Strengthen humanizer factual integrity

[CoveMB]

Title

Strengthen factual integrity and add banned-pattern reference

Summary

This MR updates the Humanizer skill from 2.5.1 to 2.6.0.

The main change is stricter factual integrity. The skill now tells the model not to add facts, names, numbers, sources, quotes, examples, prices, dates, or claims unless the user supplied them. This reduces a common failure mode where a rewrite sounds more concrete by inventing plausible details.

It also adds two new AI-writing pattern categories, separate reference file, and updates the README and WARP guidance so manual installs include that reference file.

What Changed

• Added a hard factual-integrity rule to SKILL.md.
• Added a Hard Rules section covering:
  • no invented details
  • no em dashes by default
  • no forced rule-of-three structure
  • no contrast framing
  • no rhetorical hooks
  • no fake naming
  • no self-narration
  • no chatbot wrapper text
• Added references/banned-list.md for dense cleanup passes.
• Added two new detected patterns:
  • 30. Fake Naming
  • 31. Self-Narration and Rhetorical Hooks
• Expanded signposting detection with phrases like "nobody talks about this" and "nobody tells you this".
• Changed the process so the anti-AI audit happens privately and the user receives the final rewrite by default.
• Updated the full example to avoid invented studies, named people, fake quotes, and unsupported metrics.
• Updated README installation instructions to copy both SKILL.md and references/.
• Updated README pattern count from 29 to 31.
• Updated WARP guidance to document references/banned-list.md.

Why

The previous example and process could encourage the model to make vague writing feel more concrete by adding unsupported specifics. This MR makes the safer behavior explicit: preserve only supplied facts, keep vague claims general when evidence is missing, and ask for missing information when needed.

The new reference file also keeps SKILL.md readable while still giving the skill a fuller list of phrases and patterns to scrub when the input is especially dense with AI-style writing.

Files Changed

• SKILL.md

  • Bumps version to 2.6.0.
  • Adds factual-integrity guidance and new pattern categories.
  • Updates process and output behavior.

• references/banned-list.md

  • Adds the longer banned-pattern reference used for deeper cleanup passes.

• README.md

  • Updates install instructions, overview, pattern count, references, example, and version history.

• WARP.md

  • Documents the new reference file and updates manual install guidance.

Validation

• Ran git diff --check origin/main...HEAD.
• Confirmed the upstream branch diff contains only:
  • README.md
  • SKILL.md
  • WARP.md
  • references/banned-list.md

No runtime tests were added because this repository is a Markdown-based skill with documentation and prompt changes only.