Fractional CMO. Building tools with obsessive attention to craft and details. Onadit, Juby and more.
Paris, France
Joined May 2011
- Tweets445
- Following97
- Followers386
- Likes1K
Pinned Tweet
Launched my first product today!
AI labs drop a new model almost every 2 weeks.
To switch between them (or between sub-agents) without re-explaining everything, you need memory that travels with you.
Meet Onadit: local memory on your Mac.
Conversations get saved, split and indexed (keywords + embeddings). Your agents (Claude Code, Codex, Cursor… via MCP) can pull from it.
Ask “what stack did I choose for the content pipeline, no infra?”
→ finds: “No database. No server. No build step. Just markdown, JSON, and Claude”
→ from ~63k tokens to 820 (imagine the savings + efficiency when agents work in chain)
onadit.com (first 7 days free)
First launch so would mean a lot if you try it or share it 🙏
Everyone here is building apps (me too)!
but almost nobody talks about how far you can scale as a freelancer with AI and the right setup. 100% you make money faster than with apps.
Crossed €200k in October, took all of August off with my 4 kids, and went skiing last January 😊
when did Northwind become the new Acme?
I just learned it's a 1990's fictional business from Microsoft but it seems lately models use it more and more
genuinely asking
you now can use google analytics mcp in claude
setup in claude: 1 min
setup Google Auth: 12 steps (good luck)
feels good not to have to fight GA4 UI anymore
MCP repo github.com/googleanalytics/g…
claude-copy just dropped 🔥
Instant keyboard shortcuts to copy ANY Claude Code output to clipboard. No selecting text, no scrolling. Just Cmd+Shift+C/P/A in Kitty or iTerm2.
Works for questions/plans/normal output. Perfect for pasting into ChatGPT, Gemini, Grok or whoever you trust for a second opinion.
Try it: github.com/clementrog/claude…
It works like a charm paired with @raycast snippets!
Even faster double-model workflow unlocked.
Star the repo if it saves you time ✌️
Everyone thinks Claude Opus is a cheat. It’s good, for sure.
But the actual cheat is: your favorite model’s output + sharp feedback from a second model.
Double-pass feels tedious at first…but delivers ~5× quality and slashes revisions dramatically
1/ I was burning through €100+/month on AI-assisted coding. Most of it on tasks that didn't need frontier-level reasoning.
So I built a system that uses Claude for thinking and Cursor for typing. Costs ~€40/month now. Here's how!
7/ And to me it feels promising because it reduces Opus 4.6 context (needs less compacting, more token efficient ofc), Composer 1.5 from Cursor is fast (and Cursor indexing works through CLI).
I almost code unlimited for 40€/month without noticeable quality drifts.
8/ "Separate reasoning from execution, get same results" is close to reality, especially with Cursor new Composer 1.5!
If you're one of ~2K Pilot users (my previous workflow) i'm sure you'll like this 3x more.
Repo: github.com/clementrog/claude…
Still prototyping layouts in plain HTML + CSS.
No server, no build. Just refresh.
Debug borders + tinted backgrounds until alignment is locked. Then port to real code.
Feels old school but nothing's faster for iteration.
Anyone else work this way?
Did Anthropic just quietly break OpenCode for Claude Pro/Max? Getting “Claude Code only” on every request (worked yesterday).
Anyone managed to make it work again?
Pilot v3.1.0 — stricter by design.
Git diff required on ALL verifies. Evidence mandatory. Secret reads = hard violation. Less trust, more proof.
Also in v3: batch mode, design contracts, skills system.
github.com/clementrog/pilot
