All Code Is Slop (Except Mine) by Lauren Tan
Lauren Tan has a pretty interesting talk about how she and her team code with AI (for their Grok Bot).
Some of the highlights from the talk:
The main takeaway is that we should focus on the environment around the code that helps us to improve our trust.
To improve our agents, we need to focus on improving these things:
1. Codebase improvements: We should focus on better architecture, good api designs, stronger types and invariants, clear boundaries.
Your codebase is the agent’s largest input; anything in it gets copied as “the established pattern.”
2. Static analysis: Have strict lint rules, tests, CI, and other usual things that we do to catch issues like in the good old days :)
3. Verification: Runtime and formal
4. Prose: skills, AGENTS.md, Comments - anything that can’t be enforced and requires agent judgement.
Rule: every time you correct an agent, don’t add a line to AGENTS.md — make it a deterministic check, or better, design the invalid choice away entirely.
And in the Grok Bot codebase:
- They banned the APIs like React’s useEffect which has caused lot of issue/confusion in their Dune framework.
- Context via directory structure - They’ve explictly seperated the
electron-mainandelectron-rendererin their electron based app. And have explictly banned the cross-imports. - They’ve banned comments (which agents adds a lot for quick fixes)
And her full talk