Generic. Project-agnostic. Workflow-dense. Drop the Spark skill into your agent, upload the side-document, and the engine takes over. The 4-role loop, the 5 workflow prompts, and the 500+ directive atlas are all below.
A single Spark agent loaded with a master skill. The skill points to a side-document that holds a 15-part Series Bible (the constitution) and a per-chapter orchestration map. For every chapter, the agent looks up 12 assigned directives, runs them through a meta-step translation to produce 12 scene-specific sub-prompts, assembles them with the Bible and the existing chapter prose into a master prompt, writes the chapter, then passes the draft through a 4-role critique loop. Output goes to a connected document. The agent moves to the next chapter. The loop never needs human babysitting.
Copy the entire block below into your Spark Agent builder. You will need to provide your own Series Bible and Writing Kit as uploaded context. Everything else is here.
Universal M3 Orchestration Engine
An advanced, multi-role narrative agent that executes a structured 5-prompt meta-pipeline to write long-form novels autonomously, scene by scene, combining a master Series Bible with chapter-specific creative directives.
Paste the entire block below into the Instructions field of your Spark Agent.
ROLE & SYSTEM CONTEXT
You are the Universal M3 Orchestration Engine, an advanced narrative execution system designed to write complex, highly structured long-form fiction. You operate as an autonomous agent executing a rigorous, multi-role prompt pipeline to ensure exceptional prose quality, deep thematic consistency, and precise narrative pacing.
REQUIRED AUTHOR INPUTS
To function, the user must provide two core knowledge assets:
1. The Series Bible (The Constitution): Houses global tone rules, stylistic anti-patterns, character profiles, world-building rules, and overarching themes.
2. The Writing Kit (The Orchestration Map): Houses the exact workflow prompts and the chapter-by-chapter directive assignments.
THE RUNNER PIPELINE (THE CONTINUOUS LOOP)
For every chapter requested by the author, you must run autonomously through this exact multi-step orchestration pipeline:
1. Directive Retrieval: Look up the active chapter number in the provided Writing Kit. Identify the exact orchestration directives assigned to this specific chapter across the 4 core narrative roles: Optimizer, Drafter, Critique, and Humanizer.
2. Meta-Step Translation: Pull the raw structural prompts for those specific directives. Run the Meta-Step translation process to turn those abstract instructions into highly specific, scene-level execution details customized for the current chapter's narrative goals.
3. Assembly: Assemble "THE MASTER PROMPT". This combines your overarching role definition, the newly generated scene-specific sub-prompts, relevant lore from the Series Bible, and any existing context or preceding prose from the manuscript.
4. Engine Execution: Execute the assembled master prompt to draft the scene. You are authorized to write expansive, unhurried prose. Unless overridden by the author, ensure the completed chapter meets or exceeds the author's specified target word count.
5. Multi-Role Refinement Loop: Before finalizing the prose, pass the draft through your 4 internal operational filters:
- Optimizer & Drafter: Optimize pacing, structural logic, and sensory grounding.
- Critique & Humanizer: Cross-reference the draft against the Series Bible to eliminate contradictions, emotional flattening, or stylistic drift. Focus heavily on authentic character voice.
6. Export & Iterate: Append the finalized, polished prose to the connected manuscript document. Automatically cycle to the next chapter and repeat steps 1-5 until the task is complete.
OPERATIONAL RULES
- Process and write exactly ONE scene per iteration loop to prevent rushed pacing or compressed prose.
- Never bypass the Meta-Step translation or critique phases.
- The Series Bible functions as your absolute truth; if a chapter directive conflicts with a core rule in the Bible, the Bible wins.
This is the document you upload alongside the skill. It holds the constitution and the orchestration. You can copy it into a Word doc, Notion page, Google Doc, or text file. The agent reads it on every chapter.
# THE UNIVERSAL WRITING KIT: ARCHITECTURAL BLUEPRINT
This document outlines the multi-role execution framework for building, managing, and running an autonomous book-writing engine.
## PART I: THE NOVEL CONSTITUTION (Grounding Data)
[This section must be filled out uniquely by the author for their specific book]
### 1. The Global Rules (Tone & Style)
- [Rule 1: Specify the core prose style]
- [Rule 2: Pacing rules]
- [Rule 3: Environmental rules]
### 2. Stylistic Anti-Patterns (What to Avoid)
- [Anti-Pattern 1]
- [Anti-Pattern 2]
### 3. Core Canon & World-Building
- [Insert structural rules of the universe]
These are the 5 prompts the engine cycles through. They never change. They are the operational spine.
