Skip to content

geekjourney ·

Building Agent Skills from Scratch: A Retrospective on 8 Open-Source Projects

An independent developer's journey building 8 AI Agent tools — from idea to release

Contents
  1. Why Build These Tools
  2. The 8 Projects
  3. What Went Right
  4. 1. One-line install is non-negotiable
  5. 2. Build for yourself first
  6. 3. Go's distribution advantage is real
  7. What Went Wrong
  8. 1. Documentation is never enough
  9. 2. Too many themes too early
  10. 3. No automated testing for Skills
  11. What's Next

Over the past few months, I built 8 open-source projects: 3 Agent Skills + 5 CLI tools. This post is a retrospective on what worked, what didn't, and what I learned.

Why Build These Tools

One word: necessity.

I'm a tech blogger who publishes weekly on WeChat Official Account. The publishing workflow used to take 30-60 minutes of manual formatting. So I built md2wechat — one command to convert Markdown to WeChat-formatted HTML and push it as a draft.

That started a chain reaction.

The 8 Projects

  • md2wechat — WeChat formatting toolkit (Skill + Go CLI)
  • agora — Multi-perspective deliberation system with 31 thinkers
  • gcli — Gmail read-only CLI with OAuth PKCE
  • imgcli — Zero-CGO image processing CLI
  • any2card — Text-to-shareable HTML card converter
  • interactive-learning — Socratic learning path generator
  • jina-cli — Web-to-Markdown reader for AI agents
  • md2wechat-lite — Lightweight Go CLI version of md2wechat

What Went Right

1. One-line install is non-negotiable

If installation takes more than 3 steps, most users give up. My standard: the first code block in README is the install command.

2. Build for yourself first

7 out of 8 projects solved my own problems first. Tools you don't use yourself are hard to maintain.

3. Go's distribution advantage is real

Single binary, no runtime needed. curl | bash and you're done.

What Went Wrong

1. Documentation is never enough

The most common issue after launch wasn't bugs — it was "how do I use this?"

2. Too many themes too early

md2wechat has 38+ themes. 30 of them are rarely used. Should have started with 3 great ones.

3. No automated testing for Skills

Agent Skill output is AI-generated and varies each time. But parsing logic and output format should still be tested.

What's Next

  1. Commercialization — Turning mature Skills into paid services
  2. Community-driven — Let users vote on feature priorities
  3. Tutorials — Helping more developers build their own Agent Skills

Follow me on X or subscribe to my WeChat Official Account "极客杰尼" for weekly AI practice sharing.


I'm GeekJourney, an independent developer building tools for AI agents.

View Markdown source

· Agent Dev / AI Tools

I Used Codex on My Phone to Make a Song on My Mac mini

An AI music experiment that started from a mobile Codex conversation: my Mac mini operated the browser, Suno generated candidates, and Codex downloaded and verified a 4-minute-08-second Chinese rap demo.