prose-sanitiser

name

prose-sanitiser

description

Audit and de-slop prose for AI writing tells (lexical, structural, and narrative), strip technical AI provenance marks (invisible Unicode watermarks, statistical sampling fingerprints, C2PA/EXIF/XMP metadata), and enforce UK English. Use when writing or editing public-facing content — blog posts, docs, tutorials, articles, presentations — or short fiction, or when asked to "sanitise this", "de-slop", "remove AI tells", "strip watermarks", "remove provenance", or "make this read human".

Prose Sanitiser

Strip LLM fingerprints from text — both the stylistic tells a human reader spots and the technical provenance marks a machine detector finds. Output should read as if written by a competent human with opinions, carry no invisible Unicode carriers or sampling watermarks, and ship with clean file metadata.

UK English throughout. No exceptions.

Four layers:

• Generative — apply the Section A principles when drafting new content.
• Destructive — run the Section B audit on existing text.
• Simplification — LLM-driven rewrite to plain English (simplify, declaudish strengths) for readability without manual line-by-line fixes. Useful as a first pass before the destructive audit, or standalone when the goal is readability not watermark evasion. Inspired by claudish-to-english.
• Watermark — strip technical AI provenance marks (Section E).

Quick path

1. Strip invisible marks first. Run inspect_text.py / clean_text.py to remove Unicode watermark carriers (Layer A). This is lossless and always safe.
2. Simplify (optional, fast path). If the text is dense AI prose and you want a quick first pass, run rewrite_text.py --strength simplify (or declaudish for Claude-specific tells). This is an LLM-driven rewrite to plain English that can save manual editing time. Add --context "the original question" for better results. Skip this if you prefer manual control.
3. Scan for stylistic tells. Run scripts/slop_scan.py to catch the mechanical writing tells a regex can see.
4. Fix in priority order — high-severity findings first (see the reference sections for the full catalogues).
5. Do the human read — narrative defaults, altitude, voice, and whether a sentence is actually true. The scanner is blind to all of these.
6. Statistical rewrite (optional). If the text must also defeat token-sampling detectors, run rewrite_text.py last — it changes wording, so do it after all editorial choices are final.
7. Strip file metadata. Before publication, run clean_file.py to remove C2PA manifests, EXIF, XMP, and document properties from exported files.

# Layer A — invisible Unicode carriers (lossless, always first)
python3 inspect_text.py <path>                      # report invisible chars
python3 clean_text.py <path>                        # strip them
python3 clean_text.py <path> --stats                # strip + report counts

# Stylistic tells
python3 scripts/slop_scan.py <path>                 # full report + slop score
python3 scripts/slop_scan.py <path> --severity high # only the strongest signals
python3 scripts/slop_scan.py <path> --json          # machine-readable, for CI

# Layer B — statistical watermark attack (lossy, last)
python3 rewrite_text.py <path>                      # default strength (paraphrase)
python3 rewrite_text.py <path> --strength simplify  # plain English, short sentences
python3 rewrite_text.py <path> --strength declaudish # targets Claude-specific tells
python3 rewrite_text.py <path> --strength simplify --context "What does X do?"
python3 rewrite_text.py <path> --strength humanize  # defeat AI detectors
python3 rewrite_text.py <path> --min-chars 200      # skip short texts

# File metadata
python3 inspect_file.py <path>                      # report metadata found
python3 clean_file.py <path>                        # strip metadata

# Aggregate audit
python3 audit_dir.py <directory>                    # recursive directory sweep
python3 audit_website.py <url>                      # crawl and scan a published site

The slop scanner reads .md .markdown .mdx .txt .rst, skips fenced code and blockquotes, respects the slop-ignore marker, and reports each finding with file:line and the fix. Its exit code is the high-severity count, so CI can gate a docs build on it. It sees lexical and structural tells only.

The watermark tools (inspect_text.py, clean_text.py, rewrite_text.py, inspect_file.py, clean_file.py, audit_dir.py) come from watermarks-remover. Core text cleaning needs only Python 3.10+ stdlib; c2patool, exiftool, and qpdf extend file metadata coverage when present on PATH.

Don't launder slop into new slop (second-order defaults)

The failure mode of every de-slop pass is swapping one default for another. Kill every "leverage" and the prose acquires a different fingerprint: uniform "use", staccato two-word fragments ("Fast. Actually fast."), the same inverted "X, not Y" cadence on every other line, hedges amputated until the voice reads as clipped and machine-confident. An editor can clock a de-slopped-by-AI draft as fast as a slopped one. The replacement vocabulary, applied mechanically, is itself a tell.

So the rules in the references are a detector, not a target. The replace-with column is a prompt to make a choice, not a lookup table to apply on autopilot. The only durable property is the one a default can never have: a wording you chose for this sentence and can say why. Vary the repair. Sometimes "leverage" wants "use", sometimes "lean on", sometimes the clause should be cut. If a fix introduces a new uniform default, it is not a fix.

Reference sections

Load the relevant catalogue on demand — don't hold all of it in context at once:

• Section A — Generative Principles: lead with value, show don't tell, honest trade-offs, audience framing, write from experience. Read this when drafting new content.
• Section B — Destructive Audit: the mechanical catalogue — em-dash density, "The X" headings, negative parallelism, Tier 1/2 vocabulary tables, throat-clearing, hedges, structural tells, transitions, passive voice, UK spelling, and Claudish structural patterns (B13), insider voice in external documents (B14), and preamble setup labels (B15). Read this when auditing existing text.
• Section C — Narrative Tells (Fiction): the StoryScope-derived structural defaults (thematic over-explanation, embodied emotion, single-track plots, tidy resolutions, per-model fingerprints, and more). Read this when sanitising stories or character-driven prose.
• Section D — Checklist, Output Format, and Scope: the pre-publish checklist, the report format, when NOT to sanitise, and the slop-ignore marker.
• Section E — Watermark Removal: technical AI provenance marks — invisible Unicode carriers (Layer A, lossless), statistical sampling watermarks (Layer B, lossy rewrite), file metadata (C2PA/EXIF/XMP), pixel-domain image watermarks, aggregate auditing, and the HTTP service API. Read this when stripping machine-detectable marks or preparing files for publication.