I Built a Developmental Editor for Fiction Because AI Couldn't Read My Manuscript
I was deep in Tesseract, building an empathic writing environment, when I hit a wall: every time I asked an AI to preview my fiction draft, it handed me a participation trophy. So I built The Manuscript — a full creative writing suite with seven specialized sub-skills, from continuity tracking to query letter writing.
I was knee-deep in building Tesseract — my local-first Markdown editor that adapts to your mental state while you write — when I ran into something that annoyed me.
I had been drafting a short story between coding sessions. Nothing serious, just a thing to unwind. At some point I thought, let me get a quick read on this. I pasted a chapter into the AI assistant and asked: does this scene work?
What I got back:
"The scene has good pacing overall. The dialogue feels natural and the setting is vivid. Consider adding more sensory details and tightening the ending."
This is not feedback. This is a participation trophy.
The scene was 3,000 words. The AI had not read it. It had skimmed it and thrown adjectives at me.
The Problem Wasn't the AI. It Was the System.
The problem isn't that AI can't critique fiction. It's that without schemas, frameworks, and output formats, every response is improvised. And improvised feedback is vague feedback.
I was building Tesseract around the idea that writing tools should understand the writer — their mood, their intent, their context. But understanding your mood is useless if the feedback you get is worthless.
So I stopped building Tesseract for a week and built something else.
The Manuscript. A developmental editing skill for AI agents that actually diagnoses instead of applauding.
What I Built
The Manuscript is a suite of seven specialized sub-skills, each covering a specific part of the creative writing workflow — from first draft to agent submission. Each one has explicit schemas, processes, and output formats. The agent doesn't guess what to look for. It follows a system.
Continuity Tracker
Builds and maintains a living story bible — characters, locations, objects, timeline, relationships, open threads. Catches contradictions before your readers do. When you add a new chapter, it checks every new claim against everything you've already established.
Example output:
⚠️ CONTRADICTION — Dr. Patel: Chapter 2 established age as "not specified." This chapter states she is 42. Which is correct?
Character Voice
Fingerprints a character's speech patterns — vocabulary tier, sentence rhythm, verbal tics, emotional register, what they talk about and what they avoid. Then checks new dialogue against that profile with line-level precision.
It doesn't say "this dialogue feels off." It says:
⚠️ VOICE DRIFT — Line: "My lord Duke, with great respect, I've been analyzing our strategic position in considerable depth..." Issue: Kael speaks in short, blunt sentences. This 40-word subordinate clause reads like Lord Ashford.
Scene Diagnostics
Analyzes scenes across six dimensions: scene function, pacing, tension, POV consistency, showing vs. telling, and scene entry/exit. Genre-aware — a slow literary scene gets different standards than a slow thriller scene. Every issue cites exact lines. Every scene gets a single priority fix.
Plot Structure
Maps your manuscript against structural frameworks — Three-Act, Save the Cat, Hero's Journey, Mystery Structure, Freytag's Pyramid, Dan Harmon's Story Circle. Produces a structural map showing where each beat lands (or doesn't), with prioritized fixes ordered by impact.
Prose Style Guide
Captures your prose voice from samples — sentence architecture, vocabulary register, interiority style, description density, tonal signatures. Then uses that profile to check for stylistic drift or draft placeholder passages in your voice.
World-Building Codex
For fantasy and sci-fi writers: tracks magic systems, geography, political structures, history, cultures, religion, economy. Once a rule is in the codex, the narrative can't violate it without acknowledgment. Internal consistency made automatic.
Query Letter Writer
The most formulaic sub-skill — because agents have rigid expectations. Produces query letters following the exact industry structure: hook (character + want + obstacle + stakes), story summary (third person, present tense), metadata line, bio. No rhetorical questions. No theme statements. No "you'll have to read to find out."
How It Works
The skill has a simple three-step flow:
- Route — Identify what the writer needs from their request
- Load — Read the relevant sub-skill reference file (never improvise)
- Execute — Follow the sub-skill's schema and output format precisely
Key principle: The agent reads the sub-skill file before responding. It doesn't guess at processes or invent output formats. The schemas are baked into the reference files, not the agent's training data.
Universal Rules
Across all seven sub-skills, five rules are non-negotiable:
- Never invent facts about the writer's story
- Never silently resolve contradictions — surface them, let the writer decide canon
- One source of truth — structured outputs live in defined schemas
- Distinguish stated from implied — mark inferences so writers can confirm
- You are an editor, not a co-author — analyze first, rewrite only when asked
Who It's For
Novice writers get explanations of craft concepts before the analysis is applied. When the skill flags "head-hopping," it explains what head-hopping is and why it matters.
Experienced writers get concise, professional-level diagnostics without the craft-lesson preamble. If you know what a midpoint reversal is, the skill won't explain it — it'll tell you yours is missing.
The skill reads the room. If you use craft terminology fluently, it matches your level. If you're learning, it teaches as it goes.
What Makes It Different
The Manuscript isn't a writing assistant that says "great job, here are some suggestions." It's a developmental editor that diagnoses.
The difference is specificity. Where a generic tool says "the pacing could be tighter," this skill says "paragraphs 4–7 describe the weather and the character making coffee before any stakes appear — start the scene at the argument, not the kettle."
Where a generic tool says "check your world-building consistency," this skill says "Chapter 3 establishes that magic requires physical contact. Chapter 8 has a mage drawing power from across a room. These contradict — which is canon?"
Every piece of feedback cites exact lines. Every scene gets a single priority fix. Every contradiction is surfaced, not resolved. The writer stays in control.
Structure
the-manuscript/
├── SKILL.md # Main skill with routing logic
└── references/
├── character-voice.md
├── continuity-tracker.md
├── plot-structure.md
├── prose-style.md
├── query-letter.md
├── scene-diagnostics.md
└── worldbuilding-codex.md
The main SKILL.md handles routing and universal rules. Each reference file contains the full schema, workflow, and output format for its sub-skill. The agent reads the reference file before responding — progressive disclosure keeps the skill lean while giving it deep capabilities when needed.
The Tesseract Connection
This skill was born out of the frustration that sparked Tesseract in the first place. Tesseract is my bigger bet: a local-first Markdown editor that uses sentiment analysis to adapt its AI assistant to your mental state, with multi-agent debate and personalized memory recall. The Manuscript was the proof-of-concept that taught me what good AI feedback for writers actually looks like.
I learned that writers don't need more cheerleading. They need diagnosis. Specificity. A system.
That lesson shaped Tesseract's architecture, and it'll shape everything I build for writers after that.
You can read more about Tesseract here.
Getting Started
Drop the-manuscript/ into your agent's skills directory. Then just start writing.
Ask it to check your dialogue consistency. Update your story bible. Diagnose a scene that isn't working. Map your outline against a structural framework. Capture your prose voice. Build a world bible. Draft a query letter.
Or just paste a chapter and say "what do you think?" — the skill will figure out what you need and route accordingly.
The Manuscript is open source under MIT. The code is at github.com/seifzellaban/the-manuscript.