<- all tokdocs
2026.09.22@maven_hq

Cursor's Lauren Tan shipped her code-comment ban as a public agent called Comment Sicko, and its keep-list has exactly five entries

Coding
Watch on TikTok

View on TikTok ->

The engineer in this clip is Lauren Tan, and the rule she describes is not a preference she keeps in her head, it is a published agent definition sitting in the cursor/plugins repository under an MIT license. Maven posted the 23-second clip on September 22, 2026. The frame is split in two. The top half is a dark navy mock editor where code lines are drawn as green and grey bars, and comment lines type themselves in character by character: first // moved from utils.ts, then // during the auth refactor, then // Lauren said you should never do this. As each one lands it turns from white to orange. The bottom half is her webcam feed with a lower third reading "LAUREN TAN / Engineer @Cursor". In the last few seconds a dashed line connects the offending comment to a box labeled "PR REVIEW", then a second box labeled "CI" flips its three neutral checkmarks to red X marks. The closing frame is a green "LINK IN BIO" button above the words "for the full lecture".

The on-screen code carries the argument, not the audio

The spoken line "it will often say, oh, Lauren said you should never do this" is confusing without the visuals, because Lauren is the speaker. The animation resolves it. She left a review comment on a pull request, the agent read that review comment, and the agent wrote her name and her instruction into the source file as a permanent // comment. The dashed line to the "PR REVIEW" box makes the provenance explicit. Her complaint is about a category error. A review comment is a message to a person that expires when the change lands. A code comment is a claim about the code that persists until someone deletes it. Agents that read review threads as context will happily promote the first into the second.

The second example on screen, // moved from utils.ts during the auth refactor, is the cleaner illustration. That fact is already in git. Duplicating it in the file means it will be wrong the next time the file moves, and nothing will flag the drift.

The rule is public, and it is stricter than the clip suggests

Lauren Tan is listed on the Maven lecture page as Member of Technical Staff at Cursor. The framework she teaches there is pstack, which lives in the official cursor/plugins repository. It includes a skill called /no-comments whose entire job is to spawn a read-only subagent named Comment Sicko before review. The agent file opens in character: "I hate comments. Feed me the parent scoped files or diff. If none exists, feed me the current diff against main. Narration, banners, commented-out corpses, workaround sermons. I want them all."

I checked the actual text rather than trusting a summary, and the delete list does not name "historical comments" the way the clip's framing implies. The ban is structured as a whitelist instead. Only five things survive: legal or license headers, non-obvious behavior forced by an external dependency or platform that the team cannot reshape, // prettier-ignore, doc comments defining a public API contract, and issue or RFC links explaining a constraint the code cannot express. Everything outside that list dies by default, and the file is explicit about ties: "When I am not sure a keep clause applies, the comment dies."

So // moved from utils.ts during the auth refactor is not killed by a rule that targets history. It is killed because it fails to match any of the five keeps. That is a meaningfully different and more aggressive design than "ban AI comments," and the clip does not convey it.

The 99% figure has no source

She says agents write irrelevant historical comments "99% of the time." I found nothing behind that number. There is no measurement in the pstack repository, no figure on the Maven lecture page, and no Cursor blog post quantifying it. Read it as emphasis, not data. The same caution applies to the volume claims that circulate around her work. Reporting on her talks cites both 2,000 and 2,500 pull requests shipped in a month, the two numbers come from different presentations of the same material, and neither has been audited outside her team. None of that appears in this clip, but it sets the tone for how her claims travel.

What the clip omits

The closing line, "everything you can imagine that agents are bad at, we ban," is where the promo does the most compression. The "CI" box turning red is the enforcement mechanism, and the Maven lecture has a chapter for it at 41:00 titled "Implementing Strict CI Constraints and the Dune Architecture." The clip shows the outcome without the mechanism, which leaves the impression that banning is a policy conversation rather than a lint rule someone has to write and maintain.

The clip also blurs whose rule this is. She says "we ban," and the video title says "this Cursor engineer bans." pstack's README is written in the first person by poteto, her X handle, and it calls the stack "my answer." It is published in Cursor's plugins repository, which is a strong signal of institutional blessing, and I could not find any Cursor document establishing a company-wide comment ban. Treat it as her stack, adopted broadly enough that she says "we," until Cursor publishes otherwise.

One genuine gap in the argument: the keep-list assumes you can always reshape confusing code instead of explaining it. The agent file says as much, telling the reviewer to mark surprising in-house code MUST KILL for rename, extraction, or rearchitecture. That is the right instinct and it is also the expensive path. A team that bans the comment but never funds the refactor ends up with neither the explanation nor the fix.

Key Takeaways

  • The speaker is Lauren Tan, Member of Technical Staff at Cursor, and the clip promotes her free Maven lecture "How Cursor Turned AI Agents Into Better Engineers," recorded August 12, 2026.
  • The rule is real and inspectable. pstack's /no-comments skill and its Comment Sicko subagent are in cursor/plugins under MIT.
  • The enforcement is a whitelist of five permitted comment types, with an explicit tiebreak that deletes anything ambiguous. That is stricter than "no AI comments."
  • The specific failure she names is agents promoting pull request review feedback into permanent source comments, which mislabels a temporary message as a durable claim about the code.
  • The "99% of the time" figure is unsourced. I found no measurement behind it in any primary material.
  • Transcript quality: Whisper captured all 23 seconds of speech accurately and did not cut off. The only discrepancy against the burned-in captions is tense, where Whisper has "I actually ban code comments as well" and the caption reads "I actually banned." The transcript misses everything that makes the clip legible, including the speaker's name and employer from the lower third, the three animated comment lines, the "PR REVIEW" and "CI" labels, and the "LINK IN BIO for the full lecture" end card.
  • Unverified: the 99% claim; whether the comment ban is Cursor company policy or Lauren Tan's own workflow adopted by her team; the identity of the interviewer or full context the clip was cut from; the pull request volume figures attributed to her elsewhere, which appear as both 2,000 and 2,500 and have not been independently audited.

Resources

Published September 22, 2026. Writeup generated from a favorited TikTok.