AI-Humanizer-Skill: Instructions file for Claude Code

AGENTS.md

Guidance for AI coding agents (Claude Code, Codex, Warp, etc.) working in this repository.

What this repo is

A Claude Code / OpenCode skill implemented entirely as Markdown. The runtime artifact is SKILL.md: the agent reads its YAML frontmatter (metadata) followed by the editor prompt. There is no build step for the skill itself; the only code in this repo is the validation script under scripts/.

This skill is intentionally separate from humanizer. Do not merge them, and do not add cross-repo dependencies. If a change belongs in humanizer's general AI-pattern detection instead of this skill's house style, it goes in that repo, not this one.

Key files

• SKILL.md — the skill itself. YAML frontmatter (name, description, license, metadata.version) followed by the rule set: sentence rules, word rules, structure rules, tone rules, 13 numbered patterns with before/after examples, a "what not to flag" guardrail section, and voice calibration. This is the source of truth.
• README.md — for humans: installation, usage, the output contract, the pattern tables, a full worked example, and a version history.
• .claude-plugin/plugin.json — Claude Code plugin manifest. Points skills at ./ (the repo root, where SKILL.md lives).
• .claude-plugin/marketplace.json — makes the repo installable via /plugin marketplace add johnakande/ai-humanizer-skill.
• scripts/validate-package.py — checks that SKILL.md, README.md, and plugin.json haven't drifted apart. No external dependencies.
• .github/workflows/validate.yml — runs the validator, a skill-discovery check, and a Claude plugin check on every push and pull request.

The maintenance contract

SKILL.md, README.md, and the two .claude-plugin/*.json files must stay in sync. When you change behavior or content:

• Patterns: the skill currently defines 13 numbered patterns under "Patterns to catch," headed ### 1. through ### 13.. If you add, remove, or renumber any, update the README pattern tables, the "N patterns" references, and scripts/validate-package.py's expected range in the same change.
• Banned words: the word list in the "Word rules" section is a single flat list, not numbered. If you add or remove entries, keep the list alphabetized-by-convenience (grouped loosely by theme is fine) and don't duplicate an entry that's already covered by a different phrasing.
• Version: version lives in SKILL.md's metadata.version frontmatter field, .claude-plugin/plugin.json's version field, and README's first "Version History" entry. Bump all three together. Do not add a top-level version: field to SKILL.md (it belongs under metadata:), and do not add compatibility: or allowed-tools: fields. The validator rejects both.
• Output contract: for pasted text, "deliver final text only, no draft, no summary" is the defining difference from humanizer's default (which shows a draft/audit/final loop on request). File mode is the one exception, where a short after-the-fact summary is expected because the rewrite itself isn't visible inline. Don't soften the pasted-text default without the user's explicit sign-off, since it was a deliberate, repeated instruction, not a default this repo drifted into.
• Non-obvious fixes: if you change the prompt to handle a tricky failure mode (a rule that kept getting missed, an edge case in the cut list), add a short note to the README version history explaining what was fixed and why.