How this helps you
It shows where effort pays off most when you check agent work: automatic checks catch more than extra written rules.
Look through my code review comments and past commits to find the most common mistakes agents make in this codebase. Propose a plan to restructure the code I own so those mistakes are harder to make, backed by real examples. Confirm the plan with me before you change anything.
What they did
@poteto (Lauren, on the Grok Bot team) shared an order of effort for trusting agent work: 1. better code structure, 2. lint, tests and CI, 3. prose rules and skills (sparingly), 4. human review, which is the hardest to scale. Her reasoning: agents copy what already exists and take the shortest path, so make the easy thing the right thing. Code structure and static checks (lint, tests, CI, formal verification) are deterministic and token-efficient; prose rules need agent judgment, so use them minimally; human review cannot keep up with every PR, so spend it on designing the environment. The post ends with a project prompt that has the agent study your past review feedback and commits to find common pitfalls and propose a rearchitecture plan, then confirm the plan with you before starting.
Who it's for: People using Grok Bot or other agents to write code.
How to try it
- Invest in code structure first: clearer boundaries, stronger types.
- Add lint rules, tests and CI next. They catch problems without needing judgment.
- Use prose rules and skills sparingly.
- Keep human review for what's left, and spend it on designing the environment.
What to expect
A written plan listing common agent mistakes in your code and what to change so they stop happening.
Common mistakes
- Answering every agent mistake with another written rule instead of a test or lint check.
Related tips
- What actually stops a coding Bot (and what is only advice): Of 15 operations tested, only a merge to main hit a hard stop. Skill-file rules are only advice.
- Watch coding work live and get recordings back: With Cursor connected, your Bot links you to the cloud agent live and sends screenshots and screen recordings.
- The fan-out method: three models, one brief: Launch Cursor cloud agents on three models at once and keep the best parts.
- Promote a Bot to director and let it build and run its own team: When one Bot has collected a lot of context, promote it: ask it to create three Bots that report to it, keep them away from you, and send you one Monday summary.
- A routine that triages your backlog with RICE and blast radius: Your Bot collects requests into a backlog, then a routine scores each item for reach, impact, confidence, effort and blast radius so the best small jobs rise to the top.
- Make your Bot the manager, not the coder: one rule and locked doors: Let your Bot plan and check the work while coding agents write the code. Start with one rule, and remember every Bot shares one computer and its logins.
Worth knowingThis was widely quoted as the 'verification skill' idea. Her original project prompt uses her own pstack skills; ours is a plain-language version of the same steps.Next in Skills →A /clean-start skill for long chats
Learn more: How to Use Grok Bot: A Beginner's Guide (2026) · All Grok Bot skills tips