"Fede, how do you prompt things into existence? What's your setup?"
I am approaching Kandinsky level coding experience. AI builds full SaaS apps with >20k users and manages my WhatsApp.
It’s time to share my setup. All of it.
Especially because I don't want you to suffer the way I did.
I've mass-deleted my own Mac with a bad rm -rf. Leaked API keys to a public repo. Spent $100k on AI tokens across 500 sessions. Seven days a week, for months.
(Kids won't even understand how close everyone was to a state of epilepsy when Claude Code used to have a flickering bug)
I've seen a lot. Every blocker or mistake became a rule.
Boris Cherny (creator of Claude Code) said hooks alone make you 2-3x more productive. This setup goes way beyond hooks. Every rule became a hook, a skill, or a line in a config file. My complete stack, open sourced:
- 12 safety hooks that block destructive commands before they run (rm -rf, secret leaks, wrong package manager, force pushes)
- 40 slash commands: /ship, /debug, /qa, /deploy, /morning, /recall, /launch, /pdf, /docx, /pptx, /xlsx
- Independent auditors: Gemini reviews Claude's output before it can stop. Second model catches what the first one missed, so you don't waste cycles on broken code.
- Memory system so Claude learns across sessions and stops repeating mistakes
- CLAUDE.md templates with self-audit rules, error recovery patterns from 10,000 errors analyzed
- Full cloud hosting setup: headless Chrome, tmux, safe process execution with memory caps
- WhatsApp Gmail integration, cost tracking, cron patterns
This setup powers everything I ship: FlyFast (flight tracker), OpenDraft (academic search), and 10 other open source tools (120 GitHub stars, 20,000 installs).
Hooks give you 2-3x. The full stack is probably 10x.
One command: git clone ./install.sh (or just ask your agents to do it ;))
Everything templated, no personal data. MIT license.