plainly

My AI first drafts all had a tell. The same few words (delve, robust, leverage), the "it's not just X, it's Y" reflex, the tidy list of three. I got tired of deleting it by hand, so I wrote plainly.

It's a Claude Code plugin. It scores non-fiction prose for the habits that make it read as machine-written, and offers to fix them. It is not an AI detector and I don't pitch it as one. These are just bad writing habits, whoever has them. Works on essays, docs, blog posts, marketing copy. Python 3.11+, no dependencies.

Install

claude plugin marketplace add Mapika/plainly
claude plugin install plainly@plainly

That gives you the /plainly:check command, the deslopper agent, and a writing skill. To run the engine on a file without the plugin:

python3 plainly/scripts/prescan.py draft.md

Using it

It watches on its own. A hook runs after Claude writes or edits any .md, .txt, or .rst file. If the result reads AI-generated, plainly says so and offers to clean it. Clean prose and code get no comment. Tune or switch it off in .plainly.toml.

/plainly:check draft.md prints findings by severity, each with the line, the bad span, a reason, and a rewrite. It reads the file. It never edits.

python3 plainly/scripts/prescan.py draft.md --json   # the raw scan

density is the number to watch: tell-weight per 100 words. Human prose sits near 0. Slop runs into the teens.

The deslopper agent edits the flagged spans and hands the text back. Just say "deslop draft.md". Two things matter here. It works a paragraph at a time and reverts any edit that adds a new tell, fails to lower the score, or flattens the rhythm, so the text can't come out worse than it went in. And it finishes with a before/after scorecard plus a blind judge: a cheap subagent that reads both versions in random order and picks the more human one without seeing the scores. It only claims a win when both agree.

The writing-clean-prose skill loads itself when Claude writes prose for you, so the tells don't show up to begin with.

In CI:

python3 plainly/scripts/prescan.py --diff --fail-over 4   # fail if a changed .md scores over 4

Config

Drop a .plainly.toml in your repo root.

Section	Key	Meaning
[severity]	critical, moderate	Cluster weight per paragraph to reach each tier.
[rules]	em_dash	The em-dash check. Off by default (noisy).
[burstiness]	min_cv	Flag docs whose sentence-length variation drops below this.
[concreteness]	min_mean	Flag paragraphs below this mean concreteness (1=abstract, 5=concrete).
[genre]	default	prose, docs, or marketing. Sets how hard marketing tells count.
[allow]	terms	Words never flagged (say, a product named "Tapestry").
[deslop]	judge	Run the blind judge after editing. On by default.
[deslop]	burstiness_tolerance	A rewrite's cv must stay at or above the original's times this (default 0.9).

How it scores

A stdlib engine does the counting: structural patterns (participle tails, "not just X, it's Y", tricolons, announcement clichés, puffery), a weighted word list of measured AI vocabulary and marketing buzzwords, evidence-backed emoji, sentence-length variation, and paragraph concreteness against the Brysbaert lexicon. Every entry is sourced in SOURCES.md. The [genre] setting decides how hard marketing tells count, so a punchy word in a technical doc isn't scored like a LinkedIn post. Then a haiku subagent catches the semantic tells a regex misses, with no API key because it runs inside Claude Code. Weights are low. The score is about how tells cluster, not any single word.

Does it work

I measured it instead of guessing. Full writeup in eval/.

Against 2022-era AI it separates human from machine well: AUC 0.82, 4.3% false positives, best on scientific writing. Against 2026 frontier models it drops to 0.67, because the obvious tells are fading. That is the honest result, and the reason I call it a quality linter and not a detector. The fixes hold up though: a blind judge outside the test set preferred the deslopped version 95% of the time, and plainly's score always moved the same direction.

The eval also caught the tool lying to me. On fresh frontier prose the score went flat. A textbook LinkedIn post ("thrilled to share", "double down", "the best is yet to come") scored 0.00. So I dug up the actual evidence and taught it the marketing register. Same post, before and after:

Requirements

Python 3.11+. No third-party packages.

License

MIT.