Google Prompt Engineering (Lee Boonstra)（Z-Library）

【One-Line Pitch】 A practical, vendor-neutral guide to writing effective prompts for large language models, covering configuration knobs, core techniques, and code-specific prompting—ideal for developers, product managers, and curious non-engineers who want better AI outputs without a data science background. 【Book Arc】 - **Opening (~0%–10%)**: Introduces prompt engineering as an accessible skill for everyone, then dives into LLM output configuration—length, temperature, top-K, and top-P—explaining how these sampling controls shape response variability and determinism. - **Early (~10%–30%)**: Covers foundational prompting techniques: zero-shot, one-shot, and few-shot prompting, plus system, role, and contextual prompting to steer model behavior and persona. - **Middle (~30%–60%)**: Advances to reasoning techniques—step-back prompting, Chain of Thought (CoT), self-consistency, Tree of Thoughts (ToT), and ReAct—showing how to elicit multi-step logic and improve accuracy on complex tasks. - **Late (~60%–80%)**: Focuses on code-specific prompting: writing, explaining, translating, debugging, and reviewing code, plus an introduction to multimodal prompting with images and text. - **Ending (~80%–100%)**: Consolidates best practices—providing examples, simplicity, output specificity, instruction-over-constraint, token control, and variable use—then covers JSON repair, schema handling, collaborative experimentation, and documenting prompt attempts for reproducibility. 【Key Takeaways】 - **Output configuration is the first lever** (Early): Temperature, top-K, and top-P control randomness; lower values for deterministic tasks, higher for creative ones—mastering these prevents vague or repetitive outputs. - **Few-shot prompting beats zero-shot for consistency** (Early): Providing 1–5 examples in the prompt anchors the model’s format and tone, especially for classification or structured output tasks. - **System and role prompting shape behavior** (Early): Setting a system prompt or assigning a persona (e.g., “you are a senior Python reviewer”) dramatically changes response quality without extra examples. - **Chain of Thought unlocks reasoning** (Middle): Asking the model to “think step by step” improves accuracy on arithmetic, logic, and multi-hop questions; self-consistency (sampling multiple times and voting) further boosts reliability. - **ReAct combines reasoning with action** (Middle): For tasks needing external tools or search, ReAct prompts interleave thought, action, and observation loops—critical for agentic workflows. - **Code prompting requires explicit context** (Late): For debugging or review, include the error message, expected vs. actual behavior, and language version; for translation, specify target syntax and style constraints. - **Instructions over constraints** (Ending): Positive directives (“do X”) outperform negative prohibitions (“don’t do Y”)—models follow clear actions more reliably than vague restrictions. - **Iterate and document** (Ending): Prompt engineering is iterative; track versions, outputs, and failures to build a reusable prompt library, and use JSON repair or schema validation to handle structured outputs. 【Reading Tips】 - **Skim the configuration chapter** (Early) if you’re already familiar with LLM APIs—just note the temperature/top-P trade-off; deep-read the technique chapters (CoT, ToT, ReAct) for reasoning-heavy use cases. - **Focus on the code prompting section** (Late) if you’re a developer—the debugging and review prompts are immediately actionable; skip multimodal if you only work with text. - **Treat the best-practices list** (Ending) as a checklist, not theory—apply “provide examples” and “control max token length” to your next prompt and compare outputs. - **Hard spot**: Tree of Thoughts and ReAct may feel abstract—re-read the examples twice, then try a simple puzzle (e.g., planning a trip) to internalize the loop. - **Takeaway to remember**: The book’s core message is that prompt engineering is a skill, not a talent—anyone can learn it, but iteration and documentation separate amateurs from pros. 【Coverage Limits】 This guide synthesizes the table of contents and introductory/concluding excerpts; it does not cover detailed worked examples, specific model comparisons, or advanced multimodal case studies that may appear in the full text.