@kensio/pangram-check

Runs a finished document past Pangram, a commercial AI-text detector, and reports which passages read as machine-drafted. Packaged as an agent skill.

Only the prose is sent. The script strips frontmatter, code, HTML, shortcodes, tables and headings first, then maps every result Pangram returns back to the source line it came from, so a flagged passage is a file:line reference and not a paragraph to go hunting for.

Pangram is a paid service and the skill needs an API key. It runs when a user asks for it by name.

Install

Into any agent that reads SKILL.md:

npx @kensio/skills add pangram-check

That copies the skill directory into .agents/skills/, where Codex, Cursor, Copilot, Gemini CLI and the other implementations of the specification look for one. Pass --agent claude for .claude/skills/, --agent copilot for .github/skills/, and --user to install it for every project at once.

Claude Code also takes it as a plugin:

claude plugin marketplace add KensioSoftware/kensio.ai
claude plugin install pangram-check@kensio

Or pin it in a repository as a dependency:

npm install @kensio/pangram-check

Every skill is also published as a zip on each release, for a machine with no npm reach. Unzip it into .agents/skills/ and it is installed.

Set up the key

Get one from the Pangram dashboard, then put it where every repository on the machine can reach it:

mkdir -p ~/.config/pangram
printf 'PANGRAM_API_KEY=%s\n' 'the-key' > ~/.config/pangram/.env
chmod 600 ~/.config/pangram/.env

Then check it. This hits GET /models and spends no detection call:

node skills/pangram-check/scripts/pangram-check.mjs --check-key

$PANGRAM_API_KEY and $PANGRAM_ENV_FILE both work too, and a ./.env in the working directory is the last place searched. The key is never printed, and it is redacted from any error text.

Use it

node skills/pangram-check/scripts/pangram-check.mjs post.md

post.md
  742 words of prose, an estimated 1 billable unit.

  AI Detected
  This document was likely written with the help of a large language model.
  42% AI · 35% AI-assisted · 23% human
  segments: 2 AI, 1 AI-assisted, 1 human

  reading order  ▁▂▇▄▁▁▃  one cell per window, start to end

  7 windows, worst 3 first:

  █████████░ 0.91  AI, high confidence  post.md:112  humanized 0.73
     "the opening of the passage…"

Scores are colour-coded on a terminal. --format markdown gives a table to paste into a report, and --format json carries every field with the line references added.

What it will refuse to do

Every run costs money, so the script checks the document is worth a call before making one.

• Under 50 words of prose: Pangram's own documented floor.
• Under 300 words, which is this skill's own default and is loosened with --min-words. Pangram predicts from 50 words up, with less confidence the shorter the text.
• Over a cost ceiling set with --max-units. One billable unit is each started block of 1000 words.
• Any pattern a repository has told it to refuse: --reject-todo covers the common case, where a TODO marker means a paragraph is still to be written. --reject <regex> takes anything else.

--dry-run runs all of that without sending the document. --print-prose prints the exact string that would go out.

Results are cached by a hash of the extracted prose, so re-running an unchanged document is free. There is deliberately no score history and no before-and-after delta, because both invite treating the number as a target.

Configuration

A .pangram-check.json beside the document (or anywhere above it) sets the defaults for a repository, and flags override it:

{
  "minWords": 400,
  "rejectTodo": true,
  "format": "markdown",
  "windows": 3
}

Full flag list, extraction rules and cache locations are in reference/configuration.md. What each API field means, and which signals to act on, are in reference/reading-results.md.

What a score is for

A high score marks a passage worth rereading. It is not a number to drive down. Editing to move a detector score is a different activity from writing in your own voice, and the two come apart quickly.

Pangram reports whether text reads as machine-generated, and it has no opinion on whether the writing is any good. Detectors also carry false positives, which published audits have found falling hardest on writers working in a second language. One verdict is one input to a rereading, never evidence about a person.

Related

@kensio/technical-prose-style measures the constructions that make prose tiring to read. Its own measurements show text that follows every one of its rules still being identified as machine-written, which is the clearest demonstration that style and provenance are separate signals. Anyone after concealment is holding the wrong tool.

Part of kensio.ai. Licensed under the Apache License 2.0. See the LICENSE in the repository root.