Why I'm Building PM in the Loop (and My First Agent Skill)
I’ve spent over a decade as a PM across Cloudera and Dell, most recently running a data services platform, AI inference products, and partner ecosystem work. Somewhere in the last two years, the job changed under my feet. It’s not enough anymore to have a point of view on AI — you need to actually be able to build with it. Wire up an agent, write a prompt that survives contact with real data, know when to trust the output and when not to.
Most AI-for-PM content I read falls into one of two buckets: high-level strategy decks with no hands, or engineering tutorials that assume you already live in a terminal. I wanted something in between — a place where I write down the actual skills I’m building, week by week, as I build them. Equal parts teaching and notes-to-self.
That’s what this site is. One post a week. Every post ends with something you can copy and use today.
The first skill: turning messy research into a synthesized insight doc
Here’s a task every PM has done badly at least once: you’ve got a stack of user interview notes, some support tickets, maybe a survey export, and you need a structured findings doc by end of day. The old way is skimming everything twice and writing bullet points from memory — which quietly drops the outlier finding that actually mattered.
Here’s the prompt I use to have an AI agent do the first pass, so my time goes to judgment instead of transcription:
You are helping me synthesize raw user research into a structured findings doc.
Input: [paste interview notes / ticket exports / survey responses]
Do the following:
1. Extract every distinct theme mentioned, however small.
2. For each theme, count how many sources mentioned it and quote the
sharpest verbatim example.
3. Rank themes by (frequency x severity), not just frequency — a theme
mentioned once by an enterprise customer about a blocking issue outranks
a theme mentioned five times as a minor annoyance.
4. Flag anything that contradicts our current roadmap assumptions.
5. Output as a table: Theme | Frequency | Severity | Example quote |
Suggested action.
Do not soften or round up sentiment. If the data is thin on a theme, say so
instead of extrapolating.
What this buys you: a first draft in minutes instead of an afternoon, with the tedious extraction done consistently. What it doesn’t buy you: judgment on what’s actually worth acting on. I still read every quote myself before it goes near a roadmap doc — the agent is a very fast intern, not a replacement for having sat in the interviews.
Two things I’ve learned running this repeatedly:
Severity beats frequency, and you have to say so explicitly. Left alone, these tools default to “most mentioned = most important,” which is exactly backwards for the finding that should scare you.
Ask it to flag contradictions to your existing beliefs, not just confirm them. Without that instruction, synthesis tools tend to hand you back a fancier version of what you already thought going in.
What’s next
Next week: building a weekly metrics review agent — the prompt and workflow I use to turn a raw metrics dump into a scorecard with recommended actions before Monday standup.
If you’re a PM building with AI agents and want to compare notes, this is exactly the kind of thing I want to hear about. See you next week.