Zero Slop

  Zero Slop is installed about 1,600 times a week on npm.

npx skills add manavmishra/ZeroSlop --global

Before and after

A launch post, as AI wrote it:

We're thrilled to announce that our team has leveraged cutting-edge machine learning to deliver a seamless onboarding experience, reducing setup time by 40%.

The scorer rates that draft 99.3/100 and flags four phrases: “We're thrilled to,” “leveraged,” “cutting-edge,” and “seamless.”

The rewrite, limited to the draft's stated claims:

We used machine learning to reduce onboarding setup time by 40%.

Writing score: 9.5/100  [clear]
  Flagged phrases : 0 across 10 words

The rewrite retains the draft's stated result. See four complete, reproducible pairs in examples/.

Use it for launch posts, changelogs, emails, research summaries, or batch checks. Scores describe writing patterns, not authorship.

Related MCP server: TrueVoice MCP

Install

Try the free browser editor, or install:

Ask your assistant to edit:

/zero-slop (your writing)

Inspect without editing:

/zero-slop inspect (your writing)

Score locally:

npx zero-slop score draft.md

From a cloned checkout, gate a folder:

python3 scripts/slopscore.py --batch drafts/ --gate 25

Installed checks run locally. Skill editing follows your assistant's privacy settings; MCP editing is remote.

Hosted MCP

https://mcp.zero-slop.ai/mcp

Connection options and listing status.

CLI

Edit a file through MCP:

npx --yes zero-slop@2.12.1 deslop draft.md --genre professional

Use - for stdin and --json for structured output. --require-approved exits nonzero when review is needed; the result is still printed. Files stay unchanged. Node.js 22+; offline score also needs Python 3. CLI reference and privacy.

REST API

curl --fail-with-body --max-time 75 https://mcp.zero-slop.ai/v1/deslop \
  -H 'Content-Type: application/json' \
  --data '{"text":"Maya owns the pricing review.","genre":"professional"}'

Same pipeline and result as MCP. Check status before using the edit. Free, shared capacity; up to 20,000 Unicode code points per draft after trimming. Hosted CLI editing and REST process drafts remotely without storing them.

API reference · OpenAPI contract

What the workflow adds

Zero Slop ships no model. Your AI assistant reads and edits the draft in context, using Claude, GPT, or another compatible model. The repository supplies the workflow and local tools for scoring and source checks.

What it catches

The scorer combines 294 weighted patterns with a 96-term lexicon. Examples include:

• binary contrast formulas: “It's not X. It's Y.”

• canned openers: “We're thrilled to…” and “Here's the thing…”

• vague attribution: “experts agree” and “studies show”

• significance inflation: “marks a pivotal moment” and “a testament to”

• promotional riders: “robust,” “seamless,” and “leverage” when used as hype

• repeated sentence shapes, crowded statistics, and overworked formatting

Marketing terms are scored in context, so an ordinary technical use of a word need not trigger the same penalty. references/eval.md documents all 80 checks.

Unedited AI drafts averaged 77 in bench/examples.json. Human writing scored 9 to 21 in data/corpus/must-not-flag/. These are reference points for the scorer, not authorship boundaries.

How it works

Eight responsibilities form one workflow. They are jobs, not separate models. Research supports the checks, not the number eight, which is an engineering choice.

The free web editor combines the five AI responsibilities into one response and makes at most one live model call. A single response does not provide independent review. Any final change receives one final local recheck.

If repair fails, return the safest edit with a warning. No open-ended rewrite loop.

Review the reader's side

Ask: “Review this for backend engineers. Where would they stop reading? Don't rewrite it.” Optional skim, passage reactions, and notes-only recall produce comments and revision strips. Simulations cannot establish human behavior; sequential review requires isolated contexts. The workflow explains privacy, limits, and reporting. Hosted calls are unchanged.

Inspired by First Reader, a review-only skill. Zero Slop also rewrites and checks source details. Neither has established human-reader accuracy; First Reader gets no invented rewrite score.

Evidence and limits

Same model, same 18 drafts

A saved replay ran Zero Slop and three comparable open-source instruction sets over the same drafts with GPT-5.4, high reasoning, and pinned instructions. The outputs are frozen and reproducible.

This small LLM-reviewed regression study measures repeatable behavior; it does not establish universal writing quality. The drafts, hashes, method versions, prompts, and limitations are in bench/README.md.

Zero Slop's frozen outputs came from v2.5.9; newer versions only rescore those saved outputs. The current scorer matched the prior 84.2% result on the fixed 38-item editorial panel. These fixed-sample checks are not field accuracy.

• A method-hidden editorial preference replay: bench/incumbent-blind-replay/

• External-checker clean rates: assets/bench-external-checker.png

• Method-hidden quality ranking: assets/bench-blind-quality.png

• Current-model corpus measurements: assets/bench-raid-plus.png

• Antithesis regression set: assets/bench-antithesis.png

On the 75 labelled antithesis pairs, the current reading pass reached 91.2% recall across the full set, 100% recall on shapes in reach, and 100% precision. The labels are maintainer-authored and the pairs are constructed, so this is a regression floor rather than field accuracy.

Local speed measurements cover the checks, with editing time excluded. On one Apple silicon Mac, the scorer processed 1,000 documents in a median of 1.9929 seconds (501.8 per second); the five runs ranged from 1.9614 to 2.0945 seconds. It scored a 15,201-word document in a median of 0.3223 seconds. The slowest stress case took 2.2932 seconds, and learning from an 8,000-word edit took 0.1592 seconds. The measurements and machine details are in bench/performance-results.json.

Across 12 interleaved runs against 2.7.7, we measured 0.26% lower median throughput, within the 5% regression limit. Local timing, not an SLA. The two-way replay used Zero Slop v2.6.0.

The RAID+ audit checks how the scorer responds to output from different models. Its pinned sample contains 7,627 usable generations:

RAID+ labels record which model produced each text; they do not grade writing quality. The Beemo paired-edit audit checks how scores change after human editing: raw responses averaged 30.2, expert edits 25.3, and human answers 20.0. Beemo also lacks writing-quality labels.

Documented capability audit

This chart says nothing about writing quality or which tool writes better. It records documented features at pinned commits; the data and reproduction notes are in bench/README.md.

The design follows research on predictable wording in machine text and overused vocabulary. It deliberately avoids authorship claims because detectors can misclassify non-native English.

Private learning

Learning begins only when you provide an original output and your reason-labelled edit. Zero Slop does not monitor files, browsers, or publishing tools. Private data stays under $ZERO_SLOP_HOME; it is not committed to this repository and does not retrain the model.

A profile selected by name can exempt existing watchlist words. It does not learn cadence, tone, or a complete writing style.

Repository map

Contributing and support

Bug reports, false positives, examples, and carefully tested pattern improvements are welcome. Read CONTRIBUTING.md before opening a pull request, use the structured issue forms, or start a Discussion.

For setup help and responsible disclosure, see SUPPORT.md and SECURITY.md.

Credits

Ideas from First Reader, no-ai-slop, humanizer, de-slop, stop-slop, unslop-text, and avoid-ai-writing.

License

MIT