I built a personal AI agent workstation from scratch on Windows - here's everything I learned (including the traps nobody mentions)

Description and use

The author built a self‑hosted AI workstation on Windows, overcoming network, UTF‑8, and skill management challenges, and now automatically pulls data, generates weekly reports, and maintains tunnels.

Problem

Automating AI workflows fails when the network layer, data layer, or output layer are not fully implemented, leading to abandoned projects.

Outcome

The workstation now pulls data from 10 platforms daily, matches cross‑platform flags, publishes a 6‑page weekly PDF report every Sunday, and runs unattended while the author sleeps, eliminating API bills and rate limits.

Workflow

• 1. Define four layers: agent, data, network, output.
• 2. Build agent harness to decide tool usage.
• 3. Set up self‑hosted data service to own raw material.
• 4. Create a SOCKS5 tunnel through a cheap VPS with a keep‑alive watchdog to ensure network access.
• 5. Replace Invoke‑RestMethod with curl.exe to correctly handle UTF‑8 on Windows.
• 6. Develop skills (versioned instruction files) instead of prompts for reusable IP.
• 7. Implement pipelines: daily trending list pull, weekly report generation, tunnel and data service maintenance.
• 8. Log incidents line‑by‑line for troubleshooting.
• 9. Schedule pipelines to run automatically while sleeping.

Evidence

Everyone posts "just call the API" - nobody mentions your server can't reach half the internet.

Source

On Windows, Invoke‑RestMethod hands you garbled CJK text and you will waste hours thinking your scraper is broken.

Source

A prompt is text in a chat window. A skill is a versioned instruction file with tests and a changelog.

Source