Vibe Engineering (MEAP V02) (Tomasz Lelek, Artur Skowroński)（Z-Library）

【One-Line Pitch】 A practical field guide to turning AI-assisted coding into a disciplined, end-to-end engineering process—covering the tradeoffs, guardrails, and failure modes that appear once generated code has to survive production. Best suited to working developers, tech leads, and architects who already use LLM coding tools and now need to govern them. 【Book Arc】 - **Opening (~0%–10%)**: Frames the core problem—vibe coding produces brittle, opaque code optimized for functional correctness rather than performance, security, or maintainability—and introduces vibe engineering as the "architectural plan" wrapped around a probabilistic core. (Opening) - **Early (~10%–30%)**: Builds the mental model: the end of "scale worship," the 70% Problem, edge-case cartography, and the shift from tacit craft knowledge to explicit, verifiable specifications and validation pipelines. (Early) - **Middle (~30%–50%)**: Walks through a concrete legacy modernization—migrating a JBoss/JSF/JPA application to Spring Boot and MongoDB—using an agentic IDE, with testing and incremental commits as the backbone. (Middle) - **Late (~50%–80%)**: Moves into context engineering for agents, LLM-driven data analytics and visualization, continuous AI development, and a scientific approach to validating LLM-based solutions. (Late) - **Ending (~80%–100%)**: Closes on performance engineering when assumptions mislead, ultra-tight evaluation loops for AI-powered codebases, and FinOps for LLMs—cost-cutting, confidentiality, and chip selection. (Ending) 【Key Takeaways】 - **Vibe coding is the sketch; vibe engineering is the architectural plan** (Opening): the deliberate integration of AI into a mature software lifecycle, not a replacement for it. - **LLMs are probabilistic, non-deterministic systems with no built-in model of truth** (Opening): every design choice trades accuracy, creativity, safety, cost, and latency against one another—so a provider-agnostic methodology matters more than picking the "best" model. - **The 70% Problem defines the practical limit** (Early): AI accelerates scaffolding and boilerplate but falters on the context-sensitive final 30%—edge cases, error conditions, and boundary behavior that actually define production readiness. - **The real bottleneck is procedural, not model quality** (Early): teams apply human-diff review practices to agent-produced, multi-module changes, inflating review times and erasing generation speed gains—the AI productivity paradox. - **Prompts and specifications must become versioned, executable artifacts** (Early): typed APIs, property tests, Agent Manifests, sandboxed runs, canary releases, and auditable provenance turn ad-hoc review into an engineered pipeline. - **Legacy modernization is a repeatable, AI-assisted framework** (Middle): migrate layer by layer with one commit per step, generate missing tests first, and avoid overusing LLMs for tasks static tools handle better. - **Evaluation is king** (Late): an ultra-tight engineering loop for AI-powered codebases is what separates repeatable results from alchemy. - **Cost, confidentiality, and hardware are engineering concerns** (Ending): FinOps for LLMs treats token spend, data leakage, and chip choice as first-class tradeoffs, not afterthoughts. 【Reading Tips】 - Read Chapter 1 closely—it establishes the mental model and vocabulary (70% Problem, edge-case cartography, probabilistic core) that every later chapter assumes. - Treat the legacy modernization walkthrough (Chapter 2) as a template: follow the prompt-and-commit rhythm, then reapply it to your own stack rather than copying the specific Spring/Mongo choices. - Skim the code-heavy diffs; the excerpts note that full AI responses are truncated to key changes with commit links—focus on the reasoning, not the listings. - Deep-read the validation, evaluation, and FinOps chapters if you own production systems; these are where the cost and risk tradeoffs actually bite. - Keep a running list of tradeoffs the authors flag (increment size, context carryover, model temperature vs. retrieval) and map them to decisions in your own pipeline. 【Coverage Limits】 This guide is based on stratified excerpts covering roughly the first half of the book plus the table of contents; chapters on context engineering, LLM-driven analytics, continuous AI development, scientific validation, performance engineering, evaluation, and FinOps are represented only by their titles and brief framing, so their specific arguments and examples are not summarized here.