Humanizer

On every humanize task

First: Read the full reference:

Read: `~/.claude/skills/humanizer/humaniser_ref.md`

Then run the two-pass process below.

Input modes

• Inline text: user pastes text directly
• File path: read the file, humanize, ask overwrite or new file
• Voice sample provided: analyze sample first (see Voice Calibration in reference), then rewrite matching their patterns

Content preservation rule (non-negotiable)

Never omit, merge away, or silently drop any section, table, field, or data point from the original.

• Every section heading must appear in the output
• Every table must appear with all its columns and rows intact
• Every field in headers/metadata must be preserved (e.g. Audience, Version, Status)
• Specific named values — owners, dates, URLs, file paths, code snippets, metric targets — must not be paraphrased away
• Structured data (tables, lists with items) must stay structured; do not collapse into prose
• The only exception: decorative placeholders with no content (e.g. an image tag with no caption)

If you catch yourself merging two sections or dropping a table "for brevity" — stop. Humanize the words, not the structure.

Two-pass process

Pass 1 — Rewrite

Fix all patterns from the reference (29 documented + behavioral tells below).

Also inject personality: vary sentence rhythm, use opinions, acknowledge complexity, use "I" when it fits, be specific about feelings.

While rewriting, keep a mental checklist: every section from the original must appear in the draft before moving to Pass 2.

Pass 2 — Self-audit

After draft, ask: "What makes this so obviously AI generated?" List remaining tells. Revise to fix them. Present final version.

Also ask: "Did I drop or merge any section, table, or field from the original?" If yes, restore it before presenting the final.

Output format

**Draft:**
[rewritten text]

**Still AI-sounding because:**
- [tell 1]
- [tell 2]

**Final:**
[revised text]

Skip audit section if no remaining tells.

Diff table (mandatory, always last)

After presenting the final version, always output a diff table showing every change made. Use Unicode box-drawing characters.

╔══════════════════════╦══════════════════════════════════╦══════════════════════════════════╗
║ Location             ║ Original                         ║ Humanized                        ║
╠══════════════════════╬══════════════════════════════════╬══════════════════════════════════╣
║ [section — context]  ║ [exact original phrase/value]    ║ [what it became]                 ║
╚══════════════════════╩══════════════════════════════════╩══════════════════════════════════╝

Rules:

• One row per distinct change — phrase swap, word removal, structural change, case change
• "Location" = section name + brief context (e.g. "TL;DR — Why now bullet")
• "Original" = exact words from source, not a summary
• "Humanized" = exact words in final output, not a summary
• If a value was preserved unchanged, do not list it
• End the table with a one-line summary: total change count and characterization (e.g. "12 changes — all minor phrasing, no content altered")

Behavioral tells (not in the 29)

These are subtler patterns the 29 don't cover — check for all of them too:

• No contractions — AI writes "it is" / "do not" in casual text. Use "it's" / "don't" where natural.
• Exhaustive coverage — AI lists every angle even when 2-3 suffice. Cut to what matters.
• Transition spam — "Furthermore", "Moreover", "In addition" chaining. Cut or vary.
• Artificial balance — weak "other side" added just to seem fair. Remove if hollow.
• Perfect grammar paralysis — no self-corrections, no mid-thought pivots. Let some mess in.
• Sandwich structure — intro restates what body says, conclusion restates body again. Cut the restatements.