AI Agents I Built (2/7) — Story Director: Completing a 30-Chapter Novel

Agent architecture, context management strategies, and character consistency techniques required to complete a full-length novel

What This Post Covers

  • 4 structural components of an AI agent designed for long-form narrative completion (world-building, character sheets, outline, continuity tracking)
  • Hierarchical reference strategy for connecting 30 chapters within a limited context window (last 3 chapters verbatim + prior chapter summaries + foreshadowing tracker)
  • Dialogue-sample-based character sheet design and post-chapter consistency check loops
  • Strengths and limitations of AI in long-form narrative — which tasks to delegate to AI and which must remain with the author

Design Target: Why a Creative Agent Differs from a Technical Agent

A technical agent operates on well-defined inputs and outputs: read a file, modify code, run tests. A creative agent's objective — "complete a coherent long-form narrative" — is inherently ambiguous. That distinction produces four concrete design requirements.

  1. World-building document — Maintain the novel's world, technology rules, and organizational structures as a separate, persistent document. Fixed values that must not be regenerated each chapter are externalized to a reference store.
  2. Character sheets — Define each character's personality, backstory, relationships, and speech patterns. Sheets must include sample dialogue, not just attribute lists.
  3. Chapter outline — Design the full 30-chapter arc before writing any chapter. Unplanned generation causes plot diffusion around chapter 10.
  4. Continuity tracker — Maintain a dedicated document for events that occurred in prior chapters and unresolved foreshadowing threads.

Without these four pillars, defects accumulate past chapter 10: characters behave as different people, or already-resolved events reappear.

Case: "Ghost Network" Specifications

The following are the specifications of "Ghost Network," the full-length novel completed using this technique.

  • Length: 30 chapters, 258,000 characters (approximately 800 pages in Korean)
  • Genre: Cybersecurity / tech thriller
  • Premise: A security expert tracks down the true identity of the anonymous hacking collective "Ghost Network"
  • 5-act structure: Introduction (ch. 1–5) → Development (ch. 6–15) → Crisis (ch. 16–22) → Climax (ch. 23–27) → Resolution (ch. 28–30)

These specifications serve as the reference baseline for the techniques described in subsequent sections.

Context Management — Hierarchical Reference Strategy

An LLM's context window is finite. Injecting all 30 chapters at once is impossible; providing no prior context collapses narrative continuity. The combination that works in practice uses four layers.

  • Last 3 chapters verbatim — The three immediately preceding chapters are injected as raw text. Style, dialogue rhythm, and scene-level detail must be preserved for the next scene to connect naturally.
  • Summarized prior chapters — Chapters older than that are compressed to approximately 300-character summaries. Only plot progression and character state changes are retained; description is discarded.
  • Foreshadowing tracker — Unresolved foreshadowing threads, open questions, and deferred plot hooks are maintained as a separate list.
  • Character state snapshots — Each character's current location, emotional state, and known information are updated per chapter.

This structure trades "resolution of recent context" against "span of the full narrative" within a token budget. Limitations remain. Minor setup details from early chapters are observed to drop out in later chapters, requiring manual correction. The root cause is detail loss during the summarization step.

Character Consistency — Sheet Design and Post-Chapter Verification

AI does not "remember" a character's personality when writing each chapter — it "references" the sheet. That distinction produces subtle inconsistencies. Three countermeasures address this.

  • Dialogue-sample sheets — Include 3–5 sample lines in the sheet alongside personality traits. An example like "…could be, could not be. Need to check." conveys style more precisely than a descriptor like "cautious."
  • Post-chapter consistency check — After each chapter is written, run a review pass comparing each character's dialogue and actions against the sheet. Identify and correct utterances that deviate from the defined profile.
  • Immediate sheet updates — Significant character changes — growth, betrayal, relationship shifts — are reflected in the sheet immediately after the chapter is complete. The reference baseline for the next chapter must reflect the changed state.

The limitation is most visible in secondary characters. Protagonists accumulate dense sheets and stable speech patterns through high page-count exposure; secondary characters have sparse sheet entries and exhibit voice drift across chapters. This is not fully resolved. Mandatory dialogue samples for secondary character sheets is the identified improvement point.

Strengths and Limitations of AI in Long-Form Narrative

The outcome — a completed 258,000-character novel — is a data point demonstrating feasibility, not a ceiling on quality. AI performance splits clearly by task type.

Where AI performs well - Advancing chapters consistently along a defined plot structure - Technical setting detail (hacking techniques, network architecture, etc.) - Dialogue for characters with fully populated sheets

Where AI falls short - Natural use of metaphor and symbolism - Narrative rhythm designed to deliberately manipulate reader emotion - Nuanced expression of internal character transformation

In conclusion, reaching publication-grade quality requires a human editing layer. AI is effective at driving a draft to a completed state; literary depth is added in the editing phase. That division of labor is the realistic operating model.

