TL;DR
- Do it manually first — I did this process by hand several times before building the automation
- AI is probabilistic, not deterministic — expect to debug; the agent built the skill but forgot the slash command
- ChatGPT's Prompt Optimizer improves prompt efficacy, even when using Claude
- The LLM asking questions extracts your thinking — it's a brainstorming tool
- Four phases of episode prep: Ideation, Content Development, Proof of Concept, Blind Spots
What I've Learned
This is a behind-the-scenes episode — not a Q&A. It's for builders who want to see how I actually work.
The pattern is simple: do it manually first, pay attention to what you repeat, then systematize. I prepared several Dear Ben episodes by hand before I started building automation around it. You can't automate what you don't understand.
I use ChatGPT's Prompt Optimizer even when building for Claude. It takes messy thoughts and turns them into structured prompts. Anecdotally, it genuinely improves efficacy — the LLM follows instructions more reliably.
Every's Skill Creator Agent is useful for building Claude Code skills. But here's the thing: AI is probabilistic, not deterministic. The agent built my skill perfectly — then forgot to create the slash command to invoke it. Expect to debug.
The most valuable part of this workflow is the LLM asking me questions. It forces me to articulate what I actually believe about a problem. It's a brainstorming tool that extracts thinking I wouldn't have written down otherwise.
Tools & Resources
- Claude Code — CLI for building with Claude
- ChatGPT Prompt Optimizer — improves prompt structure (works for any LLM)
- Every's Skill Creator — agent for building Claude Code skills
- Four-phase framework: Ideation → Content Development → Proof of Concept → Blind Spots