From Paper to Prompt: Treating Books as Source Code for Agents

The Skill Distillation Wave

The AI industry is currently fixated on packaging expertise into portable units. Researchers at Shanghai AI Lab recently published COLLEAGUE.SKILL, a framework that converts heterogeneous human traces into versioned skill packages with separate capability and behavior tracks; the project has attracted roughly 18.5k GitHub stars and a gallery of 215 skills from 165 contributors. Tom Tunguz, writing about his personal Pi agent, describes a nightly loop where frontier models author atomic SKILL.md playbooks that smaller local models execute, transferring procedural knowledge through markdown rather than weights. Marketplaces such as the Agent Skills Library now catalog thousands of skills for Claude, Codex, and other coding agents, treating instructions as composable infrastructure. Against this backdrop, the open-source project Cangjie Skill has chosen a narrower and more difficult raw material: the printed book.

The Anti-Summary Methodology

The central grievance behind Cangjie Skill is that reading and summarizing are human-centric activities. A summary helps a person remember; it does not help an agent act. When users paste an entire book into a context window, the model drowns in narrative, misses the underlying logic, and produces generic advice. The project’s answer is RIA-TV++, a six-stage pipeline that treats a book as unrefined ore and refuses to let most of it through.

The first stage, Adler analysis, is a deliberate throwback to Mortimer Adler’s analytical reading method, a pre-digital technique for structural deconstruction. It forces the distiller to answer four questions before any extraction begins: what is the book about as a whole, what is being said in detail, is it true, and what of it. This produces BOOK_OVERVIEW.md, a contract that prevents later stages from cherry-picking quotes out of context.

Only then do five parallel extractors begin their work. One hunts for frameworks, another for principles, a third for cases, a fourth for counter-cases, and a fifth for terminology. This parallelism is designed to prevent the single-lens distortion that plagues most summarization; a framework extractor might miss a critical counter-example that the counter-case extractor catches. The resulting candidate pool is large, messy, and overlapping.

Triple Verification then applies a brutal filter. A candidate must show cross-domain evidence—two independent sections of the book must point to the same insight, ensuring it is not a throwaway line. It must demonstrate predictive power by answering a question the author never explicitly posed, proving the insight travels beyond its original context. And it must survive a uniqueness test: if the insight is common sense, it is discarded. The documentation notes that typically only 25 to 50 percent of candidates survive. This is the opposite of retrieval-augmented generation, which retrieves everything that might be relevant; Cangjie Skill argues that most of a book is padding, and the agent should only receive the non-obvious, structurally validated core.

The RIA++ stage translates these survivors into an agent-native grammar. R (Reading) preserves the original anchor. I (Interpretation) rewrites it in operational terms. A1 and A2 attach historical and future trigger scenarios. E (Execution) provides explicit steps. B (Boundary) defines where the skill breaks. Finally, Zettelkasten linking weaves the approved skills into a dependency graph, so that an agent invoking a skill on, say, competitive moats knows to also consult the companion skill on cash-flow analysis. The result is a knowledge base that behaves like a software library, not a document.

The Factory Floor

Cangjie Skill has already processed a startlingly eclectic canon. Its generated packs include twenty investment heuristics from Buffett’s shareholder letters, twelve mental models from Poor Charlie’s Almanack, fifteen viral copywriting triggers, twelve sales-copy skills from The Copywriter’s Handbook, fifteen STEPPS propagation strategies from Contagious, thirteen trust-building methods from 1000 True Fans, ten organizational design skills from No Rules Rules, and fifteen cognitive tools from Cognitive Dividend. Even the dense, classical language of the Huangdi Neijing yielded twenty-two diagnostic and methodological skills, while the Mao Selected Works produced twenty-five skills covering cognition, strategy, organization, and execution.

This range is the point. Warren Buffett’s letters and the Yellow Emperor’s Classic of Internal Medicine share no vocabulary, no culture, and no epistemology. Forcing both through the same distillation grammar tests whether the pipeline is domain-agnostic or merely a fancy summarizer. The consistency of the output—every skill tagged with trigger conditions, execution steps, and failure boundaries—suggests the methodology is closer to a packaging standard than a literary critique.

People, Books, and Evolution

The project does not exist in isolation. Its creators describe a trilogy: Nuwa Skill distills people, capturing thinking styles and interaction DNA; Cangjie Skill handles books, capturing systematic methodologies; and Darwin Skill evolves both. The analogy is apt. Person-grounded skills carry tone and behavioral patterns, as explored in the COLLEAGUE.SKILL research, which splits packages into capability and behavior tracks. Book-grounded skills carry frameworks and decision trees. One is persona; the other is procedure. Cangjie Skill bets that procedure is more durable, more composable, and easier to validate under pressure.

The Rough Edges

The repository is candid about its limitations. The 25–50 percent survival rate means a great deal of potentially useful content is discarded, and the methodology demands a high-quality source text to begin with. It is also unclear how much of the pipeline is automated versus human-curated; the repository provides templates and extractor prompts, yet the generated skill packs feel closer to artisanal outputs than assembly-line products.

More broadly, the skill distillation field faces unanswered questions about validation and drift. Discussion around the COLLEAGUE.SKILL paper has flagged the lack of concrete protocols for auditing patches, provenance, or cross-host consistency once skills begin to evolve. Cangjie Skill’s pressure-test with bait questions is a clever defensive measure, but it is not a formal verification framework. And because these skills are ultimately markdown instructions interpreted by an LLM, their correctness depends on the agent’s ability to follow them, not on any guaranteed execution semantics.

The Outlook

Cangjie Skill is a bet against context-window maximalism. Rather than stuffing an entire book into a prompt and hoping the model retrieves the right passage, it argues for pre-extraction, pre-validation, and pre-linking of knowledge. An analysis of the project notes that this approach achieves “zero redundancy,” allowing an AI to possess high-density professional knowledge while consuming minimal context window. As agent hosts begin to meter input tokens aggressively, such density becomes an economic necessity, not merely a convenience.

In a landscape where raw retrieval is cheap but noisy, and long-context inference is expensive, pre-distilled skill packs may become the preferred unit of exchange between humans and agents. The same analysis frames this as a foundation for autonomous agents whose strength will lie not in how many books they have “read,” but in how many validated cognitive tools they can invoke in real time. Whether RIA-TV++ becomes an industry standard or remains a boutique craft depends on whether the community can automate the pipeline without diluting the verification rigor. For now, it stands as one of the most systematic attempts to turn a library into a codebase.