| name | platitude |
| description | Check drafts of user-facing English prose (tweets, posts, blog text, READMEs, marketing copy, PR descriptions) for AI slop — rhetorical shapes doing the work content should do. Use after writing or substantially editing any prose a human will read. Returns a verdict plus the exact spans and reasons, so the text gets fixed rather than regenerated.
|
platitude
Slop is a shape, not a vocabulary: the manufactured contrast, the hollow
reversal, the false profundity. You cannot reliably see it in your own
output. Measured: an open-ended model self-review flags a third of genuinely
good texts and still misses slop; platitude's closed-question protocol
scores AUC 0.86 against a blind-labeled corpus.
Run it
platitude --json "the draft text"
platitude --json --file draft.md
Or, if the MCP server is registered, call the check_slop tool.
The host needs one of: ANTHROPIC_API_KEY, OPENAI_API_KEY, a logged-in
claude CLI, or a logged-in codex CLI.
Read the output
verdict (slop|clean), slop_likelihood (0–100), and findings — each
with form, name, span, excerpt, and a one-sentence reason.
Fix loop
- Rewrite only the flagged spans, guided by each finding's reason.
- Re-check. Unchanged text replays from cache and costs nothing.
- Stop when the verdict is clean, or when the only remaining findings are
deliberate voice.
The hard rule
The negation-then-substitution family — "it's not X, it's Y", "X isn't the
problem, Y is", "the question isn't whether…", "not just X…" — is always
flagged and always slop. Rewrite it as a direct statement. Zero exceptions,
and content is no defense.
Do not
- Do not run it on code, configs, logs, or non-English text — it judges
English rhetorical structure only.
- Outside the hard-rule family, do not overrule a clean verdict just
because a form is present.
