A complete visual companion to the book From Copilot to Colleague: How AI Engineering Turns Models into Dependable Systems.
Fourteen diagrams: four that map the book and the project that produces it, and ten that open each chapter with a concrete before → after.
Every diagram is a hand-built Excalidraw file (.excalidraw, fully editable) rendered to PNG. They are designed to argue visually — each shape mirrors the concept it represents, not just label a box.
The governing thesis: Models create possibility. Scaffolding creates trust. Organizations decide whether that trust compounds.
Four diagrams for the whole project: its argument, how it is made, its core thesis, and its evidence base.
The ten chapters as a four-act dependency arc — the Problem (ch 1–2), the Scaffolding Stack (ch 3–7), the Stress Test (ch 8), and the Widening (ch 9–10). It shows the book is an argument with a shape, not a survey, and carries the throughline from possibility to compounding trust.
How a 794-video corpus becomes a source-backed book: a five-layer pipeline — Source → Synthesis → Evidence → Manuscript — governed by a Research-Org control plane that improves the upper layers through bounded, logged agent passes. Real directory names are shown as evidence.
The book's central thesis in one picture: a raw model is capable but not dependable. Five engineered layers wrap it — harnesses, evals, context, runtimes, security — and each one earns its place by a specific failure it prevents.
What 794 videos are actually about: ten themes sized by corpus count, colour-coded by the book act they feed, each mapped to its chapter — including the honest editorial call that Models & Inference is deliberately background, not a chapter.
Each chapter diagram contrasts the naive way — how AI work goes wrong today — with the engineered way — the best practice the chapter argues for. Both sides carry a concrete code or config sample, and the chapter's four strongest claims sit underneath.
The book's opening move: AI stops being a better autocomplete and becomes a candidate worker. Assistant vs delegate, side by side — and the central question the whole book answers: what must be true before teams can trust AI to act?
When generation gets cheap, judgment gets scarce. Vibe coding (ship whatever runs) vs vibe engineering (frame, constrain, review) — shown with the same intent written two ways.
Agent quality depends on harness quality. A bare prompt into a messy repo vs a prepared environment — AGENTS.md, specs, tests as guardrails, a staged plan → produce → review → ship loop.
Trust has to be measured. Demo vibes and a leaderboard score vs an operational eval loop — human-seeded cases, task-level scoring, and production traces that feed the next generation of evals.
Prompt quality is downstream of information architecture. Stuffing 180k tokens into the window vs assembling context by layer — RAG, memory, GraphRAG, intent-aware tool selection.
Long-running work needs architecture. A stateless chat loop that loses everything on a crash vs a durable runtime with checkpoints, replay, and a human approval gate.
The moment a system can act, it becomes an attack surface. A helpful agent with admin credentials vs bounded autonomy — scoped identity, sandboxing, least privilege, an audit trail, a root of trust.
Voice and robotics stress-test the thesis. Text-chat assumptions out loud (≈3.9 s of lag) vs a realtime pipeline with a sub-800 ms latency budget, streaming, turn detection, and barge-in.
The deepest gains are organizational. Buying 500 AI seats and changing nothing else vs an operating-model redesign — broaden who can create, tighten review and governance, pay down alignment debt.
The close: separate the tools that churn every few months from the operating model that survives them — constrained delegation, explicit context, eval loops, durable runtimes, bounded trust, and human judgment.
- Built with the
excalidraw-diagramskill — diagrams that argue visually, hand-authored as Excalidraw JSON and validated through a render-and-review loop. - Sourced from the book knowledge base: chapter packets, the source-backed outline, the theme map, and the project architecture docs.
- To edit: open any
.excalidrawfile at excalidraw.com (or in the Excalidraw VS Code / Obsidian plugin). - To re-render to PNG:
npx @excalidraw-skill-pack/render <path-to-file.excalidraw>
The book's diagram identity is documented in STYLE.md — palette mapping, the blue→green signature mark, title system, citation style, and defensibility rules.
Reusable skeletons live in templates/ — copy one, fill the [bracketed] placeholders, render:
inline-figure.excalidraw— a small single-idea section figureconcept-card.excalidraw— a standalone concept explainerchapter-card.excalidraw— a full-page before → after chapter diagramlayered-stack.excalidraw— "X is built of layers"flow-pipeline.excalidraw— a process or pipelinerelationship-map.excalidraw— nodes and edges (graphs, maps)
| # | Diagram | Source | Render |
|---|---|---|---|
| 1 | The Argument Spine | 01-book-argument-spine.excalidraw |
01-book-argument-spine.png |
| 2 | The Autoresearch Knowledge Machine | 02-autoresearch-machine.excalidraw |
02-autoresearch-machine.png |
| 3 | The Scaffolding Stack | 03-scaffolding-stack.excalidraw |
03-scaffolding-stack.png |
| 4 | Theme & Corpus Map | 04-theme-corpus-map.excalidraw |
04-theme-corpus-map.png |
| 5 | Chapter 1 — The Shift | 05-chapter1-the-shift.excalidraw |
05-chapter1-the-shift.png |
| 6 | Chapter 2 — Taste | 06-chapter2-taste.excalidraw |
06-chapter2-taste.png |
| 7 | Chapter 3 — Harnesses | 07-chapter3-harnesses.excalidraw |
07-chapter3-harnesses.png |
| 8 | Chapter 4 — Evals | 08-chapter4-evals.excalidraw |
08-chapter4-evals.png |
| 9 | Chapter 5 — Context | 09-chapter5-context.excalidraw |
09-chapter5-context.png |
| 10 | Chapter 6 — Runtimes | 10-chapter6-runtimes.excalidraw |
10-chapter6-runtimes.png |
| 11 | Chapter 7 — Security | 11-chapter7-security.excalidraw |
11-chapter7-security.png |
| 12 | Chapter 8 — Realtime | 12-chapter8-realtime.excalidraw |
12-chapter8-realtime.png |
| 13 | Chapter 9 — AI-Native Org | 13-chapter9-ai-native-org.excalidraw |
13-chapter9-ai-native-org.png |
| 14 | Chapter 10 — What Endures | 14-chapter10-what-endures.excalidraw |
14-chapter10-what-endures.png |