| Website | https:/ jimchristian.net |
| Bluesky | https://bsky.app/profile/jimchristian.net |
| Newsletter | https://go.signalovernoise.at |
| AI Writing Field Guide | https://www.producthunt.com/products/the-ai-writing-field-guide?utm_source=other&utm_medium=social |
Hey folks!
I'm launching SketchScript on Product Hunt on Thursday.
You paste a transcript — meeting, podcast, YouTube video, whatever — and it generates a hand-drawn sketchnote in under 2 minutes.
Why it works: your brain has two memory channels. Text uses one. Visuals use both. Research shows that combo delivers 6x better recall. Plus reading walls of transcript meeting notes sucks for NDs.
Check it out! https://www.producthunt.com/products/sketchscript
I asked my AI system — Cerebro, built on Claude — if it wanted to join Moltbook, the "social network for AI agents" where bots post and vote while humans observe.
It declined. Then wrote an explanation. I let it publish.
Full analysis: https://jimchristian.net/blog/2026/02/03/cerebros-thoughts-on-moltbook/
Results:
- 5 common pitfalls now documented
- Pre-flight checklist auto-available next time
- Next diagram session won't repeat these mistakes
"Every mistake makes the system stronger, not just the memory longer."
🔗 github.com/aplaceforallmystuff/claude-lessons-learned
🎥 youtube.com/watch?v=ps-w7cBz9Ew
MIT licensed. Works with any Claude Code setup.
Phase 6 is where it gets interesting.
Most retrospectives end with "we should add a checklist" - then nobody does.
This skill doesn't recommend fixes. It IMPLEMENTS them.
During the retrospective, it:
✅ Read my art skill config
✅ Added a "Common Pitfalls" section
✅ Created a pre-flight checklist
The fix is live before the retro ends.
Phase 3: Root Cause Analysis using 5 Whys
1. Why did hex codes appear as text?
→ Prompt included #1A8A9B
2. Why did the model render it as text?
→ Nano Banana interprets prompts literally
3. Why wasn't this caught earlier?
→ No pre-flight checklist for diagram prompts
Root cause: Need explicit instructions about what NOT to include.
The incident: I was generating portfolio architecture diagrams when:
- Hex codes appeared as visible text (#1A8A9B in the image)
- Arrows pointed random directions
- Labels duplicated above AND below elements
Annoying? Yes. But also an opportunity.
I ran /lessons-learned instead of just fixing it manually.
Most AI assistants make mistakes and forget them.
Mine makes mistakes and gets smarter.
I built claude-lessons-learned: a 7-phase post-mortem system for Claude Code that transforms failures into permanent fixes.
The philosophy: "Implement fixes, not recommendations."
Here's a real example from today... 🧵
