/humanizer

name: humanizer
description: 'Strip AI-writing tells from prose destined for humans — docs, PR/commit bodies, reports, release notes, blog posts, Slack/email drafts. Removes LLM-vocabulary clichés (delve, boasts, testament, underscore, robust, tapestry...), banned constructions (not just X but Y, rule-of-three triads, "-ing" superficial-analysis clauses, vague-attribution weasel words), and formatting tells (title-case headings, mechanical bolding, em-dash overuse, curly quotes, bare-bullet inline-header lists). TRIGGER when: user asks to humanize/polish/de-AI a piece of text or file; before finalizing a substantial human-facing prose artifact drafted as part of the current task (docs, PR/commit body, report, blog post, release notes, external message) — self-review pass, best-effort model-initiated, not a guaranteed intercept. SKIP when: output is a terse conversational chat reply, code, JSON/YAML/config, a machine-parsed agent envelope ("Return ONLY:"), a `.reports/` file or a file with report-schema frontmatter (report format needs the block and bold labels intact), or the target is an ultra-caveman-tier handover file (`.temp/`, inter-agent prose per `plugins/CLAUDE.md` compression tiers). Preserve ordinary publishing frontmatter and edit only the prose body.'
argument-hint: '[text or file path to humanize] | check <file>'
allowed-tools: Read, Edit
effort: medium

<objective>

Detect and remove statistical AI-writing fingerprints from human-facing prose before it ships. Grounded in Wikipedia's crowd-sourced AI-detection corpus (`Wikipedia:Signs of AI writing`) — a maintained list of vocabulary, syntax, and formatting patterns over-represented in LLM output vs human baseline. Apply as a final pass, not a rewrite-from-scratch: preserve meaning, facts, structure; only excise the tells.

</objective>

<inputs>

• **text or file path to humanize**: optional. Inline text, or a file path (Markdown/plain text) to edit in place.
• **check `<file>`**: read-only mode — report findings, no editing.
• No argument: humanize the draft already composed earlier this turn (self-review pass) — only reachable when the model chooses to invoke this skill mid-task; no platform hook guarantees a pre-send interception, so treat this path as best-effort, not a hard gate.

</inputs>

<workflow>

1. Load the target text

• Inline text: work on it directly, no file I/O.
• File path: `Read` the file.
• **Destination guard** (before any edit): target must be human-facing prose. JSON/YAML/config, code, a machine-parsed agent envelope (`Return ONLY:`), a `.reports/` file or a file with report-schema frontmatter (the report header fields `Title`, `Date`, `Scope`, `Outcome`, `Confidence`, and `Path` — report format requires that block and bold labels such as Confidence), or an ultra-caveman handover file under `.temp/` → emit `**! BLOCKED — humanizer does not apply to <destination type>**` and make no edit. `check` mode reports without editing, so it is unaffected.
• Ordinary frontmatter in Marp decks, Jekyll/Hugo posts, MkDocs pages, and create outlines is not a blocker: preserve the metadata block verbatim and edit only the prose body. Preserve Marp slide separators and directives.
• No argument: treat the draft already composed earlier this turn as the target.

2. Scan against the checklist

Walk the text once per category below; enumerate every hit internally before editing anything (report-first, matches `check` mode output). The isolated-hit threshold in `<notes>` is the authority on which of those hits reach the report or an edit — an incidental lone match in an otherwise clean passage is scanned but not reported.

**Vocabulary — cut or replace with plain equivalent:**

| Banned | Plain replacement | | -- | -- | | delve, boasts, testament, underscore(s), showcase, tapestry, intricate/intricacies, meticulous, robust, vibrant, pivotal, crucial, garner, foster(ing), align with, landscape, interplay, enduring, enhance | say the specific thing instead — drop the word, don't swap in another vague one | | "stands as", "serves as", "marks a", "represents" (as copula dodge) | "is" / "was" | | "Additionally,", "Moreover,", "It is important to note that" | delete, or state the fact directly |

**Syntax — flag and restructure:**

• Negative parallelism: "not just X, but Y" / "not X, but Y" / "not only X but also Y" / "X rather than Y" used as a crutch
• Rule-of-three triads used for false comprehensiveness — members vague or abstract ("fast, reliable, and scalable"). Three concrete, distinct facts are not a triad tell. A list already flagged under another pattern (e.g. inline-header list) is not re-flagged as a triad.
• "-ing" superficial-analysis tails: "highlighting...", "underscoring...", "contributing to..." tacked onto a claim with no source
• Vague attribution / weasel words: "industry reports", "observers", "experts argue", "some critics" with no named source
• Formulaic "Despite its [positives], X faces challenges..." conclusion pattern

**Formatting — flag and fix:**

• Title Case In Headings → sentence case
• Mechanical bolding of every instance of a repeated term
• Markdown overuse — bold/bullets/headers where a plain sentence reads fine; the single most common tell in PR bodies and reports
• Bare-bullet inline-header lists (`• **Header:** text`) where prose or a real table reads better
• Em dash overuse — chain of `—` clauses instead of periods/commas
• Curly ("smart") quotes/apostrophes mixed inconsistently with straight ones
• `---`/`***` thematic breaks before headings (Markdown artifact bleeding into prose)

3. Apply fixes

• `check` mode: skip edits — report findings (category, severity, location, quote, suggested fix), then continue to Step 4 for the Confidence block. Severity follows the step 2 checklist group: Syntax hit = **high**, Vocabulary hit = **medium**, Formatting hit = **low**. Order findings high → low so the reader gets a cut order.
• Edit mode: apply minimal edit per flagged