🐕guardrails, not prompts

I gave a 2023-era GPT-3.5 model a real browser and watched it lie to me — then fixed it with five guardrails and zero prompt edits. A 4-part series on building harnesses for weak models.

  1. 1.Watching a weak model lie to me
  2. 2.Structured tools — shrink the world
  3. 3.Validate, verify, and kill the lie
  4. 4.Bound the loop, own the preconditions
view the code on github →