← Back to blog AI PO

The AI Product Owner's tech stack in 2025

June 2, 20258 min

People often ask me what tools I use daily. Here’s my complete stack, with the reasons behind each choice.

For working with LLMs

Claude Code — My main tool for anything code-related and codebase exploration. The quality of instruction following and large context window make it my first choice.

Claude.ai — For long conversations, spec writing, brainstorming. I use projects to maintain context between sessions.

ChatGPT — For tasks where I need image generation (DALL-E), or when a client is in the OpenAI ecosystem and I need to test in their environment.

Mistral Le Chat — For projects with data privacy constraints. European hosting, useful for certain clients.

For agent prototyping

n8n (self-hosted) — For simple to moderate agent workflows. Visual interface, ready-made integrations, rapid deployment. Ideal for validating a use case before coding.

LangChain / LangGraph — When you need more complex agents with state and loops. More code, more flexibility.

Claude Code + MCP — For prototypes that need to interact with real systems from the start. The most powerful combination I’ve found.

For project management

Linear — For task and sprint tracking. Clean interface, GitHub integrations, smart notifications.

Notion — For agent documentation: system prompts, data schemas, architecture decisions, evaluation results.

Loom — For sharing async demos with clients. Much more effective than a meeting for showing “how the agent behaves in this case.”

For evaluation

Braintrust — LLM evaluation platform. Lets you define test datasets, run automated evaluations, compare prompt versions.

Google Sheets — For manual evaluations on small datasets. Simple, shareable, sufficient for 50-100 test cases.

What I’ve stopped using

Consumer AI no-code platforms — Seductive at first, but limitations appear quickly once you go beyond simple use cases.

Proprietary LLM wrappers — Several startups offer their “abstraction layer” on top of LLMs. In practice, this adds complexity and dependency without real added value for most projects.

My rule: if I can do the same thing directly with the LLM API in 2 hours, I don’t pay for a wrapper.

SC

Stéphanie Caumont

AI Product Owner · Learn more