Src

When to use

• Direct user request — the user pastes text and asks whether it was AI-written, or any phrasing of that question ("is this AI?", "did ChatGPT write this?", "human or AI?", etc.). This is the only auto-trigger case.
• Proactive post-WebFetch check (opt-in only) — skipped by default because every call costs the user a Pangram credit. If the user has added an instruction to their CLAUDE.md along the lines of "when researching, run did-ai-write-this on WebFetch'd pages before citing", honor that. Otherwise wait for an explicit ask — do not call this skill on every WebFetch.
• Not for stylistic AI-detection guesses based on writing patterns — those are unreliable and this skill exists precisely to replace them with a calibrated, vendor-backed signal.

How to invoke

The CLI sits next to this SKILL.md and runs inside a self-contained venv populated by install.py. Always invoke through ${CLAUDE_SKILL_DIR} so the path works regardless of where the skill is installed.

Positional argument (short snippet, one shot):

${CLAUDE_SKILL_DIR}/.venv/bin/python ${CLAUDE_SKILL_DIR}/cli.py "the text to check"

From a file (longer documents):

${CLAUDE_SKILL_DIR}/.venv/bin/python ${CLAUDE_SKILL_DIR}/cli.py --file /path/to/document.txt

From stdin (piping output of another command, common for WebFetch content saved to a temp file or var):

cat /tmp/fetched.txt | ${CLAUDE_SKILL_DIR}/.venv/bin/python ${CLAUDE_SKILL_DIR}/cli.py --stdin

Per-paragraph attribution for mixed documents — adds a windows array showing which segments drove the overall label:

${CLAUDE_SKILL_DIR}/.venv/bin/python ${CLAUDE_SKILL_DIR}/cli.py --full --file /path/to/mixed_doc.txt

Force the output format if needed: --json (always JSON) or --pretty (always one-line summary). Default behavior is JSON when stdout is captured (your Bash tool case) and pretty when stdout is a terminal (user case), so the flags are usually unnecessary.

Interpreting the output

Default JSON output (what you receive when invoking through Bash):

{"label": "AI", "fraction_ai": 0.94, "fraction_ai_assisted": 0.04, "fraction_human": 0.02, "chars": 1284}

Fields:

• label — one of "AI", "AI-Assisted", "Human", "Mixed". This is Pangram's overall verdict.
• fraction_ai, fraction_ai_assisted, fraction_human — floats in [0, 1] summing to 1.0. The breakdown explains a "Mixed" label and gives you a confidence sense even when the label is decisive.
• chars — length of the submitted text.

With --full, the response also includes:

• windows — list of per-segment classifications, each with text, label, ai_assistance_score, confidence, character offsets, word count, and token length. Use this when the overall label is "Mixed" and you need to know which paragraphs are AI.

Decision heuristic: treat fraction_ai >= 0.5 as "do not cite as human-authored". For "AI-Assisted" and "Mixed" labels, surface the verdict to the user before citing — the source may still be usable but the AI involvement should be disclosed.

Errors

The CLI exits non-zero with a stderr message on failure. Map:

Exit	Meaning	What to do
0	Success — verdict on stdout	Use the result
1	Generic / unexpected error, including unwrapped network errors from requests	Surface stderr to the user; check network
2	PANGRAM_API_KEY missing or .env not loadable	Tell the user to re-run python install.py from the cloned repo
3	Pangram rejected the API key (HTTP 401 — bad key or out of credits)	Tell the user to check their Pangram dashboard for credits and key validity
4	Pangram server error after one automatic retry (5xx)	Surface; suggest retry later. Pangram-side outage
6	Input below 50 words	Don't retry with the same text — either gather more or skip the AI-detection step entirely for this snippet

Privacy note

Text passed to this skill is sent to Pangram's API (text.api.pangram.com/v3) for classification. Don't run it on private or confidential content unless the user has authorized third-party processing of that content.