🐕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.Watching a weak model lie to me
- 2.Structured tools — shrink the world
- 3.Validate, verify, and kill the lie
- 4.Bound the loop, own the preconditions