AI Consulting Playbook v1.0.0

This handbook now has its own repository and canonical v1.0.0 release. This page preserves the original release record and exact PDF.

First edition, October 1, 2026. 138 pages.

AI Consulting Playbook takes an AI opportunity through discovery, workflow design, governance, pilot evidence, delivery, adoption and accountable handoff. It includes twelve chapters, worksheets, a worked fictional case, and a capstone that keeps failed evidence gates and missing outcomes visible.

Read the playbook. The PDF is attached below. Source-access limits and accepted visual observations are documented in the repository’s review packet.

PDF SHA-256: 4a7055b53ec4e4db7dbafacb6895513b055b4adeb02354c9f38fe53999ed92b2.

Current public edition: https://handbooks.surfaces.systems/ai-consulting/editions/v1.0.0/
Canonical source: https://github.com/mikeylong/handbooks . Historical release assets are retained.

Production AI Engineering 1.4.0

Version 1.4.0 expands the handbook’s guidance for systems that compact conversation state, change models or tools during a run, depend on model judges, and share serving infrastructure across tenants.

Compared with 1.3.1, the changelog records 16 changes: 13 additions, one deprecation, one correction, and one clarification. The edition also includes a prose edit for clearer, more direct instructions.

1. Preserve the state a continuing conversation needs. Compaction becomes an explicit state transition, with durable evidence, exact preservation of signed summaries, separate usage accounting, and a successful continuation before accepting the compacted conversation. Handbook: §2.1, pp. 7–8.

2. Verify tool results when they arrive. Late results are reauthorized against the current run and resource. Dynamic tool registries are versioned, and consequential outputs are checked against authoritative state even when their schema is valid. Handbook: §5.2, pp. 29–30.

3. Plan migrations around the whole runtime. The compatibility matrix now includes managed agent surfaces and built-in tools. Preserved reasoning state is tied to the provider, model, account, history, and API contract. Handbook: §6.2, pp. 34–36.

4. Check whether evaluation remains stable. Black-box judges are calibrated on the endpoint and release window actually used. Multimodal injection tests distinguish attempted actions from completed external effects. Handbook: §8.1, pp. 47–48.

5. Govern credentials and agent skills. API keys receive explicit ownership, lifetime, rotation, and revocation controls. Agent skills and their supporting files are reviewed and pinned as software supply-chain inputs. Handbook: §10.1, p. 58.

6. Contain failures before they reach shared infrastructure. Structured-output compilation is bounded and isolated. Media, request fields, and transformed outputs receive limits at the earliest acquisition point, so malformed requests do not poison a shared engine. Handbook: §5.1, p. 24; §10.2, p. 60.

Page references use the printed page numbers in the 1.4.0 handbook PDF. Appendix A, “Notes and Sources,” pp. 74–87, connects the guidance to primary documentation, security advisories, and original research. The complete changelog follows the Index on pp. 94–102.

The PDF has 105 physical A5 pages. Version 1.3.1 remains available as a separate release.

SHA-256: dbed169d2eb2fe6f14d43c9f72fac59113fc6de0f2a42d63de36cd75602128c0

Current public edition: https://handbooks.surfaces.systems/production-ai-engineering/editions/v1.4.0/
Canonical source: https://github.com/mikeylong/handbooks . Historical release assets are retained.

AI Evaluation Field Guide 1.0.0

This handbook now has its own repository and canonical v1.0.0 release. This page preserves the original release record and exact PDF.

AI Evaluation Field Guide 1.0.0 is the first public edition: 100 A5 pages with 12 chapters, 12 worksheets, a guided capstone, and 25 sources.

The guide takes a beginner from defining expected behavior through evidence design, human and model grading, uncertainty, agent side effects, launch decisions, and production monitoring. The complete edition removes course attribution and broadens the source base with HELM, CheckList, Datasheets for Datasets, tau-bench, AgentDojo, idempotent API guidance, and statistical references.

Read online. The attached PDF is the exact reviewed artifact served by that reader. Structural validation, complete Poppler and macOS PDFKit review, reader review, release CI, and desktop/mobile browser checks passed.

SHA-256: 91531220eb1567eb5a51bcd0083c22ccbc77b18f88c275596382b955acf762ad

Current public edition: https://handbooks.surfaces.systems/ai-evaluation/editions/v1.0.0/
Canonical source: https://github.com/mikeylong/handbooks . Historical release assets are retained.

Production AI Engineering 1.3.1

What Changed for Teams Adopting AI Engineering

Across versions, the handbook has developed from getting a model to work toward operating AI as an accountable production system. These are adopter-facing takeaways from the handbook, not claims that v1.3.1 added new engineering chapters. The reader-facing content in v1.3.1 is unchanged from v1.3.0.

1. Teams start with the user outcome. Model and prompt choices sit inside clear requirements for quality, latency, cost, permissions, and failure handling. This reduces the gap between a successful demo and a dependable product. Handbook: ch. 1, p. 1; §1.1, pp. 2–3; §1.2, pp. 3–4; §12.2, pp. 65–66.

