humanize

Anti-slop framework for Claude Code. Skill + scorer + agent + hook + always-on rule. Strips 38 AI-writing patterns. Domain-aware (academic / docs / blog / commit). Extends blader/humanizer.

What this is

humanize is six tools that work together so AI slop has no clear path into your prose:

Layer	Component	Triggered by
1	Standing rule in ~/.claude/CLAUDE.md	Loaded into every session
2	Always-on rule file at ~/.claude/rules/10-anti-slop.md	Same
3	/humanize slash command	Manual invocation
4	PostToolUse hook on Write/Edit	Fires automatically when Claude writes prose
5	humanizer-reviewer subagent	Called by /ship, /review-paper, on demand
6	humanize_score.py CLI	Fast scoring; exits non-zero above threshold

Plus four domain profiles (academic, docs, blog, commit) with their own carve-outs. Em-dashes are fine in scientific prose. Passive voice is correct in IMRaD methods. Vague attributions are flagged hard in academic but allowed (with citation) in docs. The framework adapts.

What it detects (38 patterns)

Inherits 29 from blader/humanizer (significance inflation, promotional language, em-dash overuse, rule of three, AI vocabulary, copula avoidance, sycophancy, filler phrases, hedging, generic conclusions, …) — full list with before/after in SKILL.md.

New extensions (#30-38):

#	Pattern	Why it matters
30	Citation laundering ("studies show" with no citation)	Academic-killer
31	Manuscript boilerplate ("To the best of our knowledge…")	Generic paper opener
32	Tutorial-script scaffolding ("Let's walk through…")	Doc tutorial-script feel
33	Stat parade without effect size	Frequentist hedging
34	Temporal hedge ladders	Stacked time-disclaimers
35	Polysyndetic tripleting	Stronger rule-of-three
36	AI-flavoured commit verbs ("improves", "enhances")	Commit-specific
37	Methodology pseudo-precision ("careful", "rigorous", "comprehensive")	Self-praise without specifics
38	Dissertation-grade hedging where stance is required	Academic-only

Install (Claude Code, global)

# Clone
git clone https://github.com/kimhons/humanize.git ~/.claude/skills/humanize
# Activate the agent
cp ~/.claude/skills/humanize/agents/humanizer-reviewer.md ~/.claude/agents/
# Wire the hook (one-line append to settings.json — see scripts/install.sh)
bash ~/.claude/skills/humanize/scripts/install.sh

For OpenCode users:

git clone https://github.com/kimhons/humanize.git ~/.config/opencode/skills/humanize

Install (manual, single-file)

mkdir -p ~/.claude/skills/humanize
curl -L https://raw.githubusercontent.com/kimhons/humanize/main/SKILL.md \
     -o ~/.claude/skills/humanize/SKILL.md

Usage

Manual invocation

/humanize [paste your text]

/humanize --profile=academic [paste]

/humanize --voice=path/to/sample-of-my-writing.md [paste]

CLI scoring

$ python ~/.claude/skills/humanize/scripts/humanize_score.py STAGE3/MANUSCRIPT.md
humanize_score: 38.4/100  (minor_residue)
profile:        academic  (3,420 words)
top offenders:
  - methodology_pseudo                weighted=12.50
  - significance_inflation            weighted=7.50
  - citation_laundering               weighted=5.00
  - hyphenated_pairs                  weighted=2.40
  - excessive_hedging                 weighted=1.50

# JSON output for CI / pipelines
python humanize_score.py --json --profile=docs README.md

Hooked into commit time

The hook auto-fires after any .md / .tex Edit/Write, prints a warning when score > 60. Threshold tunable via HUMANIZE_THRESHOLD=70 in your shell.

Subagent for deep review

> Have the humanizer-reviewer audit this manuscript

The agent loads the file, runs the scorer, identifies top-5 offenders by line, rewrites them, self-audits, and reports back.

Voice calibration

Drop a sample of your own writing and the skill will match your rhythm, vocabulary, and quirks rather than producing a generic "clean" rewrite:

/humanize --voice=blog-archive/2024-09-thoughts.md
[paste AI-generated draft]

Domain profiles — at a glance

Pattern	academic	docs	blog	commit
Em-dash overuse	OK in moderation	flag	flag	flag
Passive voice	OK in IMRaD methods	flag	flag in active voice	flag
Hedging	report-grade OK	flag	flag hard	flag hard
Stat parade	flag hard	flag	n/a	n/a
Citation laundering	flag hard	flag	flag	n/a
Manuscript boilerplate	flag hard	n/a	n/a	n/a
AI commit verbs	n/a	n/a	n/a	flag hard

Why six layers?

The blader/humanizer skill is excellent but passive only — Claude only humanises when you type /humanizer. Every other write goes through unfiltered. AI slop compounds across sessions because future sessions index "this is how we write here."

humanize closes the loop:

• The rule in ~/.claude/CLAUDE.md puts the directive into every session prompt.
• The hook catches AI patterns at write-time, before they're committed.
• The scorer gives a number you can grep in CI and pre-commit.
• The subagent does the heavy review on demand.
• The skill does the manual cleanup with voice calibration.
• The profiles stop us from "fixing" passive voice in a methods section or em-dashes in an academic paper.

Acknowledgments

This work would not exist without:

• blader/humanizer — MIT, the foundation
• Wikipedia: Signs of AI writing — pattern catalog
• WikiProject AI Cleanup — maintains the source

License

MIT — see LICENSE.

Author

Kimal Honour Djam (@kimhons)