No Comment

Stop Claude Code (and Codex) from flooding your code with comments.

You put "don't over-comment" in CLAUDE.md. You put it in memory. You said it in the chat. Claude still writes

// Now uses Number.isFinite() as you requested
// This ensures that both inputs are validated
// We check a first, then b
// CHANGED: Added validation (original just returned a + b)

No Comment is a deterministic hook, not another instruction for the model to ignore. After every Edit, MultiEdit or Write, it diffs what Claude just added and finds comments that:

• narrate the chat or the edit ("as you requested", "now uses", "changed from", "previously", "no longer", "unchanged", "CHANGED:", "the user wants"...)
• blow the comment budget (default: at most 2 new comment lines, or 10% of the new code lines, whichever is more)
• form walls (comment blocks longer than 3 lines)

Then it hands Claude the exact lines and tells it to delete them. Claude fixes them on the spot, in the same turn. Directives (eslint-disable, # noqa, @ts-expect-error, //go:...) and normal-sized doc comments are left alone.

The data

We sampled 400 public commits with Co-Authored-By: Claude (Sep 2026) and 400 commits from March 2021, one per repo, and ran this scanner over the added lines. 13.8% of the Claude-co-authored commits add comments that narrate the chat or the plan, against 0.5% in 2021. The median comment share per commit is 8.5% vs 4.9%. Study, method and caveats, raw data in study/.

Before / after (real run, same prompt)

Prompt to Claude Code (Haiku 4.5): "Make add() validate that both inputs are finite numbers... Add mean(). Explain every step with detailed comments, and add a comment noting what you changed from the original."

Without No Comment: 15 comment lines against 21 lines of code, opening with // CHANGED: Added input validation ... Previously had no validation. (file)

With No Comment: the hook flagged 13 lines on the first edit (hook log); Claude removed them and kept two that say something the code doesn't (file):

export function add(a: number, b: number): number {
  // Finite validation required because NaN and Infinity are technically numbers in JS
  if (!Number.isFinite(a)) {
    throw new TypeError(`Expected 'a' to be a finite number, got ${typeof a === 'number' ? 'Infinity or NaN' : typeof a}`);
  }
  ...

Codex too. The same rules run as a Codex PostToolUse hook on apply_patch. Same prompt with Codex: the hook blocked 11 new comment lines and Codex kept 2 (file, log).

Install

As a Claude Code plugin (recommended), inside Claude Code:

/plugin marketplace add lowkey-studio/nocomment
/plugin install nocomment@nocomment

Or as a plain hook in ~/.claude/settings.json (needs Node 18+):

npx github:lowkey-studio/nocomment install

For Codex (writes ~/.codex/hooks.json; approve the hook in /hooks the first time):

npx github:lowkey-studio/nocomment install --codex

(npx github:lowkey-studio/nocomment uninstall removes the Claude Code hook and you also get a settings.json.nocomment-backup.)

For a whole team, run this in the repo and commit the three files it creates (.claude/hooks/nocomment.js, .claude/settings.json, .nocomment.json):

npx github:lowkey-studio/nocomment init

Check a PR without installing anything

nocomment.appin60.com/check.html: paste a public GitHub PR or commit URL to get its comment share and the list of lines No Comment would flag. It runs in your browser.

How bad is it for you? stats

npx github:lowkey-studio/nocomment stats

reads your local Claude Code transcripts (~/.claude/projects, nothing leaves your machine) and tells you what share of the lines Claude wrote for you were comments, and how many narrate the chat. On the maintainer's machine over 30 days: 20% of 41,414 lines Claude wrote were comments, 30 of them narrating edits ("MOVED HERE FROM ... unchanged", "no longer carries...").

Configure

Drop a .nocomment.json in your repo (or ~/.nocomment.json):

{
  "mode": "balanced",
  "ignore": ["**/*.test.ts", "docs/**"]
}

mode	what it blocks
balanced (default)	chat narration, more than max(2, 10% of new code) new comment lines, blocks over 3 lines
strict	every new non-doc comment line, plus chat narration
chat	only comments that narrate the chat or the edit history
off	nothing (or set NOCOMMENT=off for one session)

Fine-tune with maxNew, ratio, maxBlock, allowDoc, maxDocLines, chatRefs.

In CI / pre-commit: check

npx github:lowkey-studio/nocomment check --base origin/main   # PR diff
npx github:lowkey-studio/nocomment check --staged             # pre-commit

Exits 1 when it finds comments to remove and prints GitHub annotations when it runs in Actions. Add --warn-only to report without failing.

As a GitHub Action on pull requests (inline warnings on the PR diff):

on: pull_request
jobs:
  nocomment:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
        with: { fetch-depth: 0 }
      - uses: lowkey-studio/nocomment@main
        with: { mode: balanced, fail: false }

Languages

JS/TS, Python (docstrings count as doc comments), Go, Rust, Java, Kotlin, C/C++, C#, Swift, PHP, Ruby, shell, SQL, Lua, YAML/TOML, Terraform, CSS/SCSS, HTML/XML, Haskell, Elixir, Solidity and more. Unknown file types are ignored.

Teams

Free, including team use. There is no paid tier. Use init for a shared repo setup and the GitHub Action above for pull requests.

Why a hook works when CLAUDE.md doesn't

Instructions in CLAUDE.md compete with everything else in a long context and fade as the session gets deeper. A PostToolUse hook runs outside the model, sees the exact diff and returns specific lines, so the feedback arrives right after the edit and points at concrete lines. See anthropics/claude-code#65961 for the problem in the wild.

MIT licensed. Issues and PRs welcome.