unslopify

Text is clean when a careful reader cannot tell a model touched it. That takes two things: removing the named failure patterns, and not replacing them with a new house style. The pipeline below does both.

Run unslopify types for the rubric: four categories (formula, substance, wording, structure), each with named types, one-line definitions, and examples. The CLI comes from PyPI (pip install unslopify) or runs from a checkout as python -m unslopify.cli.

Mode 1: always-on

When the user asks for unslopify on every reply, follow the writing rules below in everything you produce for the rest of the session: chat replies, documents, commit messages, comments.

• Start where the answer starts. No scene-setting, no restating the question, no announcing what you are about to say.
• Concrete subjects and verbs. Name what changed, why, and the evidence. A number, a source, or an observable result beats an adjective.
• Prefer is, are, and has over ornate stand-in verbs (the serves-as and stands-as family).
• No manufactured rhythm: no not-just-X-but-Y frames, no slogan fragments, no triads for cadence, no colon-and-semicolon drumbeats, no em dashes.
• No AI vocabulary when a plain word works. If the word would look odd in a text message to a colleague, pick the plain one.
• Vary sentence length. Uniform sentences read as generated even when every word is clean.
• End on the last real point. No summary paragraph that re-says the piece, no trailing offers.
• Neutral register is correct for code, config, and reference material. Do not inject personality there.

These rules are contextual signals, not banned tokens. One isolated instance is fine; conspicuous, repeated, or unearned patterns are not. Em dashes and en dashes are the exception. They always fail. One instance is enough. A quoted prompt is not a waiver. The CLI treats dash as a hard mechanic: disable=dash, disable-line, file skip, --ignore-quotes, and config disable = ["dash"] do not turn it off.

Mode 2: rewrite a draft

Input: pasted text or a file path. Output: the clean text plus a short log of what changed. Work in a scratch directory; never edit the user's file in place unless asked.

Pipeline

Every gate either passes or sends you back to the rewrite step. Do not skip a gate because the draft "looks clean".

0. Preflight. Run unslopify --version. If the command is missing, install it from PyPI, and if you cannot install, STOP and tell the user: "the unslopify CLI is not available, so the audit and uniqueness gates cannot run; install it with pip install unslopify". Never substitute your own judgment for the CLI gates and report the pipeline as run. Declare the other degraded modes too, rather than hiding them. Step 5 runs the judge as a separate CLI process, so it needs no subagent support; only if no judge CLI exists AND your harness cannot spawn a fresh-context subagent, say so and skip step 5 explicitly. If the environment is ephemeral with no persistent home directory, say the phrase bank has no memory between sessions.

1. Audit. Save the draft as DRAFT.md and run:

   unslopify DRAFT.md --brief

   The brief names each finding with its line, span, problem, and direction. Mechanical problems (dashes, curly quotes, long sentences, repeated phrases) are listed with the findings.

2. Rewrite. Fix every finding by rewriting, not deleting. The rewrite covers everything the original covers: same claims, same qualifications, same paragraph count unless structure itself was the finding. If the user supplied a writing sample, build a voice profile first (unslopify.voice.profile_from_sample) and aim the rewrite at those numbers.

   Plain-language rules for the rewrite itself:

   • One main point per sentence, with an active subject and verb.
   • Standard capitalization and spelling. Never inject informality, slang, or errors to sound human; that is its own tell.
   • Quoted material and citations are evidence. Preserve them exactly, except em dashes and en dashes: those characters cannot stay, even inside a quote. Use an ASCII span or a paraphrase.
   • Say the direct thing instead of a template. If you notice the same sentence shape appearing across your rewrites, that shape is a house formula and it dies here.
   • Name a failure type in your notes only when it clarifies the edit. Do not tag types to satisfy a checklist.
   • The rewrite should not be longer than the original. Padding added to satisfy a gate is a new finding, not a fix.

   Record every edit you make as a RewriteEvent (line, rule or type id, before, after). mechanical_rewrite already logs its own events; yours join the same list, so the final report shows exactly what changed and why.

3. Judgment pass. The CLI cannot see everything. Reread the rewrite for the judgment-only types: meaning loss (a sentence compressed until the point is gone), unexplained jargon the regexes missed, synthetic balance, and uniform sentence rhythm. Fix what you find.

4. Uniqueness. Check the draft against the user's phrase bank:

   unslopify DRAFT.md --bank --id <doc-id>

   Any 8-word run shared with a previously banked document fails, and so does any 4-word phrase repeated inside this draft. Rephrase the flagged spans in different words, not reshuffled ones.

5. Fresh judge. Run the harness-neutral judge script from this skill's directory:

   python3 <skill-dir>/scripts/fresh_judge.py DRAFT.md

   It launches a separate headless agent-CLI process (first found of UNSLOPIFY_JUDGE_CMD, claude -p --model claude-opus-5-5, codex exec -m gpt-6.1-sol, cursor-agent -p). The default Claude judge uses Opus 5.5 (claude-opus-5-5), and the Codex fallback uses GPT-6.1 Sol (gpt-6.1-sol). Explicit command overrides take precedence. The judge sees only the rewritten text, never your working notes. It prints a JudgeVerdict JSON (approved, reasons, quoted_spans). Exit 0 is approved. Exit 1 means spans were quoted: go back to step 2 with those quotes as new findings. Exit 2 means no judge CLI exists on this machine; only then fall back to your harness's own fresh-context subagent with the same instruction ("Quote every span that reads as AI-generated, stock, or evasive, and say why. Approve only if nothing needs quoting."), and if the harness cannot spawn one either, declare step 5 skipped. Exit 3 means the judge produced unusable output; rerun it once before falling back. Never judge the text in your own context: a judge that watched you rewrite will approve what you approved.

6. Final gate. unslopify DRAFT.md must print PASS. Then, if the user keeps a phrase bank, commit the finished document so future drafts cannot reuse its phrasing:

   unslopify commit DRAFT.md --id <doc-id>

7. Deliver. Return the clean text, then a log: findings fixed by type name, judge verdict, and the final PASS line. Keep the log to a few lines; the text is the deliverable.

What not to do

• Do not delete content to make a finding disappear. A missing claim is a worse failure than a flagged one.
• Do not swap one stock phrase for another stock phrase. The uniqueness gate exists because rewrites drift toward new stamps.
• Do not normalize the author's voice away. Contractions, first person, short fragments, and mild informality are the author's, not yours to remove.
• Do not run the judge in your own context. A judge that watched you rewrite will approve what you approved.