Idea Palette Pipeline

## Overview Writer's block is not a lack of ideas — it is a lack of *structured* ideas that fit the current work's constraints. hakadoru.ai's Idea Palette addresses this by generating creative suggestions through a systematic pipeline rather than open-ended brainstorming. **The Idea Palette is defined as a nine-stage pipeline (P0-P8) that combines six narrative facets — world-building, story engine, tone, relationship dynamics, and two reserved axes — with a predicate system to generate structurally diverse creative ideas while respecting brand-specific constraints.** The pipeline produces ideas that are contextually grounded in the author's existing work, genre-appropriate for the brand, and structurally varied to avoid the homogenization that plagues unconstrained LLM brainstorming. ## Background When asked to "suggest plot ideas," LLMs tend to produce variations on the same themes — a phenomenon known as mode collapse in generation. The suggestions cluster around popular tropes, lack structural diversity, and ignore the specific constraints of the author's work (established characters, world rules, genre conventions). Existing AI writing tools typically offer free-form brainstorming (ChatGPT), single-axis plot suggestions (Sudowrite's "Story Engine"), or template-based prompts (NovelAI's scenarios). None provide a multi-axis, constraint-aware idea generation system that guarantees structural diversity in its output. ## The Nine Stages ### P0: Normalization The author's request and attached context (knowledge fragments, recent scenes) are normalized into a structured input format. Ambiguous requests are disambiguated, and the scope of idea generation is bounded. ### P1: Exploration The normalized input is expanded across six narrative facets (detailed below), generating a broad space of potential directions. This stage intentionally over-generates to ensure the subsequent filtering stages have sufficient material. ### P2: Filtering Generated directions are filtered against brand constraints (content policy, age rating), author-specified exclusions, and predicate rules. Directions that violate MUST-NOT rules are eliminated; those satisfying MUST rules are prioritized. ### P3: Incident Generation Surviving directions are developed into concrete narrative incidents — specific events that could occur in the story. Each incident is grounded in the author's existing characters and world, not generic templates. ### P4: Combination Incidents from different facets are combined to create compound ideas. A world-building incident combined with a relationship dynamic creates a richer suggestion than either alone. The combination logic avoids pairing incompatible incidents. ### P5: Scoring Combined ideas are scored on novelty (divergence from the author's recent scenes), coherence (compatibility with established narrative), and diversity (distance from other generated ideas in the batch). Low-scoring ideas are pruned. ### P6: Refinement Top-scoring ideas are refined with additional detail: potential scene openings, character reactions, thematic implications. This stage transforms abstract concepts into actionable writing prompts. ### P7: Formatting Refined ideas are formatted for presentation with consistent structure: title, one-paragraph summary, key characters involved, narrative facets utilized, and estimated scene count. ### P8: Presentation The final set of ideas (typically 5-8 per generation) is presented to the author with sorting and filtering options. The author can select ideas for further development, combine multiple suggestions, or regenerate with adjusted parameters. ## Six Narrative Facets The pipeline generates ideas across six defined axes: - **Facet A (World)** — Setting-driven ideas: environmental changes, political events, natural phenomena, technological developments - **Facet B (Story Engine)** — Plot-driven ideas: conflicts, mysteries, goals, obstacles, reversals - **Facet C (Tone)** — Atmosphere-driven ideas: tonal shifts, genre blending, pacing changes, mood transitions - **Facet D (Relationship)** — Character-driven ideas: relationship developments, alliances, betrayals, revelations - **Facets E and F** — Reserved for future expansion (planned for R18-specific axes in later phases) Each facet operates independently during exploration (P1) and is combined during the combination stage (P4), ensuring that generated ideas span multiple narrative dimensions. ## Predicate System The predicate system governs what types of ideas are generated for each brand. Predicates are boolean rules organized into two sets: - **30 shared predicates** — Apply across all brands (e.g., "ideas must reference at least one existing character," "ideas must not contradict established timeline") - **18 BL-specific predicates** — Apply only to BL brands (e.g., "relationship dynamics must involve the established pairing," "ideas should explore emotional intimacy progression") Predicates carry MUST, SHOULD, or MUST-NOT strength levels. MUST predicates are hard constraints (violations cause idea elimination at P2). SHOULD predicates influence scoring at P5. MUST-NOT predicates trigger immediate filtering. ## Diversity Control A critical challenge in idea generation is homogenization — the tendency of LLMs to generate variations of the same underlying idea. The Idea Palette implements three diversity mechanisms: 1. **Facet coverage requirements** — Each generation batch must include ideas from at least 4 of the 6 facets 2. **Embedding-based deduplication** — Ideas that are semantically too similar (measured by embedding distance) are pruned, keeping only the highest-scored variant 3. **Recency penalty** — Ideas that resemble the author's recently written scenes receive a scoring penalty, pushing the pipeline toward genuinely novel suggestions These mechanisms work together to prevent the common LLM failure mode where "generate 5 ideas" produces 5 variations of the same idea. By enforcing multi-facet coverage, measuring semantic distance, and penalizing similarity to recent work, the pipeline guarantees that its output set spans meaningfully different creative directions. ## Brand-Specific Constraints The pipeline adapts its behavior based on the active brand's configuration: - **BL brands** activate the 18 BL-specific predicates, ensuring that generated ideas respect established romantic pairings and explore emotional dynamics appropriate to the genre - **Fanworks brands** apply additional predicates that check ideas against the author's declared source material, avoiding suggestions that contradict established canon - **R18 brands** will gain access to Facets E and F in later phases, enabling intimate scene ideation through the same structured pipeline - **All-ages brands** apply MUST-NOT predicates that filter any ideas involving content beyond the age rating threshold ## Comparison with Other Approaches Sudowrite's brainstorming features generate suggestions along a single axis (plot advancement) without multi-facet exploration or predicate constraints. NovelAI's scenario system provides templates rather than contextually-grounded ideas. ChatGPT brainstorming sessions lack structural guarantees about diversity and brand compliance. The Idea Palette's nine-stage pipeline with six facets and a predicate system represents a more engineered approach to creative idea generation — treating it as a constrained optimization problem rather than an open-ended prompt. ## Conclusion The Idea Palette Pipeline transforms creative brainstorming from an unconstrained, homogenization-prone process into a structured, multi-faceted generation system. By decomposing idea generation into nine stages, exploring six narrative facets, and enforcing constraints through a predicate system, hakadoru.ai produces diverse, contextually grounded, and brand-appropriate creative suggestions. The pipeline ensures that the ideas presented to authors span different narrative dimensions — world, plot, tone, and relationships — rather than clustering around a single axis of variation.