feat: Claude-specific humanizer rewrite — 34 patterns, severity ranking, mode system

[Save12sttm]

PR Title

feat: Claude-specific humanizer rewrite — 34 patterns, severity ranking, mode system

---

Summary

This rewrites the humanizer skill from a generic ChatGPT-era pattern list into a
Claude-specific editing tool. The original v2.2.0 was built from Wikipedia's
"Signs of AI writing" guide — solid baseline, but it was documented from ChatGPT
and GPT-4 observations. Claude has measurably different output fingerprints. The
patterns that give Claude away are largely absent from the original skill, and one
pattern (curly quotes) actively misdirects when applied to Claude output.

This PR fixes that, brings the file under the 500-line spec limit, and adds
structural improvements that make the skill more reliable for an agent to follow.

---

What Changed

New patterns (Claude-specific)

Pattern 31 — Claude Markdown Overuse (HIGH)
Claude's primary structural tell — more diagnostic than vocabulary. Bullet points
for continuous thought, random bold in prose, headers on short responses, numbered
lists where order is irrelevant. Reclassified from buried LOW patterns to a named
HIGH-severity pattern.

Pattern 32 — Post-Action Summaries (MEDIUM)
Claude habitually appends summaries of what it just did: "To recap:", "In summary:",
"Here's what was covered:". Distinct from generic conclusions. Flagged for deletion.

Pattern 33 — Unsolicited Ethical/Safety Caveats (MEDIUM)
"It's important to consider the ethical implications...", "Please ensure you...",
"Be mindful that..." — Claude's safety-training artifacts appearing as content.
Delete or fold into body only if genuinely relevant.

Pattern 34 — Second-Person Lock (LOW)
Claude defaults to "you/your" with extreme consistency. Flags saturation of
second-person in contexts not specifically addressing the reader.

Extended patterns

Pattern 1 — AI Vocabulary Words
Added confirmed Claude-specific vocabulary missing from original:
straightforward (explicitly acknowledged in Anthropic's own system prompts),
nuanced, comprehensive, leverage, holistic, meticulous/meticulously,
navigate (figurative), commendable, resonate, realm, swiftly, empower

Pattern 2 — Collaborative Communication Artifacts
Added Claude-specific openers: "I'd be happy to", "I can help with that",
"Let me [verb]" as sentence openers, "Happy to help!", "I'll help you with that"
The original only covered ChatGPT-style "Great question!" artifacts.

Pattern 19 — Hedges + Standalone Caveats
Extended to cover standalone Claude caveat lines appended after substantive content:
"Note: X", "Important: Y", "Keep in mind: Z", "A few things to be aware of:"

Fixed patterns

Pattern 30 — Curly Quotation Marks
Original applied this universally. Curly quotes are a ChatGPT/DeepSeek output tell.
Claude natively outputs straight ASCII quotes. Pattern now explicitly says: skip
for Claude output, apply only if source is confirmed ChatGPT/DeepSeek.

Pattern 6 — Passive Voice
Added explicit Formal mode exception. Passive voice is standard and professionally
correct in scientific papers, legal documents, and technical specifications.
Original applied it to every mode, which corrupts legitimate Formal content.

Pattern 1 — Cut instruction
"If no specific word fits, cut the sentence" was too aggressive. Changed to:
rewrite without the AI word first; cut only if the sentence has zero information
after removal.

Pattern 3 — Example statistic
"The policy reduced error rates by 12% in the 2022 pilot" — this invented stat
in an example could train an agent to fabricate specific numbers. Replaced with
[REAL FIGURE] / [REAL SOURCE] placeholders.

Patterns 25 + 26 — Boldface / Inline bullets
Reclassified from LOW to MEDIUM with a note that for Claude output these are
high-confidence tells. LOW severity meant agents deprioritized them. For Claude
specifically they should be caught in most passes.

Structural changes

Context mode system
Three modes — Casual / Professional / Formal — with explicit rules on what to
apply in each. Voice injection ("add soul", "let mess in") is now scoped to
Casual only. It actively harms Formal/Academic text and should never have been
applied universally.

Severity ranking
HIGH / MEDIUM / LOW tiers so the agent prioritizes correctly. Original had all
24 patterns at equal weight.

DO NOT TOUCH rules
Explicit list of what must not be modified regardless of how AI-sounding it is:
direct quotes, technical specs, legal language, and patterns appearing 3+ times
consistently (frequency heuristic for intentional stylistic choices).

Fabrication guardrail — two enforcement points
Never invent sources (Pattern 3) + never invent named people or anecdotes (STEP 2).
Both are explicit. Original had neither.

Self-audit clarified
Previously written as a pseudo-prompt to the user. Rewritten as an explicit
internal reasoning step the agent performs before producing final output.

Spec compliance — file split
Original file was 545 lines (spec limit: 500). Appendix and Reference block moved
to references/examples.md. version/changelog fields moved to CHANGELOG.md
(non-standard frontmatter; the skill router ignores them). SKILL.md is now 443 lines.

Description trigger fixed
Original description contained "this sounds like ChatGPT" — wrong for a skill
running on Claude, humanizing Claude output. Changed to "this sounds like AI / Claude".

---

File Structure

humanizer/
├── SKILL.md          — 443 lines, name+description frontmatter only
├── CHANGELOG.md      — full version history
└── references/
    └── examples.md   — worked example + reference sources

---

Pattern Count

Version	Patterns	Claude-specific
v2.2.0 (original)	24	0
v4.0.0 (this PR)	34	8

---

What Was Not Changed

• Core pattern logic from Wikipedia "Signs of AI writing" — still the foundation
• The 3-stage output format (draft → self-audit → final)
• Before/after examples on existing patterns
• GUARDRAILS structure

---

Testing

Tested against Claude Sonnet output samples across three content types:

• Casual blog post draft (markdown overuse, "I'd be happy to" opener, post-action summary)
• Professional product doc (passive voice correctly preserved in Formal mode)
• Wikipedia-style article (vague attributions, significance inflation, -ing tack-ons)

All new Claude-specific patterns triggered correctly on real Claude output.
Curly quotes pattern correctly skipped on Claude source, triggered on pasted
ChatGPT output.

---

References

• Wikipedia: Signs of AI writing — https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing
• ACL 2025 BEA Workshop paper on model-specific AI vocabulary fingerprints
• Anthropic system prompt documentation (Claude 4 series) — passive voice, markdown defaults,
  "straightforward" as an acknowledged Claude tell