Skip to content
Back to skills

Deslop

ASecurity

Remove AI-generated code noise from the current diff while preserving behavior. Use for deslop, AI slop cleanup, unnecessary comments, defensive checks, any casts, or needless nesting after an agent-generated change. Do not use as permission for a broad rewrite or when a real bug has not been isolated.

  • 154 stars
  • 0 votes
  • 0 copies
  • 0 views
  • Added October 8, 2026
ai-agentsrustgoc++refactoringgitsecurity

Works with

  • cursor

Security analysis

A100/100

Scanned October 8, 2026

npx -y skills add AnastasiyaW/codex-claude-code-config --skill deslop --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Deslop?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Deslop
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/anastasiyaw-deslop/badge)](https://www.skillsdirectory.com/skills/anastasiyaw-deslop)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: deslop
description: Remove AI-generated code noise from the current diff while preserving behavior. Use for deslop, AI slop cleanup, unnecessary comments, defensive checks, any casts, or needless nesting after an agent-generated change. Do not use as permission for a broad rewrite or when a real bug has not been isolated.
---

# Deslop

Review only the branch diff against its intended base and remove noise that
does not belong in the local code style. Keep the behavior and public contract
unchanged unless a clear, separately verified bug is fixed.

## Review Targets

- comments that narrate obvious code or contradict local conventions;
- defensive checks or catch blocks abnormal for a trusted path;
- `any`, unsafe casts, or optionality used only to silence a type checker;
- deep nesting that can be made clearer with early returns or a named helper;
- one-off wrappers, flags, and branches inconsistent with the surrounding
  module;
- C++ ownership or error-handling scaffolding that is redundant with the
  established RAII/contract boundary.

## Workflow

1. Inspect the base, diff, local style, tests, and ownership boundaries.
2. Classify each candidate as noise, a clear bug, or an intentional contract.
3. Remove only confirmed noise in a focused edit.
4. Run the narrow relevant checks and inspect the final diff.
5. If the structure needs a real redesign, stop deslop and use
   `refactoring-safely`, `architecture-quality`, or
   `thermo-nuclear-code-quality-review`.

Do not delete comments that explain a non-obvious invariant, security boundary,
ABI constraint, workaround with an owner, or externally required behavior.

## Gotchas

- Shorter code is not automatically clearer; preserve names and boundaries that
  carry domain meaning.
- A broad formatter run can hide behavior changes and is not deslop proof.
- Removing a defensive check without proving the trusted-path invariant can turn
  cleanup into a regression.
- In C++, exception and ownership code may look repetitive while protecting an
  ABI or lifetime boundary; inspect callers before removing it.

## Troubleshooting

| Symptom | Likely cause | Action |
|---|---|---|
| Cleanup changes a test result | Candidate was behavior, not noise | Revert that candidate and isolate the real contract |
| Diff is too broad | Tool ran over the whole tree | Restrict review to the branch diff and restore unrelated files |
| Comment seems redundant but explains a constraint | Context is outside the file | Read the owning docs/tests before changing it |
| Code remains structurally tangled | Deslop is the wrong scope | Escalate to a planned refactor with characterization tests |

## Source

Adapted from Cursor Team Kit's MIT-licensed `deslop` workflow:
https://github.com/cursor/plugins/tree/main/cursor-team-kit/skills/deslop

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.