humanize-prose

A Claude skill for revising academic and analytical prose so it reads as human-authored rather than AI-generated. Built from empirical lessons across eight drafts of a real essay, where the GPTZero AI score moved 75% → 21% → 66% depending on what each rewrite did.

The principles also work as plain good-prose principles. Evidence density, structural asymmetry, resistance to over-refinement. Humanizing and writing well converge.

Install

Claude.ai

1. Download SKILL.md from this repo
2. In Claude.ai, open Settings → Skills and upload the file
3. Claude will automatically apply it when you ask to humanize prose or lower an AI-detection score

Claude Code

curl -L -o ~/.claude/skills/humanize-prose.md \
  https://raw.githubusercontent.com/celestialdust/humanize-prose/main/skills/humanize-prose/SKILL.md

Restart Claude Code. The skill loads automatically.

Claude API

Upload via the Skills API:

import anthropic

client = anthropic.Anthropic()
with open("SKILL.md", "rb") as f:
    skill = client.beta.files.upload(("SKILL.md", f, "text/plain"))
# pass skill.id in your Messages request

What it does

Academic prose gets flagged as AI-generated even when it is genuinely human-authored. The reason is register. By 2026, AI detectors (GPTZero, Turnitin AI, Originality.ai, Copyleaks, Winston AI) score text on three stacked signals: perplexity (how predictable each next word is), burstiness (how much sentence-to-sentence variation exists), and stylometry (the shape of the writing — how ideas flow, how punctuation is deployed, how transitions are managed). Smooth, tidy, evenly-paced expository writing scores high on AI probability because its register matches LLM output.

The skill also surfaces a 2026 bias: non-native English speakers and technical/scientific writers are systematically over-flagged because their prose naturally reads low-perplexity. If that applies to you, the skill addresses it directly.

The skill walks Claude through a six-step revision workflow:

1. Diagnose before editing. Mark each paragraph as evidence-dense, analytical, or mixed. Leave the evidence-dense ones alone.
2. Cut, do not rephrase. Removing words is the single strongest move.
3. Add evidence, not refinement. A direct quote with page citation, a named scholar, a specific date.
4. Remove AI-tell patterns. Em-dashes, semicolons, "furthermore"/"moreover"/"thus," cleft constructions, causal sentence-fusion, crisp metaphor verbs. Includes a 2026 vocabulary-cluster table (abstract verbs like "delve"/"leverage," inflated adjectives like "pivotal"/"seamless," flowery metaphors like "tapestry"/"realm") — what flags is co-occurrence density, not individual words.
5. Introduce burstiness. Follow long sentences with short ones, add hedges next to flat assertions, drop a rhetorical question, permit controlled inconsistency, vary paragraph length.
6. Verify, then stop. Detector variance is ±10–20 points. If three rewrites land within 15 points of each other, ship the best.

Empirical basis

Every principle in SKILL.md has at least one draft-to-draft comparison behind it. The full trajectory is in skills/humanize-prose/references/trajectory.md:

Draft	Body words	GPTZero AI %	What changed
v3	~1,700	69%	Initial academic draft
v4	1,862	75%	Tighter phrasing, smoother transitions
v5	1,978	68%	Added "student-voice" texture
v6	1,452	21%	Cut ~500 words; added direct quote + three dates
v7	1,442	56%	Surgical refinements to yellow paragraphs
v8	1,456	66%	v6 + one evidentiary addition

The one successful move (v5 → v6) dropped the score 47 points by cutting abstract filler and adding specifics. The two refinement moves that followed regressed the score 35 and 45 points despite feeling like improvements. Know when to stop.

Contents

humanize-prose/
├── README.md
├── LICENSE
└── skills/
    └── humanize-prose/
        ├── SKILL.md                  # The skill — workflow + red-flag scan
        ├── references/
        │   └── trajectory.md         # Eight-draft empirical record
        └── scripts/
            └── ai_tell_scan.py       # Programmatic red-flag check

Using the scanner standalone

The scan script works on any markdown or plain-text draft:

python3 skills/humanize-prose/scripts/ai_tell_scan.py your_draft.md

It reports em-dash/semicolon counts, AI-tell phrase hits, weak analytic verbs, 2026 vocabulary-cluster density (abstract verbs / inflated adjectives / flowery metaphors per paragraph), causal sentence-fusion ("X, because Y"), neat tricolons, paragraph balance, and sentence-length distribution.

Scope

• Does not teach users to pass off AI-generated writing as human. The principles only work when the content and argument are the writer's own — evidence density requires actually knowing the sources.
• Does not guarantee a specific detector score. Detectors are probabilistic and vary across runs.
• Does not apply well to fiction, text under 300 words, or heavy-jargon technical writing.
• Does not defeat cryptographic watermarks (Google SynthID) or provenance metadata (C2PA). Prose-level humanization addresses stylometric detection only — if the target pipeline uses a SynthID verifier or C2PA manifest, rewriting will not change the verdict.

License

MIT. See LICENSE.