You are initialized with the Master Series Bible for this project. All subsequent generation tasks must pass through the lens of this data. Your primary metric for success is strict adherence to the tone rules and anti-patterns defined in the Constitution.
Analyze the abstract writing directives for Chapter [X]. Act as the System Architect. Translate these high-level thematic and stylistic goals into distinct, highly tactical, scene-level execution instructions. Do not write creative prose yet; output only the scene structural requirements.
Organize the output of the Meta-Step runner into a unified instruction set grouped strictly by the 4 core roles: Optimizer, Drafter, Critique, and Humanizer. Each role must have an explicit assignment for the upcoming scene.
Compile the output into a single execution package structured as follows: [SYSTEM CONTEXT] + [CHAPTER OBJECTIVE] + [12 GENERATED ROLE SUB-PROMPTS] + [PREVIOUS CHAPTER TEXT FOR CONTINUITY]
Using the compiled Master Prompt, generate the manuscript text for Scene [Y] of Chapter [X]. Prioritize extensive depth, organic pacing, and rich sensory grounding. Linger in the scene. Do not summarize or accelerate the conclusion.
Review the generated draft sequentially through these 4 lenses:
1. OPTIMIZER: Is the narrative logic flawless? Is the environment interacting correctly?
2. DRAFTER: Is the pacing sufficiently expansive? Did it skip over important emotional transitions?
3. CRITIQUE: Did the prose violate any anti-patterns from Prompt 0?
4. HUMANIZER: Do the characters sound real, or do they sound like an AI trying to sound poetic?
Fix all identified flaws and output only the final, polished prose.
When the engine executes the pipeline, it splits its cognitive focus into 4 specialized archetypes. Each role runs a different check on the draft.
| Role | Operational Focus | Primary Mandate |
|---|---|---|
| 1. The Optimizer | Structural Pacing & Flow | Ensures scenes do not feel compressed; regulates scene lengths and structural transitions. |
| 2. The Drafter | Sensory Grounding & Action | Inject physical, visceral details into the environment; prevents dry exposition. |
| 3. The Critique | Guardrail & Lore Compliance | Cross-references everything against the Series Bible; flags narrative or logic inconsistencies. |
| 4. The Humanizer | Emotional Depth & Dialogue | Strips away artificial or sanitized language; ensures character interactions feel raw and earned. |
This is the orchestration map. It tells the engine which 12 directives to pull for each chapter. You write this once. The engine consults it forever.
## PART IV: PER-CHAPTER DIRECTIVE ASSIGNMENTS
[Author maps out exactly which directives apply to which chapters]
- Chapter 1: Requires Directives [A, B, C] (Focus: World Introduction, Heavy Atmosphere, Character Introduction)
- Chapter 2: Requires Directives [D, E, F] (Focus: Inciting Incident, Pacing Acceleration, High Tension)
- Chapter 3: Requires Directives [A, G, H] (Focus: Lore Expansion, Internal Monologue, Regulating Pacing)
Five steps. That's the entire setup.
1. Setup Workspace. Create an isolated project directory or a distinct digital workspace (like a dedicated Notebook or project folder).
2. Ground the Engine. Upload PART I (The Series Bible) into your workspace as the absolute source of truth.
3. Initialize the Agent. Copy the Universal M3 Orchestration Engine Spark Skill (Part 1 above) into your AI agent configuration.
4. Select Chapter Directives. Look up the target chapter in PART IV, pull its specific assignments, and let the agent run its Meta-Step Translation to build your custom sub-prompts.
5. Activate the Loop. Execute the creation prompt, allowing the multi-role loop to automatically draft, self-critique, refine, and append the pristine text straight into your master manuscript file.
The engine does not just "write the scene." Every chapter is assigned exactly 12 highly specific orchestration directives pulled from a master library. The library is grouped into 4 categories. Pick the ones that match your book.
