Viral! Grok's Top Engineer Teaches You How to Submit 3,000 PRs a Month

Cheng Qian, reporting from WaFei Temple, QuantumBit | WeChat Official Account QbitAI
"All code is slop (except mine)."
SpaceXSI principal engineer Potato Sister (Lauren Tan) took this wildly arrogant talk title onto the stage at the AI Engineer CODE conference, and the talk went viral across the internet.
Potato previously worked at Cursor and currently works on Grok Bot development, and is known as "the person who uses Grok Bot best on the whole internet."
According to her, last month alone she submitted 3,000 PRs by herself.
What does that mean? Last month, the entire Grok Bot team submitted and shipped 11,000 PRs in total (with a revert rate of only 0.5%).

△YouTube "All Code Is Slop (Except Mine)"
The talk wasn't long, but Potato shared quite a few AI coding tips with everyone.
"When AI makes a mistake, don't rush to add rules to AGENTS.md."
"Just ban 99% of code comments outright."
"If your AI coding output is slop, then automating this process will only produce more garbage."
Potato's AI coding mini-class is now in session (tapping the blackboard).

Keep your codebase as clean and tidy as possible
Today, the most important job for software engineers is to find a way to maintain or even improve code quality while productivity rises dramatically.
Potato Sister believes that rather than building a "software factory," it's better to simply understand this as a "trust problem."
That is, the problem of "transforming a low-trust Agent working environment into a high-trust environment."
To this end, she proposed the "inverted trust pyramid model."

At the top of the inverted pyramid, taking up the largest share, is the code itself, including API design, software architecture, and code refactoring; the next layer down consists of deterministic checking tools such as Lint rules, type systems, compilers, and automated tests.
Only below that come runtime verification, formal verification, and natural-language guidance such as Rules, Skills, and AGENTS.md.
Potato Sister believes that the most effort should be invested in the top two layers.
She compares the codebase to a garden.
If the codebase is full of temporary patches, dead code, anti-patterns, and various kinds of technical debt, the Agent will treat them as existing conventions and continue to imitate and "carry them forward."
So software engineers are becoming more and more like gardeners, constantly clearing weeds and providing a good environment for new code to grow.

First, ban 99% of code comments outright.
This is a lesson Potato Sister summed up after suffering setbacks while working at Cursor.
At the time, the Cursor client used an Electron architecture, in which VS Code's imperative programming style conflicted with React's declarative programming style.
The Agent loved to use timers such as setTimeout, requestAnimationFrame, and requestIdleCallback to patch these architectural problems.
Large numbers of timers were repeatedly created and destroyed, triggering frequent garbage collection and causing the application to stutter.
Not only that, the Agent would also write piles of comments next to the code explaining why it had to write things that way.
When later Agents saw these code snippets and comments while working, they would again find them very reasonable and continue copying them.
In this way temporary patches piled up, and application performance suffered.
So Potato Sister currently bans about 99% of code comments outright in Grok Bot.
Instead, Lint rules are used to directly check whether the code has problems, and clear error messages guide the Agent to fix the root cause.
Second, don't let the Agent write features it is prone to getting wrong.
There are quite a few such designs in Dune, the client framework developed by the Grok Bot team.
Potato Sister describes Dune as "Next.js for Electron apps." It can turn experience accumulated during team development into enforceable architectural constraints and Lint rules.
For example, the team found that some features that require careful judgment to use correctly are easily misused by the Agent.
React's useEffect is a typical example. This Hook can be used to handle side effects, but to use it correctly one needs to carefully consider dependencies, cleanup logic, and other issues—so much so that "to use useEffect well you first need a PhD."
So the Grok Bot team's solution was simple: ban useEffect outright in product code and only allow it within specific APIs inside the Dune framework. Problem solved once and for all.
Finally, every time the Agent makes a mistake, turn the lesson into an automated check as much as possible.
Potato Sister emphasized (highlight this), many people, when the Agent writes wrong code, have the first reaction of opening AGENTS.md and adding a line like "don't do xxx in the future." But natural-language rules require the Agent to understand and comply on its own, so this is of little use.
Her approach is to convert these requirements into deterministic checks as much as possible.

△The team continuously accumulates engineering constraints targeting Agent errors
For example, add Lint rules for incorrect code references, use the type system to constrain illegal calls; if a certain API is always misused, redesign it to eliminate that possibility architecturally. After that, clean up the same kind of problems that already exist in the codebase.
Through this cycle, the repetitive mistakes the Agent can make become fewer and fewer.
Let the Agent verify things itself
Once the top two layers of the inverted pyramid are clean and tidy enough, it's time to look at how to verify.
Potato Sister introduced that the Grok Bot team has already shifted a large part of code review work to the verification stage.
What they mainly look at during review now are screenshots, screen recordings, state diagrams, and formal verification results.
Specifically, they start with runtime verification.
The Grok Bot team gives the Agent the ability to run the application, as well as debugging tools commonly used by human developers, such as Chrome DevTools Protocol (CDP).
The team also designed a feature map for the Agent, similar to a website's Sitemap, but from the user's perspective.
This way the Agent no longer has to guess what you mean by "that thing below the left sidebar."
In addition, the team encapsulated some operational capabilities into custom CLIs and drivers and put them into verification Skills. This way the Agent can reuse the same set of tools each time it performs a verification task, saving Tokens, and these tools themselves can also be tested independently.

Next is formal verification.
Potato introduced that the Grok Bot team recently began introducing two mathematical verification tools, TLA+ and Lean.
Among them, TLA+ is mainly used for model checking, analyzing whether the system satisfies expected constraints in different states. The other tool, Lean, can verify the correctness of specific algorithms through formal proofs.

In addition, they introduced Bombadil to carry out property-based testing on the user interface.
This way, rather than only checking the results of pre-specified operations, it will check more possible inputs and states around the properties the program should satisfy.
Potato Sister said that some developers around her spend 70% of their Tokens on verification, and she herself is currently at about the same level.

△Various testing and verification methods for software from development to post-launch
In her view, static analysis, automated testing, and formal verification are not new methods; it's just that in the past the cost of understanding, writing, and maintaining these tools was too high.
And now, "the golden age of software testing and verification is arriving."
One More Thing
So, after this series of smooth little tricks, how efficient is the Grok Bot team's development now?
In October this year, the entire team submitted 11,000 PRs, averaging over 350 per day. And as development speed continued to increase, the team's PR revert rate actually went down, currently stable at around 0.5%.
The client's release frequency has also increased accordingly, and now a new stable version is released almost every day.

△The Grok Bot team's PR delivery volume continues to grow
The team has now achieved a rather crazy automated development process. Whenever Grok Bot releases a new version, the system automatically starts a new Cursor project and calls a cloud Agent to fuzz-test the entire application.
Article title: Viral Everywhere! Grok's Leading Lady Teaches You How to Submit 3,000 PRs a Month
Body: These Agents run and operate software like real users to find bugs. Once they find a problem, they automatically create fix PRs and even complete the merge directly.
Achieving this took Tudou a full six months.
According to her, most of the product code the team merges now no longer requires humans to perform in-depth code review.
Reference link: [1]https://x.com/poteto/status/212
- AI cranks out live MV for a band on the spot—is this a hackathon or a music festival 2026-11-21
- Chinese computing embarks on a new path to the summit 2026-11-21
- 80,000 group chat messages turned into an exam: Tencent Yuanbao × KCL release the first agent memory benchmark built on group chats 2026-11-21
- After overtaking Gemini, Karpo's new work is open-sourced again—browser agents win on both speed and effectiveness 2026-11-21