The official Cursor repository cursor/plugins (11,377 stars) has published the /no-comments skill in the pstack collection: it launches the Comment Sicko subagent, which deliberately finds and marks code comments for deletion, preserving only cases from an explicit whitelist of exceptions — from licenses and public API doc comments to issue and RFC references. The subagent's report is then reviewed by the main agent: incorrect deletions are rejected, the root causes of the 'crutches' are eliminated, and constraints that cannot be expressed in code are proposed to be fixed with a test, type, or lint rule in CI. The skill appeared in pstack 0.13.0, is actively maintained, and can be adapted for environments other than Cursor.

image

What happened

The /no-comments skill was included in the pstack collection version 0.13.0 — the corresponding commit in the official cursor/plugins repository is dated August 2, 2026. Since then, the project has been actively developing: the last commit to the SKILL.md file was made on October 14, 2026, and on October 12, 2026, the subagent identifier was renamed so that Grok could launch it — the intensity of edits noticeably reduces the likelihood that this is an abandoned experiment or a random pull request. The architecture is two-stage. First, the Comment Sicko subagent, in a given 'hostile' role, goes through the code and marks comments for deletion; only licenses, public API doc comments, issue and RFC references, the prettier-ignore directive, and behavior imposed by external dependencies survive the cleanup. Then the main agent reviews the subagent's report: it rejects incorrect deletions and fixes the root causes of the 'crutches', and proposes to transfer rules that cannot be expressed in code to a test, type, or lint rule in CI.

Context

The problem is well known to teams with multi-agent development: crutch comments like 'temporary fix, we'll remove it later' are perceived by each subsequent agent as an accepted practice and copied further. One such marker turns into technical debt and constant noise in the context that agent tools reread at each iteration. The approach itself is not new: the combination of 'adversarial subagent plus orchestrator verification' is a long-known pattern, and /no-comments does not claim to be a scientific result or a new method, but is a neat engineering instantiation of this pattern for a specific task. By nature, this is a prompt configuration, not a product with an API and SLA: the cleanup is formatted as a reproducible process with a whitelist of exceptions, review by the main agent, and transfer of inexpressible rules to machine-checkable artifacts.

Why this matters for the industry

For the industry, the value lies in a formalized template for AI-generated code hygiene. In pipelines where several agents sequentially edit code, the accumulation of garbage is usually held back only by human discipline; here the process is packaged into a distributable artifact: a hostile subagent-auditor with an explicit whitelist, review by the main agent, and translation of constraints into tests, types, and lint rules. Transferring knowledge from comments to machine-checkable forms hits exactly the mechanism of technical debt accumulation: the next agent will no longer be able to copy a 'temporary fix', because the invariant is now checked by CI. The approach is open and free, so AI code hygiene becomes much cheaper — although the skill itself remains a utility without a buyer or pricing, and it is impossible to build a company on it. The assessment for the future is a forecast, not a fact: the viability of the pattern will probably lead to its replication in other skill collections and for other classes of garbage — TODO debt, dead code, naming conventions — and in a horizon of about two years, 'hygienic' auditors may become a standard layer of code generation pipelines, an analogue of the mandatory linter of the pre-AI era.

Why this matters for users

You can try it right away. The SKILL.md and comment-sicko.md files are in the open cursor/plugins repository: it is enough to download them and ask your agent to adapt them to your environment — you will get a ready-made comment cleanup process with a clear list of exceptions (licenses, public API documentation, issue and RFC references, service cases like prettier-ignore and behavior of third-party dependencies). Cursor users have it even easier: the /no-comments skill is published in the Cursor Marketplace and is connected as a regular skill. The implementation itself is a prompt configuration, not a server deployment, so rollback and isolation are simple: if you don't like the result — the skill is removed, and nothing in the repository will be broken. A mandatory condition for production launch is adaptation to your environment, as the solution was originally created for Cursor.

What is still unknown / limitations

There are no efficiency metrics in open sources: the proportion of false deletions (including the preservation of comments from the whitelist in disputed cases) and the frequency with which the main agent rejects the subagent's report have not been published. There is also no 'before/after' comparison by comment density, code quality, or frequency of relapses — the viability of the structure is currently based on the reasonableness of the architecture, not on measurements. The skill was created for Cursor, so transferring it to other agents and environments requires adaptation without compatibility guarantees. Finally, the word 'reproducible' here should be understood in an engineering, not scientific, sense: repeatability on other codebases still needs to be checked.

Sources

Author

Look at AI, editorial team