Iteration speed · Kill your darlings · Ecosystem emergence
What you thought was a side project was actually a laboratory.
SCROLL TO START
The 'right way' isn't always the right way. MCP was well-documented, standard protocol — everything a good engineer would reach for. 5 days of rigid callbacks, opaque errors, zero flexibility.
“The right way for others isn't always the right way for your problem.”
MCP
Google Workspace
Python
direct
“The moment you kill the wrong bet, the timeline resets.”
Decision in one commit. Complete migration to Python direct. Everything accelerates immediately — first new skills in 2 days.
cfa0ce3: Remove linkedin-jobs skill — not used and ineffective
4 days of work · Deleted in 1 commit
“No sunk cost fallacy. A feature that doesn't work → delete it.”
389 entries · 52 days of active intelligence
X / Twitter
214
GitHub repos
52
Total entries
389
Not just bookmarking — active competitive intelligence. Every architectural decision validated or invalidated in real time by what others were building.
“Your competitive advantage is how fast you learn, not how fast you code.”
FIRST SESSION
6:34 AM
Telegram bot live. Your assistant is no longer tied to a chair. Mobile, in bed, at 6:34am before the first coffee.
“Constraint removal > feature addition.”
BEFORE
Code first. Regrets later.
AFTER
Spec first. Code after.
259 commits in 17 days. But more importantly: the shift from 'code first, regrets later' to 'spec first, code after'. linkedin-jobs would have been avoided. The 4 bot refactors would have been 1.
The system is stable. It can now handle complex requests without breaking.
REMOVED
ADDED
“Every deletion is a statement about what matters.”
February 28 — the most intense non-debugging day. Three new systems born in parallel, all from real needs identified during 2ndBrain.
voiceCLI
TTS + STT CLI — Qwen3, Chatterbox, Whisper. Born from Telegram's need for voice.
roxabi-plugins
Open-source marketplace — 1 → 11 plugins in 4 days.
roxabi-claude-config
Shared config + alias lyra='ssh user@machine'
“Ecosystems don't emerge from planning. They emerge from compounding.”
BEFORE
~15s
~15 seconds / message
AFTER (qwen-fast)
~2-3s
~2-3 seconds / message
First commit: fix PyTorch for RTX 5070 Ti (Blackwell). By end of day: voice generation viable for real-time Telegram conversation.
“The right hardware investment pays off in one day.”
“The first time you ask your creation to explain itself — that's when it becomes real.”
TESTED
Solene
THE ANSWER
Lyra
The free electron — non liée à un atome fixe.
Solene appeared in a midnight voice brainstorm. She lasted a few hours. Lyra was the answer that had been there all along.
“The name is the concept. Get the concept right.”
6 repos. All feeding each other. None of it was planned.
“You don't build an ecosystem. You build things that deserve each other.”
462
Commits (2ndBrain)
389
Knowledge entries
194
Bot sessions
52
Days to ecosystem
11
Open-source plugins
6
Active repos
“Each number is a decision made, not a feature shipped.”
March 1: first commit. March 4: SimpleLyraAgent receives first end-to-end message. Hub, 2 channel adapters, CI, 8 ADRs, full documentation.
“52 days of learnings compressed into 4 days of build. That's what a laboratory is for.”
“2ndBrain was a laboratory. Everything else is the engine.”
Every 'mistake' was information. Every refactor was a discovery. The debt wasn't failure — it was the price of fast learning.
Wrong bets are data
Kill darlings without regret
Compound early — invest in shared foundation
Your lab is your competitive advantage
The ecosystem you see today didn't start as an ecosystem. It started with one wrong bet — and the discipline to kill it in 5 days.
“What's your laboratory?”
Links & contact