gptzero-check

name	gptzero-check
version	2.0.0
description	Run GPTZero AI detection on one or more review files via the app.gptzero.me dashboard. Requires Chrome running with --remote-debugging-port=9222 and the user logged into GPTZero. Use when user says "check ai rate", "gptzero", "检测AI率", "run gptzero", or wants to verify that review prose passes AI detection.
license	MIT
compatibility	claude-code opencode
allowed-tools	["Bash","AskUserQuestion"]

gptzero-check

Run GPTZero AI detection on review files. All logic lives in research_harness/stages/external/gptzero_check.py — this skill just drives that script.

Install from zero (one-time)

# 1. Python CLI (auto-installs openprogram)
pip install research-agent-harness

# 2. Symlink the skill
ln -s <path-to-research-agent-harness>/skills/gptzero-check ~/.claude/skills/gptzero-check

# 3. One-time: open Chrome with CDP and log into GPTZero
/Applications/Google\ Chrome.app/Contents/MacOS/Google\ Chrome \
  --remote-debugging-port=9222 \
  --user-data-dir=$HOME/.openprogram/chrome-profile &

# Then in that Chrome window: open https://app.gptzero.me, log in. Cookie
# persists in the user-data-dir; subsequent runs just need Chrome up.

When to use

• User says "run gptzero", "check AI rate", "检测AI率", "verify AI detection"
• After peer-review or humanize-paper-review to verify results

Required inputs

• File paths — one or more review files. Supported formats: .md, .txt. Ask via AskUserQuestion if not provided.

Format conversion (non-text inputs)

If the input file is not .md or .txt, convert it first:

PDF:

cd $RESEARCH_HARNESS_DIR && \
python -m research_harness.stages.review.pdf_to_markdown "<file.pdf>" > "<file.md>"

DOCX:

cd $RESEARCH_HARNESS_DIR && \
python -m research_harness.stages.review.docx_to_markdown "<file.docx>" > "<file.md>"

Then pass the converted .md file to the scan step below.

Prerequisites check

Before running, verify Chrome CDP is reachable:

curl -s http://localhost:9222/json | python3 -c "import json,sys; print('OK:', len(json.load(sys.stdin)), 'tabs')" 2>&1

If it fails, start Chrome:

/Applications/Google\ Chrome.app/Contents/MacOS/Google\ Chrome \
  --remote-debugging-port=9222 \
  --user-data-dir=$HOME/.openprogram/chrome-profile &
sleep 4

Then verify again. If Chrome is up but GPTZero isn't logged in, tell the user to open app.gptzero.me in that Chrome window and log in.

Run the scan

cd $RESEARCH_HARNESS_DIR && \
python -m research_harness.stages.external.gptzero_check \
  "<file1>" "<file2>" "<file3>"

Output is one JSON object per line. Each line has:

• ai_pct, human_pct, mixed_pct — percentages (0-100)
• verdict — full sentence from GPTZero
• confidence — "highly" / "moderately" / "somewhat"
• status — "ok" / "no_result" / "error"

Report results

Present as a table:

File	AI%	Human%	Pass?
...	...	...	✓ / ✗

Pass threshold: AI% ≤ 20% (COLM 2026 / ACM MM standard).

If any file fails (AI% > 20%), suggest running /humanize-paper-review.

Error handling

• status: error with "CDP not reachable" → start Chrome (see Prerequisites)
• status: error with "Could not open GPTZero tab" → user needs to log in
• status: no_result → GPTZero rate-limited; wait 30s and rerun that file