humanize-en

Humanize EN

Strip AI writing tells from English prose. Preserves meaning, structure, code blocks, links, anchors, and frontmatter — rewrites only the flagged phrasing.

Additional context from the user: $ARGUMENTS

Scope

This skill removes AI slop. Default mode (no -f): the goal is a clean, direct, human-edited register that preserves the source voice — universal AI tells stripped, no opinion injected. If the source is an opinion piece and the user explicitly asks for voice, references/voice.md covers the optional voice-calibration pass.

Under -f <voice-doc>, the brand voice is the primary contract. Prescan and validation enforce mechanically detectable rules; the LLM reviews the remaining semantic rules. A mechanical clean result covers only those detectors, not every brand rule. Preserve source facts, code and quotations throughout.

Brand voice integration (optional)

When $ARGUMENTS starts with -f <voice-doc>, load a BRAND-VOICE.md (typically produced by /brand-voice) and treat its rules as the primary contract on top of the universal 32 patterns.

$SKILL_DIR = this skill's folder — ${CLAUDE_SKILL_DIR} in Claude Code, the directory containing this SKILL.md elsewhere.

Workflow:

1. Strip -f <voice-doc> from the head of $ARGUMENTS. The remainder follows the Input modes table below as usual.
2. Verify <voice-doc> exists and is readable. Missing explicit brand input blocks brand-aware completion; report the exact path/error instead of silently substituting universal-only success.
3. Resolve once. Find the installed brand-voice skill through the harness, then its scripts/extract_rules.py; known fallback locations are the sibling skill and ~/.agents/skills/brand-voice or ~/.claude/skills/brand-voice. Run python3 <extract_rules.py> --resolved-json <voice-doc> > <temporary-rules.json> and read that JSON as the LLM's rule contract. It resolves voice.extends and _replace/_remove through the existing resolver. On failure, preserve stderr and stop brand-aware completion. If the resolver is absent and the voice has no inheritance, read its local YAML and use --brand <voice-doc> for both checks. Inherited input requires the resolver: report npx skills add coroboros/agent-skills --skill brand-voice, then the exact extraction and rerun commands; never claim parent coverage from child-only input.
4. Run the brand-aware prescan: python3 "$SKILL_DIR"/scripts/prescan.py --rules-json <temporary-rules.json> <file>. Use the same resolved JSON for the LLM and both mechanical checks; do not rescan the original child path. Local-only fallback uses --brand instead. Brand hits carry source: "brand" and a rule_id alongside the 8 mechanically detectable universal patterns.
5. Cite per source. Name each hit by source: pattern numbers for universal (#14), brand rule_ids for brand ([no-hedging-imperative], [forbidden_lexicon:game-changing], [all_caps_emphasis]). Brand rules win on direct conflict — a voice that requires em-dashes overrides pattern #14.
6. Validate the authorized edit with python3 "$SKILL_DIR"/scripts/validate.py --rules-json <temporary-rules.json> [--baseline <pre-rewrite-hits.json>] <file> (or the local-only --brand input selected above). clean ends the mechanical pass; include semantic coverage before completion. residuals → surface and iterate within the cap; regression → repair the offending edit. Audit-only may validate a temporary proposed rewrite, leaving the target unchanged.
7. Pseudo-tables (```text or unspecified-language fences) are scanned the same as prose under -f — see Preservation rules. Real code (```python, ```bash, etc.) stays verbatim.