Product management · how I work
How I use AI across the product lifecycle
Honestly, it’s woven through my whole week now — discovery, PRDs, prototyping, and a lot of the grind I used to just push through. Three modes, really: a thinking partner, a build tool, and a co-worker that does the clicking.
Thinking partner
Synthesising research, drafting PRDs and stories, poking holes in my own edge cases, reading impact data, getting stakeholder comms to land.
Build tool
Clickable prototypes for concept validation, and scrappy local tools that hit other teams’ and third-party APIs.
Co-worker
Walking through user flows, scraping, and the one-off data entry that discovery and sanity checks always seem to need.
I’ve kept everything here generic — no real product or project details.
Foundation
A private skill that actually knows our system
This is the thing everything else leans on.
Before I do any discovery or spec work, I load a private Claude skill I built and keep current. It’s basically our system’s context in one place — architecture, service boundaries, the data model, and the constraints and tech debt that never make it into a Confluence diagram.
So when I’m writing acceptance criteria or reasoning about a flow, it’s anchored to how the system actually behaves rather than my best guess — and it pushes back when I describe something the architecture wouldn’t actually support.
The loop
From a friction point to a measured outcome
Roughly the path I run for each improvement — discovery through to a measured result. Open a step for where AI fits and what it produces.
The work
I start on the quant side. I instrument my own events through GTM, then read the behaviour across custom reports in GA, Looker / Data Studio, or the CRM, and in Clarity — session recordings, heatmaps, rage-clicks, funnel drop-off. Looking for where users hesitate, backtrack, or bail, and trying to tell real friction apart from noise.
Where AI comes in
Claude helps me shape the GTM tags and the GA / Looker report definitions, and Clarity’s built-in Copilot summarises the recordings and surfaces behaviour patterns I’d take hours to spot by hand. I use all of it to shortlist the frictions worth taking to the backlog.
The work
Getting it to something engineering can pick up — problem statement, user stories, acceptance criteria, the non-functional bits, and the edge cases. All written against the system skill so it lines up with how things really work.
Where AI comes in
Claude summarises the research, gives me a first draft of the PRD, and is genuinely good at enumerating states and failure modes — the stuff that otherwise surfaces mid-sprint or in QA.
The work
Getting the critical items — decisions, risks, trade-offs — in front of leadership in a form they can actually action, and pulling the right inputs back without a week of back-and-forth.
Where AI comes in
I draft the update, then run it through the Gemini browser extension to tighten it — cut the throat-clearing, lead with the ask. The escalation reads like an escalation, and I get a decision back instead of three clarifying questions.
The work
Getting the flow concrete enough to react to. It’s far easier to align people on something they can click through than on a Figma frame or a wall of text.
Where AI comes in
Claude builds a working prototype straight from the spec — fast enough that it’s disposable, so I’m not precious about binning v1 after the first round of feedback.
The work
Pressure-testing the experience against how other teams’ services and third-party APIs actually behave — in a lower environment, before it hits grooming or a tech review.
Where AI comes in
Claude builds me small local tools that call those APIs directly, so I can see the real responses, error states, and latency. I tighten the flow against that, and walk into tech and QA conversations already knowing which questions are coming.
The work
After rollout, checking whether the change did what the PRD said it would — pre / post on the target metric, the guardrail metrics, and the segments that matter. Not just “did it ship”.
Where AI comes in
Claude helps me build and read the GA, Looker / Data Studio, and Tableau views, compare pre and post cleanly, and write the readout so the outcome — or the miss — is clear to stakeholders.
Alongside
Always on, not once a cycle
Two things that just run in the background of everything else.
Tracking my own progress
Claude is hooked into my Jira board and a tracking sheet. It pulls ticket status, flags what’s moved and what’s been sitting, and drafts my standup notes — so I’m reporting progress at 9am, not reconstructing it.
Claude + Jira + Google SheetsAutomating the busywork
Claude for Chrome in co-worker mode takes the repetitive stuff off my plate — clicking through user flows to check they still work, scraping data I need for research, and the one-off data entry that testing and sanity checks always seem to require.
Claude for Chrome · co-worker modeGuardrails
How I keep it honest
A few rules so this stays useful and doesn’t bite me later.
The system skill keeps specs and analysis anchored to how the product actually works — and Claude calls it out when I get a detail wrong.
Numbers, claims, edge cases — checked against the source before they go into a PRD or in front of leadership. AI is confidently wrong often enough that this isn’t optional.
No customer data, no PII, nothing unreleased goes into a prompt. I work on the shape of the problem, not the raw records.
Stack
The stack
What each one is actually for.
Claude
My main workhorse — research synthesis, PRDs, prototypes, local tooling. Runs the private system skill and is connected to Jira and a tracking sheet.
Claude for Chrome
Co-worker mode for the browser work — flow checks, scraping, one-off data entry.
GA + GTM
My own product-event tracking — GTM to instrument events, GA or the CRM for custom funnel and behaviour reports.
Looker / Data Studio · Tableau
Dashboards and impact reporting — measuring whether a rolled-out change actually moved the metric.
Microsoft Clarity + Copilot
The qualitative side of discovery — recordings, heatmaps, rage-clicks. The built-in Copilot summarises sessions and surfaces the behaviour patterns.
Gemini
Browser extension I use to sharpen stakeholder and leadership comms before they go out.
Different tools, same habit: start from how the system actually works, verify what matters, make the call myself.
The line
What I don’t hand over
The parts of the job AI doesn’t do — and that I don’t want it to.
Discovery will hand me twenty problems. Knowing which one moves the north-star metric is the job.
When to escalate, who needs to be in the pre-read, how hard to push on a trade-off before it costs me capital.
“The model suggested it” isn’t a line I get to use in a retro.
Prioritisation only means something when someone’s accountable for what didn’t make the cut.