Limitations and Corrective Directions

Three corrective points emerged from the design process. Each is directly applicable to any designer facing the same problem.

  • Outline first — Attempting chapter-by-chapter improvisation without a prior outline causes plot diffusion around chapter 10. Establishing the full structure before filling in chapters produces a stable result.
  • Pre-calculate context budget — Unconditionally injecting the last N chapters verbatim will hit the token ceiling. "Last 3 chapters verbatim + all prior chapters summarized" is the empirically validated balance point.
  • Maintain the completion structure — A temptation to close the story early appears around the midpoint (chapter 20). Holding to the planned 30-chapter structure yields better overall quality, because unresolved foreshadowing threads from the first half still require payoff in the remaining chapters.

Applicable Scope and Open Questions

This technique set is most effective for long-form narrative with complex world-building: cybersecurity thrillers, SF, fantasy, mystery — genres where state management of world, characters, and foreshadowing carries high weight. Conversely, short fiction, prose poetry, and essays — forms where micro-level sentence quality dominates — show limited benefit.

Two questions remain open.

  • How to recover detail lost during summary compression. Early-chapter setup dropping out in later chapters is a fundamental limitation of the summarization strategy. Whether a retrieval-augmented recall (RAG) layer can compensate requires further experimentation.
  • How far secondary character voice can be sheet-codified. Unlike protagonists, secondary characters appear infrequently, leaving their sheets thin. The minimum viable number of dialogue samples and the cost-benefit break-even point need to be quantified.

AI can function as a tool for driving a long-form narrative to completion. What determines the quality ceiling is the design of the sheets, structure, and editing loop.

Series overview: Series index

๋Œ“๊ธ€

๋Œ“๊ธ€ ์“ฐ๊ธฐ

์ด ๋ธ”๋กœ๊ทธ์˜ ์ธ๊ธฐ ๊ฒŒ์‹œ๋ฌผ

Agent Memory Engine (2/10) — Building an AI Agent Memory System with SQLite Alone

← 1/10 SQLite Memory Engine f… ๐Ÿ“š Series Index 3/10 memcore Completed → A memory engine that runs without daemons, without dependencies, anywhere ํ•ต์‹ฌ ์š”์•ฝ This post covers how to design an AI agent memory engine (memcore) on a single-file SQLite backend. - Structure: 25 modules, ~6,464 lines, 22 DB tables - Principles: no daemon, single-file portability, no required external dependencies, host-agnostic - Core techniques: 4-layer hybrid search (topic · vector · FTS5 · LIKE), U-tag dialectic-based confidence evolution, ingestion-layer bias rejection rules What This Post Covers One approach to building AI agent memory without a server or cloud vector DB. Specifically: (1) the limitations of text-file-based memory, (2) vector DB alternatives compared and selection criteria, (3) how to layer search tiers, (4) how to evolve confidence scores over time, and (5) how to block biased memory accumulation at the ingestion stage. Design Principles: Why Single-File SQLite Text-file-based ...
Read more »

"ML Foundations (9/9) — PyTorch vs TensorFlow, and the Road to Local LLMs"

์ด๋ฏธ์ง€
← 8/9 Deep Learning Architec… ๐Ÿ“š Series Index (series end) The final part. So far we've focused on models. Now we focus on the tools that actually run them . We'll lay out the philosophical difference between PyTorch and TensorFlow, and trace how Hugging Face transformers , llama.cpp , MLX , and Ollama built the bridge to running large language models on your own machine. By the end you should have the full mental model of "download a pretrained LLM and serve it locally." 0. Learning Objectives Explain the eager-vs-graph difference between PyTorch and TensorFlow. Explain, in graph terms, how autograd automates the backward pass. Use Hugging Face transformers ' AutoModel , AutoTokenizer , and pipeline abstractions. Describe llama.cpp's GGUF quantization, INT4 inference, and CPU-first flow. Describe Apple MLX's use of unified memory and how it differs from PyTorch. Run a local LLM with Ollama and call it through an OpenAI-compatible API. ...
Read more »

"RAG Core Study (14/26) — Evaluation Sets with RAGAS & DeepEval"

