Skills for AI Agents (for memsa memsa) (Lucas B. Nicolosi Soares) (z-library.sk, 1lib.sk, z-lib.sk)

【One-Line Pitch】 A practical guide to “skills”—structured markdown files that tell AI agents how to execute real work—for developers, technical leads, and operators who want agents to do tasks reliably rather than just talk about them. 【Book Arc】 - **Opening (~0%–10%)**: Frames the core thesis through a personal case study—a course-update workflow that dropped from ~25 hours to under an hour once captured in a SKILL.md file—and poses the question of why this works now. - **Early (~10%–30%)**: Traces two parallel histories: humans writing down expertise (apprenticeships, SOPs, wikis) and engineers making machine instructions more human (machine code → assembly → high-level languages). The two tracks collide with ChatGPT and the rise of prompt engineering. - **Early–Middle (~30%–45%)**: Exposes prompting’s ceiling—models can describe work but not perform it—then introduces tool calling (Toolformer, OpenAI function calling) as the bridge that gives models “hands.” - **Middle (~45%–55%)**: Walks through the five-step tool-calling loop (describe tools, receive question, model requests tool, code executes, result returns) and the shift to non-deterministic models orchestrating deterministic systems; touches on structured outputs and tool-description quality. - **Late (beyond excerpt coverage)**: The table of contents points to skill design for the context window, building skills, composition and advanced patterns, business and domain-specific skills, plus appendices (SKILL.md schema, templates, troubleshooting)—but the excerpts do not cover this material. - **Ending (beyond excerpt coverage)**: Applications and reference appendices are listed as unavailable in the sample; the guide cannot summarize their content. 【Key Takeaways】 - **Skills turn documents from descriptions into executable instructions** (Opening): The central claim is that a markdown file like SKILL.md can now be read and acted on by an agent, collapsing the gap between knowing and doing—the book’s organizing idea. - **Expertise transfer has always been the bottleneck** (Early): Apprenticeships, manuals, and SOPs all assumed a human reader who interprets and executes; agents make that assumption optional, which changes what it means to write something down. - **Prompting has a hard ceiling** (Early–Middle): Better prompts improve output quality but cannot issue refunds, run queries, or touch external systems—the model remains “a brain in a jar.” - **Tool calling is the structural turning point** (Middle): The five-step loop lets a non-deterministic model choose actions while deterministic code executes them—a new software contract of judgment plus infrastructure. - **Tool descriptions are the new prompts** (Middle): When models misbehave, the fix shifts from crafting prompts to crafting clear tool schemas and descriptions—an early signal of the book’s skill-design philosophy. - **Structured outputs matter alongside tool calling** (Middle): Predictable data formats make agent pipelines reliable; the excerpts note this capability matured in tandem with function calling. - **The book is practice-oriented, not theoretical** (Opening): Real-world examples and hands-on structuring of tasks are promised throughout, aimed at reengineering AI applications around business and personal requirements. 【Reading Tips】 - Read Chapters 1–2 closely: they carry the conceptual load (why skills exist, where they sit in the tooling stack). The historical sections are skimmable if you already know the prompting-to-tool-calling timeline. - Treat the five-step tool-calling loop as the mental model to internalize—everything later about skill design builds on the model-decides/code-executes split. - Since this is an Early Release, expect rough edges and unavailable chapters; check the O’Reilly page or GitHub repo (mentioned as forthcoming) for updates before relying on the appendices. - When the full book is available, go straight to the SKILL.md schema and templates (Appendices A–B) after Chapter 2—they are the most directly actionable artifacts. - Keep a concrete workflow of your own in mind while reading; the book’s value comes from translating your tacit process into a skill file. 【Coverage Limits】 This guide is based on stratified excerpts covering roughly the first half of the book (Chapters 1–2 and part of the tool-calling discussion); later chapters on building skills, composition patterns, business/domain applications, and all appendices are listed but not covered in the sample.