slopcheck

by vicnaum·1 month ago

Keep an AI agent's writing readable for one specific reader, with named rule breaks instead of a score. A standard-library Python checker plus five Claude Code hooks, one of them a per-turn reminder of the rules: an advisory line under every long reply, a report into the agent's context after it writes any document, a deny gate on artifact publishing, and a deny gate on Slack, pull-request, Linear and Notion text until the list is empty. Checks the hard shape rules (sentence length, comma count, semicolons, em-dashes, lists run inline, paragraph and bullet length, fact-carrying parentheses), the tone tells regex can see (achievement words for unfinished work, agreement openers, sign-post phrases, contrast frames, filler words), and an advisory list of terms whose first use has no gloss. Optional model layer, off by default, through the Anthropic API or the TypeSafe Jev API, for inflation, unexplained jargon, bare numbers and metaphors. Optional structure layer, installed by `slopcheck install-structure`, that parses each sentence with spaCy in its own venv and names what regex cannot see: a sentence the parser cannot read, no verb, an appositive wedged between subject and verb, four nouns in a row, a seven-word subject, and GPT-2 surprisal spikes on request.