2. Optimization begins with diagnosis. Teams measure the full workflow before changing caching, routing, quantization, or serving. This helps avoid local improvements that quietly increase cost, latency, or quality risk somewhere else. Handbook: §8.1, p. 43; §11.2, pp. 61–62; §12.3, p. 69.

3. Agent adoption comes with ownership. Agents need distinct identities, limited authority, tracked jobs, stop conditions, and recovery paths. Teams can automate meaningful work without losing control of who acted, what was allowed, or whether the task completed. Handbook: §5.2, pp. 26–28; §6.1, pp. 30–32; §12.2, p. 67.

4. Evaluation reflects real production behavior. Teams assess the path an agent took, the external effects it caused, and how it behaves under hostile or impossible conditions. A good final answer alone is no longer enough to pass a release gate. Handbook: §6.1, p. 31; §8.1, pp. 43–46, especially p. 45.

5. Retrieval becomes a trust decision. Teams evaluate whether evidence is relevant, authorized, private, and useful for the decision it supports. This reduces the risk of treating a plausible answer as a grounded one. Handbook: §7.1, pp. 36–38; §7.2, pp. 38–41, especially pp. 39–40.

6. Operations shape the design earlier. Provider changes, overload, degraded service, tenant isolation, monitoring, and incident controls are planned before launch. Teams can update the system without turning every model or SDK change into a production surprise. Handbook: §6.2, pp. 32–34; §9.1, pp. 48–50; §§10.1–10.2, pp. 54–58; §§12.1–12.2, pp. 64–68.

The practical shift is from a collection of AI techniques to a clearer way of deciding what to build, measure, approve, release, and operate.

Page numbers refer to the printed pages in the v1.3.1 handbook PDF. The chapter citation labels resolve to the primary references and marked author synthesis in Appendix A, “Notes and Sources,” pp. 70–81.

Current public edition: https://handbooks.surfaces.systems/production-ai-engineering/editions/v1.4.0/
Canonical source: https://github.com/mikeylong/handbooks . Historical release assets are retained.

Production AI Engineering 1.3

Production AI Engineering 1.3 adds practical guidance on cache reuse, KV memory measurement, asynchronous tool jobs, model migrations, overload recovery, monitoring, and token-input safety.

The edition includes seven new production controls, four source-trail updates, and an evidence-review date correction. No guidance was deprecated or removed. The PDF includes the complete changelog with primary-source links.

Download the versioned PDF attached to this release.

Current public edition: https://handbooks.surfaces.systems/production-ai-engineering/editions/v1.4.0/
Canonical source: https://github.com/mikeylong/handbooks . Historical release assets are retained.

Production AI Engineering 1.2

Production AI Engineering 1.2 updates the handbook for current production AI practice.

This edition includes 27 reader-facing changes: 21 additions, one deprecation, three corrections, and two clarifications. The updates cover:

• safer MCP authorization, transport, schema, and migration boundaries;
• workflow-level serving, cache-cost attribution, and energy measurement;
• retrieval confidentiality, freshness, and interaction-aware evaluation;
• stronger acceptance gates for agents, automated research, and runtime isolation.

The complete source-linked changelog is included at the end of the PDF.

Published August 31, 2026. Release asset: production-ai-engineering-v1.2.0-2026-08-31.pdf.

Current public edition: https://handbooks.surfaces.systems/production-ai-engineering/editions/v1.4.0/
Canonical source: https://github.com/mikeylong/handbooks . Historical release assets are retained.

Production AI Engineering 1.1.1

Production AI Engineering 1.1.1 is the August 27, 2026 compatibility release.

What changed

• Fixed chapter illustration rendering in Apple Preview and iOS PDF viewers.
• Normalized all 12 print illustrations to one opaque image each, eliminating the soft-mask and duplicate-paint structure that affected Chapters 3, 5, and 9.
• Preserved the PDF 1.4 header through finalization and added structural regression checks to the handbook production pipeline.

The handbook's reader-facing guidance is unchanged. No sections were added, removed, deprecated, corrected, or clarified, so this edition does not include a changelog page.

The attached PDF is 78 A5 pages.

SHA-256: b89e9c96c766345aedad90f73de799c5586a2c4c87a5760cbbede1f774f9e15a

Current public edition: https://handbooks.surfaces.systems/production-ai-engineering/editions/v1.4.0/
Canonical source: https://github.com/mikeylong/handbooks . Historical release assets are retained.

Production AI Engineering 1.1

Production AI Engineering 1.1 is the August 24, 2026 edition.

What changed

• Added guidance for KV-cache reservation, deterministic tools, MCP migrations, serving-engine isolation, provider-resolved boundaries, and supporting source trails.
• Corrected protocol and session testing, route compatibility, and provider, MCP, and tool references.
• Clarified OAuth and MCP discovery, registration, SSRF, and redirect-hop controls.

No sections were removed or deprecated.

The attached PDF is 79 A5 pages. The full changelog is on the final page.

SHA-256: 086133c723b8cf818eb2dc65e033cc81b863830c31a6734f4e89e94761b8e4a4

Current public edition: https://handbooks.surfaces.systems/production-ai-engineering/editions/v1.4.0/
Canonical source: https://github.com/mikeylong/handbooks . Historical release assets are retained.