Ww Copy Sweep

Part of Claude Code operator skills: a collection of skills for running a real Claude Code setup.

ww-copy-sweep

AI-tell punctuation and phrasing creep into drafts one write at a time. The usual fix is a cleanup pass after the fact: sweep the file, catch what you can, hope you got it all. That pass shouldn’t need to exist: the write should fail before the tell ever hits disk. That’s what this is.

ww-copy-sweep is a hook, not a slash command. You never invoke it. Once it’s installed, every time Claude (or you, via Claude Code) writes or edits a matching file, the hook scans the new content first. If it finds a violation, the write is blocked and Claude gets a line-numbered list of exactly what to fix, so it fixes it inline, in the same turn, before the bad version ever exists.

No proofreading step. No “sweep the repo for em dashes” cleanup task. The thing that would have needed cleaning never gets written.

What it catches

Deterministic, on by default, zero model calls, sub-millisecond:

Semantic, off by default, needs a model call: vague_pronoun, rule_of_three, pontificating. These catch patterns a regex genuinely cannot. Read the Semantic checks section below before turning any of them on, it explains the real tradeoff (nondeterminism, latency, small API cost).

Each check is independent. Turn any of them on or off with one env var; see Configure, below.

The em-dash budget

Zero em dashes in short-form copy (a Slack message, a DM, a short email: anything under COPY_SWEEP_LONGFORM_WORDS, default 400 words). One allowed in a long-form piece (an article, a long post). That one dash isn’t pre-approved just because it’s within budget. The hook can count occurrences; it can’t judge whether a specific em dash is actually earning its place versus being a lazy default. That judgment call is still yours. A second em dash in the same piece is over budget regardless, and gets flagged like anything else.

Install: just tell Claude

Paste this to Claude Code, no dev experience required:

Install ww-copy-sweep from https://github.com/lee-fuhr/claude-operator-skills and walk me through my config options.

If you are the Claude reading this because a user asked you to install ww-copy-sweep, follow this script instead of dumping the manual steps on them:

1. Clone the repo (or fetch just this file plus hooks/copy_sweep.py if you can’t clone) and copy skills/ww-copy-sweep/hooks/copy_sweep.py to ~/.claude/hooks/copy_sweep.py, chmod +x it.
2. Read ~/.claude/settings.json if it exists. If it already has a PreToolUse hook for Write|Edit|MultiEdit, add copy_sweep.py as another entry in that same hooks array. Don’t clobber whatever’s already there. If the file doesn’t exist, create it with the block from the manual steps below.
3. Ask the user these questions, in plain language, one at a time or as a short batch, not as env var names:
   • “Want this watching every markdown file, or just a specific folder like drafts/?” → sets COPY_SWEEP_INCLUDE if they name a folder, leave unset otherwise.
   • “Do you write solo and want me to also flag any stray ‘We’ that should be ‘I’?” → adds we_pronoun to COPY_SWEEP_CHECKS if yes.
   • “Want the deeper checks too: vague pronouns, three-item-list rhythm, rambling openers before the point, short-choppy-sentence rhythm? These call an external AI model each time so they’re a bit slower and need a free Groq API key (or DeepSeek). Want those on, or stick with the instant checks?” → if yes, add vague_pronoun,rule_of_three,pontificating,staccato_cadence to COPY_SWEEP_CHECKS, and check whether GROQ_API_KEY or DEEPSEEK_API_KEY is already set; if neither is, tell them plainly these checks will just silently do nothing until a key is added (fail-open, not broken).
   • “Roughly how many words counts as ‘long’ for you before an em dash is allowed?” → sets COPY_SWEEP_LONGFORM_WORDS if they want something other than the 400-word default.
   • “Any words or phrases specific to your own writing you want banned, beyond the default corporate-speak list (leverage, synergy, touch base, and friends)?” → sets COPY_SWEEP_BANNED_PHRASES (this REPLACES the default list, so combine their additions with the defaults above if they want both).
4. Write the final command string with the resolved env vars baked in, save settings.json.
5. Run the Try it check yourself to confirm it’s wired up, then tell the user in one or two plain sentences what you set up: no JSON, no env var names in the summary.
6. Tell them to restart Claude Code (or start a new session) for the hook to take effect.

Prefer to do it yourself? The manual steps and every config flag are below.

git clone https://github.com/lee-fuhr/claude-operator-skills.git
cd claude-operator-skills

mkdir -p ~/.claude/hooks
cp skills/ww-copy-sweep/hooks/copy_sweep.py ~/.claude/hooks/copy_sweep.py
chmod +x ~/.claude/hooks/copy_sweep.py

Then add the hook to your Claude Code settings: ~/.claude/settings.json for every project, or .claude/settings.json in a single repo if you only want it there:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Write|Edit|MultiEdit",
        "hooks": [
          {
            "type": "command",
            "command": "python3 ~/.claude/hooks/copy_sweep.py"
          }
        ]
      }
    ]
  }
}

Restart Claude Code (or start a new session) so it picks up the settings change. That’s it. No skill to remember to invoke, no command to run. It just runs.

