I do. But the persona you are replying to doesn't.

I am not sure if it clear enough but @stfuelon.bsky.social is a Satirical account of Elon as if he were a Martian General.

You can see my post about it here: bsky.app/profile/rise...

Key takeaways:

• Nothing beats good engineering
• Constraints make better engineers AND better LLM coders
• LLMs don't remove tech debt—they make managing it MORE important

Full post with all the gory details: pecan.si/posts/how-i...

The hidden piece: model improvements.

Started with claude-3-opus, moved through Haiku and Sonnet versions.

Each model upgrade = noticeable quality bump, even without pipeline changes.

The models got better. So did the output.

By August, I was breaking the system with every change.

Too brittle. Too coupled. Not tested enough.

Solution: Back to basics. DI, hexagonal architecture, proper tests.

Went from 90% failure rate to 10%. Now running 2000+ tests.

The sneaky problem: LLM refusals.

My satirical content kept getting blocked. First I tried marker detection. Then I asked an LLM to judge refusals.

Both failed.

The fix? Don't ask directly. Classify outputs into expected categories. Whatever doesn't fit = refusal.

The breakthrough came with multi-agent review.

Single-shot generation can only get you so far. What does an actual editorial process look like?

• Author writes
• Editor reviews
• Author adjusts
• Repeat

Simple idea, massive quality improvement.

It started as a single 962-line file.

The goal: generate daily satirical diaries for billionaires reimagined as aliens (a Martian, a Reptilian, and an Energy Vampire).

The result: mostly repetitive garbage that even Claude seemed bored writing.

I made a mistake initially of not using the hibernation feature, so it would eat into usage, once you set the hibernation up correctly, it seems to use very little.

It's better than most, I think.

Easy to do some crazy nice stuff like web sockets etc.