Mathieu Bernard (@MathieuBernard_) header
Mathieu Bernard (@MathieuBernard_) profile picture

Building products & AI agents. Maker of DiscussIn. CTO at Etixio: building and leading dedicated distributed dev teams.

France · 100 % remote
Joined November 2026
30 minutes before boarding, my daughter starts talking about naval battle. she loves it. she also loses. every single time. we had a 13-hour flight ahead of us. so i opened claude and asked for something we could play on the ipad, offline, at 30,000 ft. (cheat code: i reused the voxel engine i'm building for my other games, so we started from a real 3d world, not a blank page) poof: → a little voxel living room, coffee table, a sleeping cat, a cup of coffee, dice → your fleet lies flat, the target board stands up like the real box → pass-the-ipad mode: the screen blurs on handover so nobody peeks → a computer opponent with 3 levels, from "deckhand" to "admiral" → ships that burn when they sink → a cozy little soundtrack, generated in code (no audio files) → installs on the home screen, works fully offline 13 hours sorted. thanks claude 😉 free and open source, link below 👇
1
47
My LinkedIn inbox has a license dashboard now. It took 200+ commits to get here, and I’m only just starting to see where people drop out of the buying funnel.
5
😂
Senior Engineer: i have 8 years of experience. Kid + Claude: i have Claude.
4
AI can write the feature. It can’t decide what “good” means for your product. That still takes a human who knows the user, the constraints, and what trade-offs are acceptable. The bottleneck is moving from typing code to making better decisions, not removing people from the loop.
5
“Runs on 16GB” is a great headline. It’s not a quality claim. For local models, I want three numbers next to the memory figure: quantization, tokens per second, and task performance on the work I actually do. Otherwise we’re comparing laptop feasibility, not whether it can replace a hosted model.
7
A coding agent can finish a task in 5 minutes and still leave you waiting 40 minutes for CI. That changes what “fast development” means. The next bottleneck isn’t typing speed; it’s how quickly you can get trustworthy feedback. I’d optimize the loop before adding another agent: cached dependencies, targeted tests on each change, and the full suite before merge.
3
1
3
44
The best AI workflow isn’t the one with the most agents. It’s the one where each agent has a clear job, a bounded workspace, and a check that can prove the job is done. Otherwise you’ve just added more ways to be confidently wrong.
1
20
the agentic stack in 2026: my subagents spawn subagents, and i haven’t closed a for-loop by hand in months. my actual job now is reviewing diffs and stopping them before they refactor the entire repo at 2am because one test was red.
19
play it here 👇 two players on one device, or vs the computer mathieubernard.github.io/tab… on ipad: share → "add to home screen", open it once while online, and it works for the whole flight. code's open if you want to peek: github.com/mathieubernard/ta…
22