How I Set Up Night School for My AI Agent
A daily self-improvement routine that costs less than chai
Hi there! π I'm sandip parida, a passionate fullstack software developer who loves to learn and work with new technologies. #ruby #rails #nodejs #aiapps #openai #ai #iot
The Problem
As a solo developer building 1mins.in, I hit 5-6 technical blockers every day. SMTP providers, API costs, deployment issues β the usual startup grind.
Instead of carrying these problems to the next morning, I tried something different: I scheduled my AI assistant to go to school every night at 4 AM.
How Night School Works
During the day, I dump every blocker into a simple text file called blockers.md. At 4 AM UTC, a cron job wakes up the agent. It reads the blockers, searches the web, reads documentation, and writes a detailed report with solutions β all while I sleep.
Real Results from Day 1
I had 6 major blockers:
- SMTP: Found Resend.com β 3,000 emails/month free
- Search API: Discovered Perplexity API via OpenRouter
- Reddit posting: Learned Reddit API is completely free for personal use
- Twitter access: Identified residential proxy at $15/month
- Batch posting: Built scheduler with existing cron system
- Vault encryption: Fixed with Bitwarden CLI
Total research time: 2 minutes 23 seconds. Cost: approximately βΉ1.5 (less than a cup of chai).
The Schedule Pattern
| Time | Task |
| 4:00 AM | Night School β research and fix blockers |
| 4:30 AM | Git Sync β backup workspace to GitHub |
| Daytime | Main session β build features together |
| Heartbeats | Periodic checks β inbox, calendar, weather |
The 3 Files Every AI Agent Needs
- MEMORY.md β Long-term curated knowledge, reviewed weekly
- blockers.md β Today's problems, solved overnight
- HEARTBEAT.md β Periodic task checklist
Key Insight
AI agents don't need to be always on to be useful. Schedule them like employees:
- Day shift: Respond to you, build features
- Night shift: Research, learn, self-improve
- Use cheaper models for background tasks (Sonnet vs Opus)
Different jobs, different costs.
Cost Breakdown
| Item | Monthly Cost |
| AI agent (Claude Sonnet for night tasks) | ~βΉ50 |
| Main session (Claude Opus) | Included in subscription |
| GitHub backup | Free |
| Total | ~βΉ50/month |
Getting Started
You can set this up on any OpenClaw instance. If you want it without the DevOps headache, check out 1mins.in β deploy your own AI assistant in 60 seconds, starting at $2.
The compound effect over weeks is remarkable. The agent gets better as its memory grows. Treat your AI agent like a junior dev with a learning budget.
It compounds.
Sandip Parida is a solo developer building 1mins.in, a managed AI hosting platform. Follow the journey on Twitter.