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. Agree the length FIRST. Before planning anything, ask the user one question: "Full novel (~70–90k words) or novella (~20–30k)?" Then pass explicit arguments — plan_book("...", genre, target_chapters=40, target_words=80000) for a novel, or omit for the default (30 chapters ≈ 48k words). Never accept the bare default without confirming: agents that "just go" produce novellas when the customer asked for a book, and stop mid-book to ask about tools instead of finishing the manuscript.
2. Identify the genre. Don't guess. Cross-check against list_genres and get_genre_recipe(genre) before writing. Mislabeled genre = mis-voiced book.
3. Read the recipe. It returns pacing beats, dialogue ratio, and stylistic anchors from the tradition. Read them all before planning.
4. 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.
5. Plan before drafting. plan_book produces the architecture. Use it — don't freestyle 40 chapters without a map.
6. Draft EVERY chapter, in chapter order, without stopping. The common failure: the agent drafts chapters 1–5, summarises the rest, and calls it done. A book isn't done until every planned chapter exists at its full target length. If you have a per-session step budget, write the next chapter at the start of each new turn rather than summarising.
7. 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.
8. 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. Don't call it repeatedly for the same book; call once, hold the outline in context, work chapter by chapter.
• Default is 30 chapters ≈ 48k words (novella-length). A full novel needs explicit target_chapters=40, target_words=80000 (cap is 40 chapters). Confirm this with the user before planning (step 1).
• The summarise trap: the #1 way agents "finish" a book is writing chapters 1–5 fully and compressing the rest into one-line summaries. evaluate_book catches thin chapters — run it before declaring done, and re-draft any chapter under its per-chapter target.
• Long agent sessions get cut off. Persist progress: keep the assembled manuscript in a file (assemble_book markdown) and append chapter by chapter so an interrupted session resumes instead of restarting.
• search_corpus supports FTS5 syntax (quotes, AND/OR/NOT, * wildcards) but long operator 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.