It's still rough — building it in the open. If you're fighting the same thing and want to try it, reply or DM and I'll share it. And if you think the approach is wrong, tell me — I'd rather hear it now.
Hot take: your AI agent isn't unreliable.
You just gave it authority.
You wrote the workflow in CLAUDE.md.
The model read it.
Agreed with it.
And then ignored it.
The fix for non-deterministic agents wasn't a smarter model. It was taking control away from the model. (opened a bug on Anthropic about this first: https://t.co/p4OIJYwYL7)
Prompts are NOT contracts. They are suggestions.
I stress-tested Claude Code to see how to actually enforce agent behavior. 6 empirical tests on hooks, security blocks, and bypasses not found in the docs.
Check: 👇 https://t.co/ldVhJmDUpL
#ClaudeCode#AI#Anthropic