The AI Engineering Playbook
138 copy-paste prompts for teams shipping LLM features — evaluation and benchmarking, prompt test suites, RAG and chunking, agent workflows, model cards, governance and conversational design.
Key Capabilities & Features:
- •The System Primer: paste your use case, model, data shape and failure tolerance once — every prompt after it reasons about your actual risk profile
- •Evaluation prompts that build a test set from your real failures rather than from synthetic examples that pass
- •A retrieval-versus-generation diagnostic, since RAG failures are routinely fixed in the wrong layer
- •Explicit guidance on when an agent, a chatbot or a model is the wrong architecture for the problem
- •138 prompts across 9 sections — evaluation & benchmarking, prompt test suites, rag & retrieval, agent workflows and more
- •Works with ChatGPT (GPT-5), Claude Opus 5 and Gemini 3.7 Flash — fill the bracket, paste, done
One-time payment • Lifetime access & future updates
What's Inside
9 modules138 promptsEvery module below is included in the single purchase.
- 1
Evaluation & Benchmarking
22 promptsDefining what good output means before building, building an eval set from real failures, LLM-as-judge and its limits, regression testing, and metrics that are not vibes.
Replaces a system previously sold separately
- 2
Prompt Test Suites
18 promptsVersioning prompts like code, test cases covering edge and adversarial inputs, detecting regressions after a model change, and A/B testing prompts honestly.
Replaces a system previously sold separately
- 3
RAG & Retrieval
20 promptsChunking strategy against your document shape, embedding and reranking choices, retrieval evaluation separate from generation, and diagnosing whether retrieval or generation failed.
Replaces a system previously sold separately
- 4
Agent Workflows
18 promptsDecomposing a task into steps a model can reliably complete, tool definitions, failure and retry handling, loop and cost guardrails, and when an agent is the wrong architecture.
Replaces a system previously sold separately
- 5
Conversational Design
16 promptsFlows that recover from misunderstanding, disambiguation, setting capability expectations early, handoff to a human, and the failure message that does not blame the user.
Replaces a system previously sold separately
- 6
Voice & Multimodal
12 promptsDesigning for speech where there is no screen to scan, confirmation without tedium, error recovery in audio, and what genuinely suits voice rather than a form.
Replaces a system previously sold separately
- 7
Model Cards & Documentation
14 promptsDocumenting intended use and known limitations, training and evaluation data provenance, performance across subgroups, and the failure modes users must be told about.
Replaces a system previously sold separately
- 8
Governance & Risk
12 promptsAcceptable use policies, human-in-the-loop decisions, logging and auditability, incident response for a bad model output, and where a model must not be the decider.
Replaces a system previously sold separately
- 9
Knowledge Structure
6 promptsOntology and entity modelling for retrieval, structuring a knowledge base a model can use, metadata that improves retrieval, and keeping a corpus current.
Replaces a system previously sold separately
Test Workflow Execution in Your Browser
Verify actual prompt structure, token speed, and reasoning depth before downloading.
Frequently Bought Together
Technical Specifications
Architecture & System Overview
Most LLM features do not fail loudly. They work in the demo, pass the handful of examples someone tried, and then produce quietly wrong output for months because nobody built a way to notice.
The AI Engineering Playbook is a 138-prompt playbook covering evaluation and benchmarking, prompt test suites, RAG and retrieval, agent workflows, conversational design, voice and multimodal, model cards and documentation, governance and risk, and knowledge structure.
Start with the System Primer — paste in your use case, your model, your data shape and how much a wrong answer actually costs — and every prompt after it reasons about your real risk profile rather than treating a support bot and a medical triage tool the same way.
Evaluation comes first because it is what teams skip. The prompts build an eval set from failures you have actually observed rather than synthetic examples that pass by construction, and they are direct about the limits of using a model to judge its own output. The RAG section separates retrieval failure from generation failure, since fixing the prompt when retrieval returned the wrong chunk wastes weeks.
Several prompts will tell you an agent is the wrong architecture, which is more often true than the current enthusiasm suggests.
Compare with Similar AI Templates
| Specification | ||||
|---|---|---|---|---|
| Customer Rating | 4.92(36) | 4.91(43) | 4.93(45) | 4.91(38) |
| AI Models | GPT-5, Claude Opus 5, Gemini 3.7 Flash | GPT-5, Claude Opus 5, Gemini 3.7 Flash | GPT-5, Claude Opus 5, Gemini 3.7 Flash | GPT-5, Claude Opus 5, Gemini 3.7 Flash |
| Formats Included | PDF Prompt Playbook | PDF Prompt Playbook | PDF Prompt Playbook | PDF Prompt Playbook |
| Lifetime Updates | Yes | Yes | Yes | Yes |
| Creator Rating | Dr. Lucas Vance (99.1%) | Dr. Lucas Vance (99.2%) | Marcus Sterling (99.3%) | Elena Rostova (99%) |
Prompt Architect & Author Spotlight
Works with teams putting language models into production. Mostly concerned with evaluation, because that is where shipped AI features quietly fail.
Community Reviews & Ratings
Verified buyer ratings & prompt benchmarks
Write a Review
Sign in to leave a review for this system.
Verified User Feedback
We had been fixing the wrong layer for a month
The retrieval-versus-generation diagnostic showed the correct chunk was being retrieved but ranked fourth and falling out of context. We had been rewriting the prompt. Fixed the reranker instead and the problem went away.
Eval set from real failures
Our test cases were examples we wrote, which all passed. Building the set from logged failures instead surfaced three systematic problems within a day.
Told us not to build an agent
Described our workflow and it argued a deterministic pipeline with one model call would be more reliable and far cheaper. It was right. Refreshing given how much pressure there is to ship an agent.
Assumes some maturity
The evaluation material is excellent but assumes you are logging failures already. If you are not, start there — which the toolkit does say, just later than I would have liked.
Frequently Asked Questions
How is this different from the separate systems you used to sell?
It is all of them, in one purchase. LLM Evaluation & Benchmark Plan, Prompt Engineering Test Suite, RAG Retrieval & Chunking Strategy, Agent Workflow & Task Planner, Model Card & Documentation Builder, Ethics & Governance Framework, Chatbot & Conversational Flow, Voice Assistant & Skill Designer, Personal AI Agent & Task Plan, Knowledge Graph & Ontology are now sections inside this toolkit. Bought individually they came to $199.
Does it assume a particular model or framework?
No. The System Primer asks which model and provider you use, and the prompts adapt where behaviour genuinely differs. Nothing is tied to LangChain, LlamaIndex or any specific framework — the material is about method rather than library.
Will it help me build an agent?
Yes, and it will also tell you when not to. The agent section covers decomposition, tool definitions, retry handling and cost guardrails, but several prompts explicitly test whether a deterministic pipeline would do the same job more reliably — which is often the honest answer.
Is the governance section legally sufficient?
No. It covers acceptable use policies, human-in-the-loop decisions, logging and incident response as engineering and operational practice. Regulatory compliance in your jurisdiction and sector needs a lawyer, and the prompts flag where you have crossed into that territory.
How is it delivered?
You receive a PDF with all 138 prompts organized into 9 sections, delivered within 24 hours of purchase. No app or account required — copy a prompt, fill in the bracketed variables, and paste it into ChatGPT, Claude, or Gemini.
Can I use this for client work?
Yes. Your purchase includes a single-user commercial and personal license, so you can use every prompt in your own business or for client deliverables without restriction.