AI cookbook generator produces 53 books from hundreds of millions of tokens

The Robot Book Club project began with three hard requirements: generation input should fit on a book cover, output should be good enough to justify printing (no "slop"), and the entire process should be end-to-end automated. Over several months, the system generated 53 cookbooks through more than 500 attempts, consuming hundreds of millions of tokens. The books are available free on the web and as PDF and EPUB downloads, though the developer acknowledges they are not yet worth printing; the recipes and curation, however, are strong enough to stand by publicly.

The system is organized into three cookbook series, each defined by a program that takes simple inputs—a book title, back-cover description, and author voice notes—and produces a finished book. The Diaspora Series features recipes from second-generation American households (Japan, Cuba, Spain), the At Home Series adapts restaurant cuisines for home cooking (Jewish Deli, Chinese Takeout), and the Diet Series targets specific eating plans (Whole 30 Quick Meals). The developer hand-wrote the first program for the Diaspora Series through trial and error, then extracted an agent skill from it to power all subsequent series. The system now skims random book PDFs after generation, feeds change requests back to the agent to update the series program, and editions are tracked on the website—for example, Taiwanese is on Edition 13 (though the real count is closer to 200 because it was hand-written), while Vietnamese is only on Edition 3 since it used the most recent program version.

A core design decision was to keep the top-level control plane rigid—a workflow of agents rather than a single autonomous agent. The developer resists yielding overall control to an LLM, preferring to maintain a specific research, writing, editing, typesetting, and publishing process, while allowing many agents to operate within that framework. This rigidity supports better error handling, recovery, logging, and observability, which has been critical to scaling output volume. Within that structure, workflows—encapsulations of common LLM processing patterns that can be invoked repeatedly—have been the biggest factor in scaling quality. One workflow handles image generation, addressing the challenge that many regional foods are poorly represented in training data; another reduces repeated tropes across chapter introductions through editorial passes. The developer can reference workflows by name when asking an agent to write a new series program, meaning Version 1 of a new program immediately benefits from tricks developed across 30 versions of a prior series.

Caching emerged as essential for progress. Each book generation takes many hours, and errors in final stages (especially PDF layout) used to destroy entire runs. Without caching, the developer could iterate once per day on a side project, or three times per day as a full job. With caching, regenerating a book takes about ten minutes, dramatically increasing the number of iterations possible in a session. The trade-off is an append-style development approach favoring new data over co-generation to avoid cache invalidation. Output stability across editions proved difficult to manage because thousands of autonomous decisions per generation mean that minor prompt or workflow changes cascade into enormous downstream effects—entire cookbook chapters have been added or destroyed by small adjustments to author tone. The developer has adopted strategies like clever cache-key management to pin particular generations across editions despite changes to the true cache key.

Good sandboxing proved essential to quality improvement. Each book generation happens within a custom sandbox containing logs, LLM input/output snapshots, and a versioned filesystem, allowing the developer to rewind to before or after a particular agent fired and inspect exact state. The entire generation workflow involves thousands of nodes per book, accumulating nearly a quarter terabyte of logging data, invaluable for post-mortem analysis. While visualizing the generation process remains unsolved, the developer found that noting a flaw and providing those notes to a coding agent along with sandbox folder references often yields debugging threads within minutes.

The improvement loop is notably non-recursive. Rather than waiting for self-improving recursion, the developer automated three skills that run after every generation: an output critique skill identifies issues in performance, resilience, and quality; a bug writing skill curates critiques into deduplicated bug tickets; and a bug triage skill suggests the top five bugs to fix. This automation has replaced hand-reading logs and hand-coding updates, automating about 95% of the development process. The developer now simply reviews the top five improvement suggestions every once in a while, adds context or constraints, and lets an agent work on them.

In closing, the developer argues that frontier models are already capable of generating the next great novel, but they need the right harness. LLMs remain poor at long-form content when given only a prompt, but Robot Book Club demonstrates that the right harness can yield at least an order of magnitude improvement in output quality, possibly two. A good writer, prolific reader, and agent whisperer could build a viable fiction-writing harness in about a year of dedicated full-time work with today's models—not for every kind of book, but for specific categories like "Emerging Technology Thriller" or "Coming of Age Romance." The developer believes that fiction generated under such conditions would be worth reading.

The project reflects a broader insight about how large language models perform best under constraints. The developer chose cookbooks specifically because they minimize the long-form coherence challenges that trip up LLMs—recipes are largely episodic with few cross-dependencies, and cookbooks are fundamentally curation and translation of existing cultural knowledge rather than pure invention. This strategic choice reveals that LLM output quality is not solely a function of model capability but of task design.

The technical infrastructure supporting Robot Book Club—workflows, caching, versioned sandboxes, and automated critique—addresses the operational challenges of iterative AI-driven content generation. Caching proved essential because early errors late in the generation pipeline (particularly in PDF layout) would destroy an entire run; without caching, the developer could only attempt one or two iterations per day. The sandboxing system, which logs thousands of nodes per book across nearly a quarter terabyte of data, enables post-mortem debugging and has become the primary tool for quality improvement.

The developer's conclusion about fiction writing is notably measured: frontier models may already be capable of generating compelling novels, but only if embedded in a purpose-built harness that maintains balance and quality across a long generation process. The claim is not that models are ready today, but that a skilled team could build such a harness in about a year of full-time work for specific fiction categories. This suggests the frontier of AI-generated long-form content may be less about model capability and more about the engineering scaffolding around it.