When managers introduce AI into their routines, three symptoms show up regularly: reaching for summarization first; covering too much ground at once; one-shotting the whole task then endlessly refining the prompt. All three share a root problem. They skip the work of building context and jump to an answer.
Summarizing is deciding.
When a manager is facing down the long road of ingesting volumes of raw material (Slack threads, peer feedback pulls, 1on1 conversation notes, etc.) the instinct is to ask AI to condense it. Condensing the raw material is judgment, and that’s your job. You can’t audit what got dropped or disagree with what was prioritized. Just as I argued in The Transcript Is The Work, the raw evidence is the asset and everything derived from it is disposable. People default to summarizing because they’re trying to get the whole job out of one prompt.
Engineer your prompt. People think this means being clever with role-based prompting “pretend you’re a …” or they think it means stuffing as much context as possible into the prompt. It’s neither; it’s about how you build the thing. Treat it like feature development with lots of milestones, tests and iteration on increments before assembling. Don’t try to zero to one it.
Try this:
Frequent snapshots along the way - At short intervals, use AI to pull from Slack, PRs, Docs and your notes into one place, verbatim. Zero summarization or curation. Think weekly 1on1 preps.
Roll snapshots into longer horizons - still no summarization, interpretation, or rating. You’re building evidence logs, not intermediate judgments. Quarterly development check-ins.
You annotate, comment, edit and enhance these evidence logs.
AI updates its understanding from your log edits and highlights threads it couldn’t properly trace. You fill the gap, direct it at a source or note why thin evidence is fine there.
Guide the evidence into your format - competency maps, reflection templates, scope, impact, complexity, etc.
Mandate a “not included” section - the output plus not-included should be equivalent to the log. If it isn’t, something was lost along the way. That’s your audit.
Interrogate AI along the way. Where did it struggle? Where couldn’t it find evidence? Where does it need organization guidance? Where is it least confident? This is how you check its work.
The shaped evidence log is now your first-order artifact. You’ve spent this entire process building something you (and AI) haven’t written a word of feedback from yet and that’s the point.
Next: you write it and AI judges it.