Strip AI-isms from docs and memory files

Rewrites natural-language memory files (CLAUDE.md, AGENTS.md, todos, preferences, docs) so they sound human-written - stripping AI-isms and adding burstiness - while preserving every code block, URL, path, command, and heading exactly.

Two modes: --deterministic (a fast regex pass stripping canonical AI-isms and tightening tricolons, no API call or ANTHROPIC_API_KEY needed, best for batch processing and CI), and LLM mode (the default, calling Claude via the Anthropic SDK or a claude --print CLI fallback for a full rewrite that engineers burstiness and restructures performative paragraphs - slower but higher quality). Three intensity levels via --mode: subtle (stock vocabulary only), balanced (the default - sycophancy, hedging, transitions, stock vocab, authority tropes, signposting, performative balance, and an em-dash cap), and full (balanced plus filler-phrase replacement and negative-parallelism tricolon stripping, for marketing copy or slop-heavy LLM output).

humanize --deterministic --report audit.json doc.md     # writes audit + humanized
humanize doc.md                                         # optional LLM polish on top

A two-pass audit runs the deterministic pass first for a report (audit.json lists every rule that fired, each before/after pair, and counts by rule), then optionally an LLM polish on top - useful for reviewing exactly what the regex changed before trusting the diff. A FILE.original.md backup is always written before the original is overwritten; re-running after editing the backup regenerates the humanized version.

Removed categories include sycophancy openers ("Great question!", "Certainly!"), stock vocabulary (delve, tapestry, leverage as a filler verb, robust, seamless, cutting-edge, and similar), hedging openers ("It's important to note that", "Generally speaking"), authority tropes ("At its core,", "Fundamentally,"), signposting announcements ("Let's dive in", "Here's what you need to know"), transition tics ("Furthermore,", "Moreover,"), performative balance (a reflexive "however" appended to every claim), and em-dash pileups (more than two per paragraph); --mode full additionally strips filler phrases ("in order to" to "to", "due to the fact that" to "because") and negative-parallelism tricolons ("No guesswork, no bloat, no surprises."). Tightening rules collapse unnecessary three-item lists to two, merge repetitive bullet soup into one sentence, and vary paragraph length to avoid a mechanical five-paragraph-essay shape.

What's preserved exactly (byte-for-byte): fenced and indented code blocks, inline code, URLs and markdown links, file paths, shell commands, technical terms/proper nouns/API names, dates/version numbers/numerics, and environment variables - plus structural elements like headings, bullet/numbered-list hierarchy, tables (cells compressed but structure kept), and YAML frontmatter. The critical rule: everything inside a fenced or inline code region is strictly read-only - no comment edits, whitespace changes, or line reordering; humanization only touches the prose between code regions.

Processing steps: locate the SKILL.md's adjacent scripts/ directory, run python3 -m scripts <absolute_filepath> (add --deterministic for the regex-only pass), detect file type, write the .original.md backup, humanize, validate (a preserve check plus an AI-ism residual check), retry with a targeted LLM fix call up to twice on validation failure, then report success (humanized file path plus backup path) or restore the original and exit with an error on final failure.

Use it to de-slop a natural-language memory or documentation file while guaranteeing every code block, path, command, and structural element survives untouched. It only operates on .md/.txt/.markdown/.rst/extensionless prose files - never on .py/.js/.json/.yaml/.env/etc. or an already-processed *.original.md file, and it refuses sensitive paths (.env*, *.pem, *.key, ~/.ssh/, ~/.aws/) and files over 500 KB before any read or API call.

Input is a filepath to a natural-language file, plus optional --deterministic and --mode flags. Output is the humanized file (overwriting the original), an .original.md backup, and, in deterministic-report mode, an audit.json listing every rule fired with before/after pairs and counts.

LLM mode calls Claude via the Anthropic SDK or a claude --print CLI fallback; its AI-ism vocabulary was cross-referenced against blader/unslop's skill listing and Wikipedia's "Signs of AI writing" taxonomy, with a full comparison in docs/research/IMPLEMENTATION_TRACE.md.

Teams maintaining CLAUDE.md-style memory files or documentation who want AI-generated tone stripped out automatically, with a guarantee that code, commands, paths, and structure are never touched in the process.