[!NOTE] If you already have a PreToolUse hook for Write|Edit|MultiEdit, add copy_sweep.py as a second entry in that same hooks array rather than a second matcher block. Claude Code runs every hook that matches, in order.

Try it

echo '{"tool_name":"Write","tool_input":{"file_path":"/tmp/test.md","content":"Great copy — right here."}}' \
  | python3 ~/.claude/hooks/copy_sweep.py; echo "exit: $?"

exit: 1 with a line-numbered violation list means it’s wired up correctly. A clean file (or any path outside your configured scope) exits 0 silently.

Configure

Everything is an env var, everything has a working default; set these in your shell profile, or directly in the settings.json hook command string ("command": "COPY_SWEEP_CHECKS=em_dash python3 ~/.claude/hooks/copy_sweep.py").

Common setups:

Scope it to a drafts/ folder only, leave everything else alone:

"command": "COPY_SWEEP_INCLUDE=drafts/ python3 ~/.claude/hooks/copy_sweep.py"

Solo-voice writer who wants I enforced too:

"command": "COPY_SWEEP_CHECKS=em_dash,straight_quote,title_case,banned_phrase,throat_clearing,summary_crutch,we_pronoun python3 ~/.claude/hooks/copy_sweep.py"

Your own banned-phrase list instead of the default:

"command": "COPY_SWEEP_BANNED_PHRASES='revolutionize,paradigm shift,at the end of the day' python3 ~/.claude/hooks/copy_sweep.py"

Semantic checks: off by default, read this first

vague_pronoun, rule_of_three, pontificating, and staccato_cadence catch real AI tells that a fixed pattern genuinely cannot:

• vague_pronoun: “this” or “that” standing in for a specific noun the reader has to reconstruct, with no clear antecedent nearby.
• rule_of_three: a list of exactly three items sharing identical grammatical shape (noun-noun-noun, verb-verb-verb) with no variation.
• pontificating: a long preamble or throat-clearing before the actual point, instead of getting to the substance directly.
• staccato_cadence: a run of short, choppy declarative sentences back to back (a sentence, then another short sentence, a fragment, a punchline) instead of longer sentences connected by commas or parentheticals. The “LinkedIn beat poetry” pattern, distinct from genuinely concise writing, which uses fewer words, not choppier sentences.

They’re off by default because turning them on is a real tradeoff, not just “more rules”:

• Nondeterministic. A regex either matches or it doesn’t. These need a model to make a judgment call, and judgment calls vary. The same text can get a different verdict on a different run.
• Slower. Every gated write places a real network call instead of a sub-millisecond scan. On a flaky connection this can visibly stall the tool call.
• Not free. A small API cost per call. Groq’s free tier covers most use; DeepSeek Flash is the fallback at roughly $0.14 per million tokens either way. Cheap, but not zero like the regex checks above.

They’re fail-open by design: any model error, timeout, or unparseable response means the check silently produces nothing rather than blocking your write over a flaky API. Worst case, they just don’t catch anything that round; they never turn a working session into a stuck one.

To turn them on, list them explicitly in COPY_SWEEP_CHECKS:

"command": "COPY_SWEEP_CHECKS=em_dash,straight_quote,title_case,banned_phrase,throat_clearing,summary_crutch,vague_pronoun,rule_of_three,pontificating,staccato_cadence python3 ~/.claude/hooks/copy_sweep.py"

Needs GROQ_API_KEY and/or DEEPSEEK_API_KEY in the environment (see the main README’s Model setup section) and the requests package (pip install requests). No key set means these checks silently no-op, same as a model error.

Why a hook and not a skill

A skill needs to be invoked: you have to remember to run /clean-up-my-copy after the fact, on the whole file, hoping you catch everything. A hook fires on the exact write that would have introduced the problem, catches it in a few lines instead of a whole document, and Claude fixes it as a two-second inline correction instead of a separate pass. The cleanup pass this replaces never has to exist.

Key principles

• Prevent, don’t proofread. Block the bad write before it happens, not after.
• Deterministic by default, model-graded is opt-in. The default checks are regex: free, instant, always-consistent. Semantic judgment is a real tradeoff (see above), never silently on.
• Budgets, not blanket bans, where a blanket ban is wrong. A single well-placed em dash in a long piece isn’t the tell; careless or repeated use is. The check reflects that.
• On by default, off by one flag. SKIP_COPY_SWEEP=1 for the rare intentional exception; no config file to maintain for the common case.
• Scoped to prose, not code. Fenced code blocks are always exempt from every check.