Deterministic scan
The CLI audits locally and records stable findings, input SHA-256 and a ruleset/config fingerprint.
ACTIVE SKILL · PYTHON · MIT
Audit writing patterns. Preserve the facts.
A source-traceable local skill for detecting AI-style writing risk and guiding full-article repair. It combines deterministic scanning with an agent-led editorial workflow without claiming authorship probability.
1. Deterministic auditLocal rules · protected text · stable finding IDs
detect↓
2. Editorial diagnosisStructure · specificity · evidence · repetition
risk↓
3. Agent-led repairFull rewrite when explicitly requested
repair↓
4. Post-repair auditRe-scan and report unresolved fact gaps
verify01 / WHAT IT DOES
Risk signals, not authorship guesses.
The tool measures patterns that can make writing feel formulaic or AI-generated. It explicitly does not output an “AI probability” and does not claim to know who wrote the text.
Fact-preserving repair
When the user asks for a full polish, the Agent rebuilds the article while preserving supplied facts, protected text and factual limits.
Tool-agnostic Skill
The same SKILL.md is designed for Codex, Claude Code and Hermes Agent without requiring an MCP server or provider-specific API.
02 / COMPOSITION
Traceable ideas, deliberately integrated.
A project counts as an active component only when its ideas or rules are actually integrated. Merely reviewed repositories are not presented as part of the skill.
avoid-ai-writingv3.23.1 · workflow/detector reference and core pattern provenance
humanizerv2.9.1 · pattern taxonomy and selected language / structure rules
no-ai-slopLocked main snapshot · sentence-level patterns, minimum-effective-edit and voice-preservation gates
Native layerai-writing-audit 0.4.0 · Chinese, B2B marketing, SEO/GEO, deterministic scoring and repair policy
03 / USE
Run the scanner locally.
The CLI implements detect mode. Full repair is intentionally an Agent workflow rather than a fake CLI mode that silently behaves like detection.
python3 scripts/audit.py article.md \
--mode detect \
--language auto \
--profile general \
--format markdown
Protected content includes quotations, code, tables, YAML front matter, HTML attributes, legal text and attributed third-party text. Findings may be reported there, but repair rules must not rewrite it casually.
General by design
Company-specific brand facts, product constraints, legal policies and private editorial rules do not belong in this public skill. Those are intentionally left to downstream integrations or private overlays.