I'm Yash, a product manager who built an entire PM operating system inside Claude Code: personas that pressure-test every spec, persistent product context, and eval pipelines that run on real call data. It turns a rough idea into reviewed, validated, working software: not slide decks.
My Claude Code workspace isn't a chat window. It's a structured repository (ground-truth context, role personas, analysis scripts, cross-session memory) that every piece of work runs through. Same source of truth, same constraints, compounding for six months.
Every session reads the product context, dealer-economics model, and domain knowledge first. No hallucinated capabilities.
6 personas modelled on real people pressure-test every spec before a human sees it.
7 Python scripts turn 450+ weekly calls into intent-level scorecards and ROI-tied findings.
Decisions, open items, feedback rules carry across sessions. I never rebuild context from scratch.
Pick a real spec I wrote. Watch the stakeholders tear into it, before a single human reviewer saw the doc.
These are the actual objections my persona files raised, pulled from real sessions: not a live model talking. On the follow-up spec, Mike caught the missing hot-lead handoff; on the CRM spec, Sahil's persona flagged the null-handling and write-back risks before the real Sahil reviewed it, so the human review focused on architecture, not basic gaps.
Three products, built end-to-end with the system above. Each one deployed, each one validated with real users: not mocked up.
Email, Slack, support tickets, meetings, a ten-tab planning sheet: most of my day was spent scanning all of it just to find the few things that actually needed me. Otto is a scheduled agent that reads every channel, surfaces only what needs me, drafts replies in my own voice, runs my weekly plan, and learns from every correction I make. It never sends on its own: every outbound waits for my one click.
As we scaled Spyne's AI sales agent to 30+ dealerships, the ROI was invisible to dealers, making renewals harder and expansion conversations vague. I built the dealer-facing console myself (Overview, Action Items, Appointments, Leads) and validated the designs directly with dealer GMs and BDC managers.
Not a transcript read aloud: a two-person show. The book's author "joins" as a guest, and an AI host interviews them, with every question drawn from your role and what you need from the book. NotebookLM and the rest generate the same episode for everyone; this one is a conversation shaped around you.
The output is good because the rules are enforced, not suggested. A few that run on every session:
Straight from our CEO's framework: a car on the lot costs ~$50/day. If a feature doesn't cut days-in-inventory or protect gross margin, it doesn't ship. Claude knows this before I say a word.
Plain language, active voice, no filler. "3 days → 4 hours", never "significant improvement." I built a /humanize command that rewrites any doc to the style guide.
Scores agent calls across 6 dimensions (information accuracy, tool usage, conversation quality, escalation, conversion) with Python to aggregate and compare against human notes.
Every requirement pushes to Jira and Notion with a sync map tracking each task. Claude flags any spec drafted but not yet pushed to both before a session ends.