No Slop — De-Slop Writing Engine

AI-written text has recognizable tells — em dashes everywhere, 'it's not just X, it's Y' constructions, hedging opens, hype words — and mechanically swapping them out isn't the whole fix: word-level detectors are already being replaced by ones that read sentence-structure burstiness, so a tool that only launders vocabulary will still read as generated.

One dependency-free engine in src/deslop.js does the real work — regex-driven word and phrase swap tables, opener and transition detection, and a separate flagsFor() pass that only detects the antithesis family of constructions rather than trying to auto-fix them. That same engine backs a live browser diff view, a POST /api/deslop endpoint served by a custom Vite plugin rather than a separate server, and a CLI for piping text through the terminal.

Fixes are tiered on purpose: AUTO rules (punctuation, filler openers, vocabulary, hype phrases) get rewritten automatically, FLAG rules (the 'it's not X, it's Y' antithesis family) are only surfaced with a fix instruction because rewriting them safely takes judgment, and a further HUMAN tier covering structural issues is advisory-only, documented in no-slop-rules.md rather than encoded as code. The slopometer score weights that antithesis flip pattern hardest of any signal, since it's treated as the single most reliable tell.

Past the mechanical pass, a Rewrite tab calls a local Claude or Codex CLI to fully rewrite a draft in a learned voice — built from a hand-picked seed example plus a growing corpus of past examples, retrieved by plain lexical cosine similarity rather than embeddings, and assembled into a few-shot prompt. The tool tests itself against a real corpus of the author's own draft captions, tagged by pipeline stage and sorted worst-first by slop score, so the examples browser doubles as both a regression set and a gallery of the voice it's aiming for.

django_api/ ports the same engine line-by-line to pure Python (no Django import in engine.py), fronted by Django REST Framework: a custom PlainTextParser so raw text/plain bodies work like the Node API, a serializer enforcing the same 4MB cap with a real 400, and AnonRateThrottle standing in for the Node origin-header guard. Parity wasn't assumed — four representative inputs were run through both engines and diffed field-by-field (clean text, fix counts/categories, slop score, signal breakdown), and a fresh adversarial review pass caught a real divergence anyway: Python's banker's rounding disagreed with JS's Math.round() at exact .5 boundaries, fixed with a _js_round() helper and pinned with its own regression test. The same review pass also deleted a dead serializer whose docstring falsely claimed test coverage, and fixed a throttle test that would have passed even if throttling were completely broken. 14 DRF tests now guard the port, verified green from a fresh venv install.

A single dev-server process serves the browser tool, the API, and the CLI identically at a pinned local port, with an origin check guarding the API against anything but localhost. It runs as a standing part of a personal writing workflow — piped from the clipboard, wired into an editor shortcut, given to a coding agent as a rule to run every final draft through before handing it back, or called as an MCP tool directly from an agent's own context.