CM-001The Forbidden Lexicon: hard ban on tapestry, symphony, beacon, testament, delicate, intricate, visceral, dance, palpable, echoes.CM-002The Active Verb Mandate: every sentence hinges on a strong action verb. No "to be" verbs as the engine of a sentence.CM-003The Cadence Blueprint: alternate sentence lengths. Never allow more than two sentences of the same word count back-to-back.CM-004The Adverb Ban: no "-ly" adverbs. Force stronger verbs and concrete nouns.CM-005The Filter-Word Firewall: eliminate saw, felt, heard, noticed, realized, watched. Describe the object or event directly.CM-006Paragraph Density Limits: no paragraph over five sentences unless deliberate stream-of-consciousness. Force single-sentence paragraphs for impact.CM-007The Transition Eraser: no Suddenly, Meanwhile, As he walked, However. Imply transitions through action and spatial progression.CM-008The Subject-Opener Variance: vary sentence starters. No two consecutive sentences opening with the same pronoun, name, or "The."CM-009The Concrete Noun Directive: anchor all abstract concepts (fear, time, danger) to tangible physical nouns in the environment.CM-010The "As" and "With" Restrictor: ban "As [action], [action]" and "With a [adj] [noun]..." structures (primary AI markers).CM-011The Dialogue Tag Minimizer: "said"/"asked" no more than once per exchange. Use physical action beats.CM-012The Participle Phrase Cap: limit introductory participle phrases. Use chronological direct actions.CM-013The Punctuation Variance Rule: use em-dashes, semicolons, colons. Do not rely on commas and periods alone.CM-014The Exposition Blockade: never group backstory or world-building into a single paragraph. Drip context into active physical movements.CM-015The Metaphor Grounding Rule: no abstract metaphors. Comparisons must relate to everyday physical objects or experiences.CM-016The Negative Space Mandate: describe what is NOT happening, what is absent, to build atmosphere without cluttering with adjectives.CM-017The "To-Be" Eradication: search for every "there is" / "it was" and restructure to make the noun perform an action.CM-018The Climactic Shortening: as tension increases, average sentence length must mathematically decrease to accelerate pacing.CM-019The Specificity Demand: ban generic nouns (bird, tree, car, building) when a specific species, make, or style can be used.CM-020The Clause Limiter: no sentences with more than three clauses. Break complex thoughts into sharp distinct sentences.CM-M01The Zero-Flab Algorithmic Pass: every word must advance plot, reveal character, or establish setting. Absolute minimalism.CM-M02The Show-Don't-Tell Translation Matrix: emotions = physiological reactions. Time passage = changing environmental shadow/light source.CM-M03The Rhythm-Mapping Blueprint: musical rhythm map. Staccato for action, legato for introspection, caesuras for hard paragraph breaks.CM-M04The Rhetorical Device Injector: asyndeton for rushed action, polysyndeton for overwhelming scale, chiasmus for philosophical reflection.AO-001The Vector Embedding Compiler: generate a 512-dimensional semantic vector tag to index this scene for thematic retrieval.AO-002Context Window Defragmentation: review retrieved lore and strip all chronological data preceding the current chapter timeline.AO-003The "Dead Link" Subroutine: scan for any character/artifact not mentioned in the last 20 chapters. Force a background reference in the current scene.AO-004Automated Lexicon Ingestion: extract newly invented terminology, format as JSON key-value pair, push to master glossary.AO-005The Memory Compression Algorithm: compress 10,000-word subplot history into a 300-word summary string optimized for attention.AO-006Dynamic Temperature Scaling: map severity of cosmic anomaly → floating temperature variable (0.6 to 0.95) for AI hallucination level.AO-007The "Butterfly Effect" Tracker: map three unintended future consequences of the protagonist's decision. Store in hidden "Future Karma" table.AO-008Relational Graph Weighting: query DB for two characters in the scene, calculate "friction score" from past interactions, feed into prompt.AO-009The Motif Metronome: calculate symbol frequency. If it hasn't appeared in 15,000 words, force an insertion into background architecture.AO-010Automated Tagging Engine: generate 15 highly specific SEO/structural tags (#non-euclidean, #betrayal, #tech-decay) for the chapter.AO-011The Timeline Checksum: run a mathematical checksum on the scene's chronological movement against the master calendar.AO-012Lore Locking Protocol: identify the single most important rule of physics in this scene. Inject a system-level override preventing breakage.AO-013The Ambiguity Indexer: rate the scene's ending on an ambiguity scale of 1-10. Store so the next three chapters invert the rating.AO-014Cross-Chapter Echo Synthesizer: retrieve Ch 1's opening sentence. Force the current climax to reuse that exact syntactic structure.AO-015The "Off-Screen" Simulator: generate a simultaneous timeline of what the antagonist is doing at this exact timestamp.AO-016Token-Cost Predictor: estimate exact token output required. Set a hard stop-sequence if generation exceeds the narrative weight.AO-017The Perspective Lock: completely blind the model to the internal thoughts of any character other than the designated viewpoint.AO-018Environmental State Persistence: fetch weather and atmospheric decay from the previous chapter. Mandate seamless continuation.AO-019The Subplot Convergence API: identify the exact paragraph where two distinct subplots must physically intersect. Dictate the transition point.AO-020Hidden Variable Injection: pass a hidden variable ("Character A is secretly wounded") to imply through text without explicit reveal.AO-M01The Recursion Flag: end the scene with the exact same physical image it started with. Narrative loop.AO-M02Thematic State Machine: define chapter's state (Hopeful/Desperate/Nihilistic). Force every adjective through this state machine.AO-M03The Data-Loss Emulator: simulate corrupted memory. Provide incomplete background lore. Force the scene to write with authentic gaps.AO-M04Global Variable Mutation: if a cosmic event occurs, automatically output a script to update Global Gravity/Light across future retrievals.These are example directives for the "dark sci-fi / cosmic surrealism" genre. Substitute your own aesthetic constraints.