์ด๋ฏธ์ง€
Series overview: Series index ๐Ÿ“š Series Index Without an evaluation set, every RAG improvement is a story, not evidence. RAG systems fail in more than one place: retrieval, context selection, answer generation, and grounding. That means "looks better" is not a metric. Part 14 explains how to build a golden dataset , how RAGAS and DeepEval fit into the workflow, and why teams must separate retrieval quality from answer quality if they want tuning decisions to hold up. 0. Prerequisites Part 13 reranking — system quality now depends on multi-stage ranking. Part 1 grounding — a correct-looking answer is not enough. Part 12 Hybrid — multiple retrieval settings need comparison on fixed data. 1. Learning Objectives Build a small but useful golden dataset for RAG. Distinguish retrieval metrics from answer metrics . Use RAGAS and DeepEval in the right roles. Avoid the common traps in judge-model-based evaluation. 2. ํ•ต์‹ฌ ์š”์•ฝ An evaluation set for RAG is ...
Read more »

"ML Foundations (8/9) — Deep Learning Architectures: CNN, RNN, Attention"

์ด๋ฏธ์ง€
← 7/9 Deep Learning Training ๐Ÿ“š Series Index 9/9 PyTorch vs TensorFlow → What did we keep adding on top of an MLP? Vision went CNN, sequences went RNN/LSTM, and eventually both converged on Attention and the Transformer. Each architecture is easier to remember if you read it as what it gave up to gain something else . This part is the one-line summary of the decisive inflection points: LeNet → AlexNet → ResNet → LSTM → Attention → Transformer. 0. Learning Objectives Explain why CNN's convolution, pooling, and weight sharing match images so well. Trace what got unblocked from LeNet → AlexNet → VGG → ResNet as depth grew. State BPTT for RNNs and the vanishing/exploding gradient problem. Write the LSTM gate equations (forget, input, output) with intuition. Write the attention formula in query/key/value form and the scaled dot-product variant. State the precise reasons Transformers replaced RNNs (parallelism + long-range dependencies). 1. ํ•ต์‹ฌ ์š”์•ฝ CNN : weight sharin...
Read more »

"ML Foundations (7/9) — Deep Learning Training: Optimizers, Regularization, Initialization"

← 6/9 Neural Networks ๐Ÿ“š Series Index 8/9 Deep Learning Architec… → Building an MLP, as we did in Part 6, is not the same as getting it to train well . The right optimizer, regularization, initialization, and learning rate — when these line up, deep networks converge. When they don't, the network refuses to learn at all. This part is the catalog of those crafts, with formulas, paper citations, and code in one place. 0. Learning Objectives Compare and write the update rules for SGD, Momentum, Nesterov, and Adam. Explain how Dropout, BatchNorm, and LayerNorm work and where they belong in a model. Derive the variance formulas for Xavier and He initialization and match them to activations. Implement step, cosine, and warmup learning-rate schedules in PyTorch. Explain why gradient clipping is effectively required for RNNs and Transformers. Diagnose the most common training failures (NaN, plateau, overfitting) and apply first-line fixes. 1. ํ•ต์‹ฌ ์š”์•ฝ SGD : \(w \leftarro...
Read more »

OpenClaw to Hermes Migration (2/13) — What to Preserve, Partially Port, or Discard

← 1/13 Current Structure Inve… ๐Ÿ“š Series Index 3/13 What Moves to Hermes → A code review record classifying operational scripts into three tiers: preserve, partial-preserve, and discard. What This Post Covers An asset classification framework (preserve / partial-preserve / discard) for migrating a legacy agent system to a successor A concrete case applying the Strangler Fig + Subset Migration strategy to a memory pipeline Classification results evaluated across LOC, external dependencies, and replaceability axes A method for slimming down a bloated single file (1,310 LOC) using the migration as a natural entry point Problem Definition: Port Everything or Cut Deep? The OpenClaw inventory comprises multiple agents, multiple scheduler/daemon processes, and multiple operational scripts. When moving to the successor system, Hermes, the choice is not binary. A full port carries accumulated technical debt along with it; a minimal port discards hard-won domain logic. Code-revie...
Read more »

AI Agents I Built (5/7) — Building an Automated Blogger API Publishing System

← 4/7 My Life Organizer Agent ๐Ÿ“š Series Index 6/7 Dual → Designing a Markdown-Based Auto-Publish Pipeline with Blogger API v3 + OAuth 2.0 Key Summary Blogger API v3 + OAuth 2.0 enables a pipeline that converts Markdown files into publishable HTML posts blogger_publish.py handles the full flow — frontmatter parsing → body cleanup → HTML conversion → CSS injection → API post — in a single script For bulk publishing, quota management (delay, retry, TOC suppression) is the critical design concern Platform Selection Criteria API access is mandatory for AI agent-driven content publishing. A comparison of platforms: Platform API Cost AdSense Decision WordPress.com REST API Paid plan required Paid only Cost overhead Ghost REST API Self-hosted or paid Manual setup Ops overhead Blogger v3 API Free Native integration Adopted Blogger is free, exposes a REST API, and integrates natively with AdSense. For personal blog automation, this combination i...
Read more »