anti-prose-slop

Stop AI writing agents from shipping generic prose.

Stop Making Prose Slop

Write distinctive, publishable longform with the Proseify MCP — a curated public-domain corpus plus a guided drafting pipeline.

Prerequisite: connect the Proseify MCP first. No key, no tools.

{
  "mcpServers": {
    "proseify": {
      "type": "http",
      "url": "https://mcp.proseify.xyz/mcp",
      "headers": { "Authorization": "Bearer YOUR_KEY" }
    }

Get a key at https://proseify.xyz (sign in → pick a plan → paste your config).

The one rule

Genre before prose. Every generation decision after the first sentence depends on picking the right tradition. A gothic ghost story and a Regency romance want different sentence lengths, different climates, different endings. Call get_genre_recipe before writing; don't use plan_book to skip it.

The workflow

1. Identify the genre. Don't guess. Cross-check against list_genres and get_genre_recipe(genre) before writing. Mislabeled genre = mis-voiced book.
2. Read the recipe. It returns pacing beats, dialogue ratio, and stylistic anchors from the tradition. Read them all before planning.
3. Search the corpus. Use search_corpus for 3–5 model passages that match the register you want. Quote a line to yourself; match its rhythm.
4. Plan before drafting. plan_book produces the architecture. Use it — don't freestyle 40 chapters without a map.
5. Draft chapter by chapter with write_chapter (or write_book for one-shot). Match the recipe, not your defaults.
6. Self-edit between drafts. Anti-slop pass, every time:
   • Cut "seemed to", "began to", "settled into".
   • Kill adverb-stacking ("quietly, almost reverently…").
   • Delete sentences that only restate the previous one.
   • Check chapter openings — vary them. Never six chapters in a row that open with weather.
   • Read the last page out loud; if it sounds like a summary, rewrite it.
7. Verify with evaluate_book. If the score is below your own bar, don't ship it — revise the weakest chapters and re-evaluate.

Anti-slop checklist (per chapter)

• Voice matches the recipe, not "ChatGPT normal".
• Dialogue keeps its register (era, class, region).
• Scene boundaries do work — each has a job, not just a location change.
• No repeated sentence-opening constructions within three paragraphs.
• Concrete nouns beat abstractions. "The bell rang" beats "A sound occurred."

Traps the corpus protects against

• The thesaurus trap. Don't swap synonyms for variety; swap for precision.
• The cliffhanger trap. One per book, not one per chapter.
• The filter-words trap. "She felt / saw / heard / noticed" — cut them.
• The dialogue-as-exposition trap. Nobody explains the plot over tea.

Pitfalls

• plan_book result is large (~10k tokens). Don't call it repeatedly for the same book; call once, hold the outline in context, work chapter by chapter.
• The 30-chapter default is a floor, not a ceiling — most books want 30–40.
• search_corpus supports FTS5 syntax (quotes, AND/OR/NOT, * wildcards) but long operators chains are clamped to 10 terms.

Licensing

Books sourced from Project Gutenberg and similar public-domain corpora; safe commercial use. See the corpus page on https://proseify.xyz for per-book details.