An agent skill that strips the statistical fingerprints of LLM writing out of prose before it goes out the door: emails, PR descriptions and review comments, docs, blog drafts, essays. It runs as a plain Markdown skill (SKILL.md), so any harness that supports skill-style instructions can use it.
Personify tracks its own pattern list on its own schedule. It started as a structural cousin of blader/humanizer (same idea, same MIT license, credited in LICENSE), but the taxonomy, wording, and version history here are independent and not synced against that project.
/plugin marketplace add smartwatermelon/personify
/plugin install personify@personify
Once installed, invoke it as /personify:personify.
Clone or copy SKILL.md into the project's skill directory so it's available to collaborators who check out the repo:
mkdir -p .claude/skills/personify
cp SKILL.md .claude/skills/personify/Copy it into your user-level skills directory instead, so it's available in every project:
mkdir -p ~/.claude/skills/personify
cp SKILL.md ~/.claude/skills/personify/Reload or start a new session after installing.
Claude Desktop reads skills from its own skills directory. Copy SKILL.md into a personify/ folder there, matching the layout above, then restart Claude Desktop to pick it up.
The entire runtime artifact is SKILL.md. Any agent harness that loads Markdown-based skills can use it by copying that one file into wherever the harness expects skill definitions.
Ask for it directly, or point it at a file:
Personify this text: [paste text]
Personify the writing in docs/launch-post.md
SKILL.md covers the standard taxonomy of AI-writing tells (inflated significance, empty vocabulary clusters, copula avoidance, filler and hedging, chatbot residue, and more), plus patterns added after reviewing specific pieces of writing that leaned on AI-adjacent techniques without being AI-written. It also has a dedicated section for GitHub PR descriptions and review comments, where the tell is usually structural (unearned headers, defensive completeness) rather than prose-level. It documents what not to flag too, so a clean human writer who hits one of these patterns once isn't treated as a false positive. See the file itself for the full pattern list and provenance notes.
Personify on its own makes prose non-robotic but not distinctive: clean, competent, and anonymous. The other half is an optional voice guide.
At load time (SKILL.md Step 0), Personify looks for a VOICE.md, checking in order: the PERSONIFY_VOICE environment variable, then ~/.config/personify/VOICE.md (honoring XDG_CONFIG_HOME), then the skill's own directory for repo-local development. The first one found wins, is read in full, and is treated as authoritative: where the general pattern list and the voice guide disagree, the voice guide wins. If none is found, Personify runs in generic mode and says so.
VOICE.md describes how one specific person actually writes, compiled from a corpus of their own writing (blog posts, essays, long-form email, docs). It is personal and git-ignored, exactly like .env. The committed VOICE.example.md documents the structure and how to build one, without containing anyone's actual voice.
Install the skill however you like, then put your voice guide at the stable path so upgrades never touch it:
mkdir -p ~/.config/personify
cp VOICE.example.md ~/.config/personify/VOICE.md # then edit, or have an agent build itDon't keep your real VOICE.md inside the installed plugin directory: plugins install into a version-pinned path that is replaced on every upgrade, so a guide kept there is lost the next time the plugin updates.
MIT. See LICENSE for the full text and provenance note.