personifylisted
Install: claude install-skill twistedmelonman/personify
# Personify
Write text that does not read as machine-written, then verify that claim against
a detector instead of asserting it.
This is the 2.0 break, and it is a large one. Version 1.x carried a lettered
taxonomy of AI-writing tells, an A/B harness that ran two rule sets against every
text, a blind reviewer, and an evidence corpus. All of it is gone. What replaces
it: draft in the writer's own register guided by their voice guide, apply current
model judgment about what reads as machine-written, then submit the result to
Pangram's detector. A Human verdict passes. Anything else stops and reports.
Two failure modes to avoid equally: leaving AI tells in, and over-correcting into
generic dryness that has no writer behind it.
One consequence, stated up front because it will surprise a caller. A draft
written by a model and edited by a model still reads as model-written, and the
detector says so. Most Claude-drafted artifacts fail this check. That is the
expected result, not a malfunction.
## Step 0: Load the voice guide first
Before applying anything below, you MUST actually check disk for the voice guide with a real tool call (Read, or `ls` via Bash). Do not infer its presence or absence from conversation context, memory, or a prior turn. Check these locations in order and use the first that exists:
1. The path in the `PERSONIFY_VOICE` environment variable, if set.
2. `VOICE.md` under the user's config directory: `$XDG_CONFIG_HOME/personify/VOICE.md`, or `~/.config/pers