means more than another person’s paragraph of caveats.
The unit of measurement is deviation from self, never comparison across people. Build it any other way and you’ve built a personality detector.
Confidence alone is too blunt. Track two things per update:
Axis 1 — Confidence. How committed is the language? “Done Thursday” is high. “Should land this week” is medium. “Hoping to get to it” is low. This is the obvious axis.
Axis 2 — Specificity. How concrete is it? Does the update name a date, a number, a next action, a blocker? Or has it gone abstract — “making progress on the integration”?