A content and skills repository for editorial review of AI-writing patterns. It packages two Claude Code skills, an offline benchmark data contract with a rank-only evaluator, analysis protocols, and historical writing artifacts.
The skills review style patterns and source faithfulness. They do not detect AI, prove authorship, or certify publication readiness — that framing is deliberate and load-bearing throughout the repository.
There is no build step, package manifest, or server. Markdown is the primary product; the Python is dependency-free, offline by default, and never calls external detector or model endpoints.
Install into Claude Code:
mkdir -p ~/.claude/skills
git clone https://github.com/zekusmaximus/humanizer.git ~/.claude/skills/humanizerThen invoke it in Claude Code:
/humanizer
[paste your text here]
Or simply ask Claude to humanize text directly. The skill reviews 24 stable, numbered patterns (see the table in Humanizer/README.md) and proposes source-faithful revisions. Anything that would add facts, opinions, experience, emotion, or voice requires explicit author input or approval.
Humanizer/SKILL.md is the source of truth; the README and WARP.md are kept in sync with it.
A deeper editorial workflow for English narrative Markdown, defined in aiproofing/SKILL.md. It runs 6 phases and 18 canonical tasks (literal IDs 1–16 plus 6.5 and 14.5), each backed by a protocol in aiproofing/protocols/ and tuned by presets in aiproofing/presets/domain_presets.md.
Initialize a run against a manuscript:
python aiproofing/scripts/aiproof_runner.py --help
python aiproofing/scripts/aiproof_runner.py Boundary/Boundary.md --preset narrative --max-edit-pct 15 --min-faithfulness 4 --require-semantic-reviewThe runner validates the manuscript and the machine-readable contract in aiproofing/scripts/task_manifest.json, then writes versioned workflow state and unsigned revision-audit scaffolding. It does not edit the manuscript itself — the editing happens through the skill workflow.
The workflow's terminal status is Internal editorial checks complete. That means the selected internal checks and required fidelity items were resolved — nothing more.
Benchmark v2 is an offline measurement contract with four non-interchangeable tracks (detector validity, editorial quality/faithfulness, span localization, cooperative provenance). It ships synthetic fixtures only; the checked-in data supports no external claims.
Migrate the synthetic v1 fixture and evaluate:
python aiproofing/benchmark/migrate_v1.py --input aiproofing/benchmark/data/example_runs.csv --output-dir tmp/benchmark_v2 --strict
python aiproofing/benchmark/evaluate.py --mode validate-rank-only --input tmp/benchmark_v2/detector_runs.jsonl --samples tmp/benchmark_v2/sample_revisions.jsonl --output tmp/benchmark_v2/summary.json --seed 20260831
python -m json.tool tmp/benchmark_v2/summary.jsonTextless legacy rows migrate as excluded stubs, raw thresholds are never invented, and rank-only mode emits no confusion-matrix metrics. See aiproofing/benchmark/README.md for the full contract, track boundaries, and allowed/disallowed claim language.
Run the full standard-library test suite from the repository root:
python -m unittest discover -s tests -p "test_*.py" -vTests cover task/file parity, failure paths, deterministic migration and bootstrap behavior, schema cross-record checks, redaction, card rendering, and historical notices. A GitHub Actions workflow (.github/workflows/p0-offline-tests.yml) runs them offline on push.
humanizer/
├── Humanizer/ # Skill 1: pattern-based humanizer (SKILL.md is source of truth)
├── aiproofing/ # Skill 2: deep narrative AI-proofing workflow
│ ├── protocols/ # 24 protocol files with declared manifest roles
│ ├── presets/ # narrative/technical/academic/business tuning
│ ├── scripts/ # aiproof_runner.py + task_manifest.json (canonical contract)
│ └── benchmark/ # offline four-track measurement contract (schema, migration, metrics, cards)
├── tests/ # standard-library unit and parity tests
├── ENHANCEMENTS.md # living roadmap
├── archive/reviews/ # historical repo reviews (preserved verbatim)
└── <artifact dirs>/ # real writing samples (see below)
Boundary/, Test story/, Mnemosyne_Cycle/, Tempus_Dimittere/, and The_Meaning_Coefficient/ hold manuscripts and historical reports. They are worked examples, not validated benchmark evidence. File-name suffixes encode roles: _AIP/_report for analysis reports, _HUM/_revised for revised output, _MERGED for consolidated versions. Reports predating Schema v2 carry a dated historical/non-reproducible notice.
- Pattern matches are style heuristics, never authorship evidence. Every check carries an evidence label:
STYLE_HEURISTIC,MEASURED_FEATURE, orHUMAN_REVIEW_REQUIRED. - Revisions must stay source-faithful. New facts, quotations, experiences, feelings, opinions, or voice require source support or explicit author approval.
- The benchmark does not validate detector evasion, provide a universal detector threshold, or combine tracks into an "authenticity" score.
- A lower detector score is one noisy measurement, not proof of human authorship and not a success criterion.
SKILL.mdfiles are the source of truth. Change behavior there first, then propagate to the READMEs (andWARP.mdfor the Humanizer).- Keep the 24 pattern numbers stable and bump skill versions in sync with their version-history entries.
- No new Python dependencies — standard library only, offline by default.
- Full conventions live in CLAUDE.md; the roadmap lives in ENHANCEMENTS.md.
- Pattern background: Wikipedia: Signs of AI writing (WikiProject AI Cleanup)
- Benchmark contract and limitations: aiproofing/benchmark/README.md