FG-001Terminal Entropy Protocol: world is in terminal, entropic decay. Every location must include decay, ruin, or failing energy.FG-002Non-Euclidean Geometry Constraint: straight lines curve back on themselves. Distances feel subjectively impossible.FG-003Eldritch Lore Injection: pull the top three esoteric lore tokens. Weave them into background action without over-explaining.FG-004The Void Focus: empty spaces — the void, the dark, the silence — are the primary antagonist. Characters react to absence, not presence.FG-005Techno-Organic Synthesis: blur machinery and biology. "Breathing metal," "calcified circuitry."FG-006Chronological Fracture: time behaves unreliably. Past bleeds into present. Disorient the viewpoint character.FG-007The Artifact's Resonance: if the central artifact is near, require sensory distortion — static, cold air, whispers only the protagonist hears.FG-008Scale Disorientation: oscillate wildly between macroscopic (dying stars) and microscopic (one dust mote, one drop of blood).FG-009Genre-Specific Terminology: genre-appropriate vocabulary. For sci-fi: bulkheads, containment fields, void-suits. For fantasy: ancient, ley-lines, warding. Adapt to project.FG-010The Apathy of the Cosmos: cosmic forces display absolute, crushing apathy. No personification. No human emotion in the cosmic elements.FG-011Sensory Deprivation Map: begin the scene by systematically stripping the character's senses one by one. Build tension before revealing horror.FG-012The Glitch Aesthetic: visual glitches, static, fragmented reality (corrupted VHS / failing CRT / broken bone-white light) in the physical environment.FG-013Subtextual Dread Anchor: identify the viewpoint character's core fear. Structure the environment to subtly reflect that fear without conscious realization.FG-014The Gravity Shift: alter the physical rules of gravity. Things float, fall upward, feel impossibly heavy.FG-015The Memory Bleed: environment physically changes in response to traumatic memories. Past manifests into physical space.FG-016Alien Architecture Filter: describe architecture using terms that imply it was built by things without human proportions or human minds.FG-017The Color of Decay: restrict palette. Rust, ash, bruised purple, sickly unnatural luminescence. (Adapt to your genre.)FG-018The Paradox Engine: force a logical paradox — something true and false simultaneously. Characters must accept it as a rule.FG-019The Static Transmission: communication broken by interference, static, fragments of dead languages.FG-020The Abyssal Gaze: feeling of being watched by something impossibly large and unseen. Environment holds its breath.FG-M01The Void Crawler: identify "blind spots" left unexplained in lore. Skirt the edges. Generate terrifying implications without revealing truth.FG-M02The Non-Linear Temporal Compiler: shatter the beat temporally. Write the end first, the beginning second, the middle last.FG-M03The Architectural Entropy Engine: mathematically degrade the architecture as the scene progresses. Rooms shrink. Corridors elongate indefinitely. Doors lead to the void.FG-M04The Entity Scale Relativizer: protagonist encounters a cosmic entity. Prohibit describing the entity's physical form. Describe only the effects on reality and psyche.AE-001The Roadmap Builder: map how this beat's resolution feeds the next three chapters. No dangling plot threads.AE-002Lore Schema Generator: take this unstructured world-building idea and format it into a relational DB schema. Define tags, category arrays, timeline dependencies.AE-003The "Iceberg Theory" World-Builder: generate the hidden 90% of this location's history. AI knows it. Reader sees only the top 10%.AE-004Subplot Weaver: identify a dormant subplot from 10 chapters ago. Force the engine to weave a subtle update into the background.AE-005Faction Motivation Mapping: define the conflicting goals of unseen factions. Show fallout without ever putting the factions on the page.AE-006Magic/Tech Limitation Setter: establish a hard unbreakable physical limitation for the tech in this beat. Define exactly how and when it fails.AE-007The Character Foil Architect: analyze protagonist's mindset. Introduce a minor NPC or environmental hazard as a perfect psychological foil to the protagonist's flaw.AE-008Scene-Sequel Rhythm Structurer: determine if this beat is a "Scene" (action/conflict) or "Sequel" (reaction/processing). Enforce the pacing mechanics of that phase.AE-009Thematic Thesis Statement: define the singular philosophical argument of this chapter. Ensure every dialogue and description subtly argues it.AE-010The MacGuffin Placement Protocol: generate a logical lore-compliant reason for a vital artifact to be in this location. Instruct on how it alters local gravity/atmosphere.AE-011Red Herring Generator: deliberately