Every "AI tools for PMs" list I've read is either a recycled list of ChatGPT, Notion AI, and Midjourney — or a VC's wish list of enterprise software. Neither is useful.
This is the actual stack I use, updated as of mid-2025, with specific reasons why each tool is in and what I tried before that didn't make the cut.
The framing: three layers
I think about my stack in three layers:
- Understanding — research, synthesis, thinking
- Building — designing and shipping AI-powered products
- Operating — running experiments, measuring, iterating
Most "AI PM tools" lists only cover layer 1. If you're actually building AI products — not just using AI for note-taking — you need all three.
Layer 1: Understanding
Claude (Anthropic) — Primary LLM
Best instruction-following of any model I've tested, especially for structured output. When I need a PRD section, competitive analysis in a specific format, or user stories that follow a schema, Claude returns it clean.
How I actually use it:
- First-draft PRD sections from bullet notes
- Synthesizing user research transcripts into themes
- Stress-testing product reasoning ("argue against this feature decision")
- Generating edge cases I haven't thought of
Perplexity — Research
Real-time web search + synthesis. For competitive research, market sizing checks, and "what's the current state of [technology]" questions — faster than building a search → read → synthesize loop manually.
NotebookLM — Document synthesis
When I have 10 PDFs or product docs and need to ask questions across all of them, NotebookLM is unmatched. I load a competitor's entire documentation site and interrogate it.
Layer 2: Building
Claude API — LLM backbone
When I'm building something like InstantPlan, Claude is the backbone. I default to Claude Sonnet for most tasks, Claude Opus for deep reasoning or evaluation. I avoid using the most powerful model by default — cost compounds fast.
n8n — Agent orchestration and automation
The best non-code-heavy way to build multi-agent pipelines. I use n8n to orchestrate LLM workflows: trigger → fetch data → process with LLM → format output → deliver.
What I've built with it:
- Content repurposing pipeline: LinkedIn post → email → blog outline
- Lead research automation for Zerton
- Listen2RE content ingestion and processing pipeline
Why not Zapier or Make: n8n is self-hostable, has direct HTTP request nodes, and LLM integration is first-class. For serious automation, it's the better choice.
Cursor / Claude Code — Implementation
I'm a PM, not an engineer. But I can direct AI to implement things I've designed. Cursor with Claude lets me build production Next.js code through conversation. Claude Code goes further — it can execute, test, and iterate autonomously.
I review everything. I understand the structure even if I couldn't write it from scratch.
Vercel — Deployment
Zero-config deployment for Next.js. Push to GitHub, site is live. Preview deployments per branch are invaluable for testing before shipping.
Supabase — Database for AI apps
Postgres with a generous free tier, built-in auth, and a REST API out of the box. For AI products that need to store conversation history or evaluation results, Supabase is the fastest path to production.
Layer 3: Operating
PostHog — Product analytics
Open source, self-hostable, generous free tier. Session recording is the fastest way to understand where users get confused.
What I track for AI features:
- Generation trigger rate (how often the AI feature is used)
- Acceptance rate (output kept vs regenerated)
- Time-to-first-value
- Drop-off point
LLM-as-a-Judge (custom) — AI output quality
For AI products, you need a way to evaluate output quality at scale. One LLM generates output, another evaluates it against criteria. This is how I track quality without manual review of every response.
Evaluation dimensions I use:
- Relevance
- Completeness
- Format adherence
- Hallucination risk
What's NOT in my stack (and why)
GitHub Copilot — I use Claude Code instead. Better for my workflow.
Notion AI — I use Notion for docs but write with Claude in a separate tab and paste results in. Native Notion AI doesn't match Claude's output quality.
Jasper / Copy.ai — Outdated category. Claude does everything these do, better.
LangChain — Too heavy for most of what I build. n8n covers 80% of my orchestration needs without the Python complexity.
The one-sentence version
Claude for thinking, n8n for orchestration, Cursor/Claude Code for building, Vercel for shipping, PostHog for learning.
Everything else is context-specific.
Sujit Chankhore is an AI Product Manager and founder based in Pune, India. Open to Senior AI PM roles globally.