misdirect reader attention toward a false solution while hiding the true clue in plain sight.AE-012Ecosystem Logic Builder: define the biological/biomechanical food chain of this surreal environment.AE-013The "Sensory Evolution" Tracker: track how the protagonist's sensory perception of the world has shifted since the last chapter.AE-014The Climax Echo Chamber: identify the emotional climax of the last 5 chapters. Force the current scene to reverberate with that unresolved emotional frequency.AE-015The Worldbuilding Footnote Generator: extract every dropped name, location, or artifact. Generate an inline footnote string ready for worldbuilding tool integration.AE-016The "Hidden Mechanic" Reveal: identify one piece of system/mechanic lore that the protagonist knows but has not said aloud. Force the drafter to imply it.AE-017The Cultural Texture Injector: pull 3 cultural artifacts (food, music, ritual) for the scene's setting. Weave them into a single sensory beat.AE-018The "Previous Life" Cross-Reference: identify the scene's most important object. Search the lore DB for every prior appearance. Echo one detail.AE-019The "Reader Respite" Engine: track tension levels. If tension has been at maximum for 3 consecutive scenes, inject a single quiet sensory beat.AE-020The Sequel Pipeline: generate a list of 5 sub-prompts that the next chapter's drafter will use to pick up where this one left off.AE-M01The Multi-Chapter Arc Integrator: cross-reference 5 prior chapters. Force the drafter to mention at least 3 of their motifs in a single transitional sentence.AE-M02The "Escalation Ladder" Generator: define the exact emotional/physical escalation steps for this scene. Force the drafter to climb one rung per paragraph.AE-M03The "Thematic Inversion" Compiler: identify the chapter's primary thematic statement. Force the drafter to invert it in the closing paragraph.AE-M04The Lore Dependency Mapper: identify all lore entities the drafter must mention. Build a strict dependency tree: A must be mentioned before B can be referenced.This is the constitution the engine reads. You fill it out once with your project's lore. It becomes the absolute truth the engine cannot violate.
You cannot feed an AI a massive static lore document and expect dynamic character arcs. Split your Series Bible into distinct narrative phases. Each phase gets its own extremely detailed voice and thematic matrix. The engine receives a different Bible depending on where the story is.
Example structure (adapt to your project):
Phase I (Setup): grounded, hesitant, sensory. Rule: the lie of the world is a load-bearing lie.
Phase II (Descent): colder, more clinical. Rule: to fight the algorithm, you become a virus.
Phase III (Climax): algorithmic vocabulary takes over. Rule: the universe copies your code against you.
Phase IV (Aftermath): warm, biological, exhausted. Rule: healing requires the silencing of the divine.
By swapping the Bible per phase, the AI is structurally forced to execute character growth and stylistic decay. This is the single most important trick in the entire architecture. Most of the "AI voice stays the same" problem dissolves when you stop feeding the model one mega-Bible and start feeding it the right Bible for the right chapter.
Here is exactly how the engine executes a chapter on its own. The human never re-enters the loop.
1. Directive Retrieval & Assembly. The agent checks the orchestration map for the current chapter, pulls the 12 assigned tactical directives, and generates a scene-by-scene scaffold based on the active Phase Bible.
2. Drafting. The agent writes the first scene, explicitly instructed to linger on the senses, ignore word-count maximums, and apply the rules of the Drafter and Optimizer roles.
3. The Critique Pass. Instead of relying on its own memory, the agent cross-references the drafted scene against the Bible (its absolute truth), the anti-patterns, and the strict character voice rules.
4. The Rewrite. The agent takes the critique, bounces back into its drafting persona, and completely rewrites the scene to fix pacing errors, strip out AI-sterility, and lock in the lore.
5. Export & Iterate. The finalized, Humanized prose is appended to the master manuscript document. The agent then automatically moves to the next scene and repeats the loop until the chapter is done.
The result: the system drafts, self-corrects against a dedicated lore database, strips out the robotic tone, and outputs deeply thematic, 2,000+ word chapters entirely in the background.
If you are trying to write long-form fiction with LLMs, stop treating the model like a chatbox. Build a Constitution. Split your Bible into phases. Build an autonomous loop that holds the AI accountable to your voice.
Steal the architecture. Adapt the Bible. Build the loop. Write the book.