Case Study 04 · AI Workflow / Process Design

Designing a Gated AI Production System Instead of “Prompting a Book”

A long-form AI workflow built around stage gates, canonical inputs, scope controls, revision checks, and human approval points. The objective was to make AI useful across a large project without allowing speed of generation to outrun structural coherence.

The problem

Long-form AI projects create a specific failure mode: generation is cheap, so the system can produce large amounts of polished material before the underlying architecture is stable.

That creates downstream problems that are expensive to repair:

  • chapters duplicate or contradict one another
  • later ideas leak backward into sections that are supposed to establish foundations
  • tone and terminology drift across sessions
  • interesting tangents gradually redefine scope
  • AI reconstructs prior decisions imperfectly from memory
  • supporting artifacts are generated before their conceptual purpose is clear
Design response

Treat the project as a production system with controlled state transitions. AI can expand, draft, revise, and render—but only after the relevant upstream specification has been approved.

Core operating principle

Do not use AI autonomy where unresolved human decisions still determine the shape of the work. Resolve architecture first; delegate production after the decision boundary is stable.
Specify Approve Expand Generate Evaluate Advance

The six-phase production workflow

Phase 01

Canonical structure

Lock project identity, thesis, major sections, sequence, and overall logic before detailed generation begins.

GateCan the project be described coherently at the whole-system level?
Phase 02

Section / chapter skeletons

Define each unit's purpose, key questions, core claims, exclusions, required examples, and tone constraints.

GateNo prose drafting until the skeleton is approved.
Phase 03

Core-ideas expansion

Develop the skeleton into detailed internal architecture: arguments, definitions, contrasts, examples, transitions, research anchors, and drift controls.

GateIs the unit detailed enough that drafting no longer requires inventing its structure?
Phase 04

Full drafting from canonical inputs

Draft one complete unit at a time using the approved skeleton and bullet architecture as structural source files.

GateDo not override uploaded structure or reconstruct it from conversational memory.
Phase 05

Revision and alignment

Review the generated material against explicit failure categories rather than relying on a generic “make it better” revision prompt.

ChecksDrift, redundancy, tone inconsistency, overstatement, historical sloppiness, accidental reductionism, conceptual bleed.
Phase 06

Supporting artifacts

Only after the conceptual structure is stable, generate appendices, diagrams, quick-reference tools, tables, case studies, and other derivative assets.

GateSupporting artifacts inherit the approved terminology, structural logic, and purpose of the core work.

Control mechanisms

Control
What it prevents
Stage gates
Prevents AI from solving downstream tasks before upstream human decisions are settled.
Canonical files
Prevents conversational-memory drift. Drafting is grounded in the actual approved skeleton and architecture files.
One unit at a time
Limits blast radius. Each chapter, diagram, or artifact remains independently inspectable before the workflow advances.
Explicit exclusions
Prevents scope expansion caused by plausible but misplaced material.
Output contracts
Controls what the model should return: prose only, specification only, one artifact only, no extra commentary, or another defined surface.
Revision taxonomy
Turns quality review into targeted evaluation instead of vague polishing.
Human approval
Keeps irreversible or structurally consequential transitions under deliberate control.

Canonical-input discipline

Common failure mode

A long-running conversation accumulates context. The model is then asked to continue “from what we already decided.” Small reconstruction errors compound over many generated sections.

Workflow response

The drafting stage required the actual structural files for each unit:

  • chapter name
  • approved skeleton
  • approved bullet architecture

The rule was explicit: use those files as structural source material; do not invent replacement structure and do not draft from memory.

Quality review as evaluation

Revision was treated as a separate evaluation stage with named failure classes. This is important because an AI can improve sentence-level style while making the larger system worse.

Drift

Did the output move away from the approved purpose or thesis?

Redundancy

Did separate units begin doing the same conceptual work?

Tone inconsistency

Did the voice or level of explanation change across the project?

Overstatement

Did generated prose become more certain than the evidence or approved model allows?

Conceptual bleed

Did material cross boundaries and appear before or outside its intended location?

Integrity failures

Did simplification, synthesis, or stylistic fluency distort an important distinction?

Extending the same logic to artifact generation

The workflow was later applied to diagrams and supporting visuals. Instead of jumping directly from idea to image, generation was split into separate states:

Name artifact Write structural specification Human review Generate artifact

Specification contains

  • purpose and conceptual role
  • layout / composition
  • major zones and node contents
  • relationships and flow direction
  • visual hierarchy
  • what must be obvious at first glance
  • what must be avoided

Why separate specification from generation?

Image or artifact generation creates a high-cost branch in the workflow: once visual details are rendered, it becomes tempting to edit the artifact instead of questioning its structure. Reviewing the specification first keeps structural correction cheap.

Human–AI division of labor

Human responsibility

  • define project purpose and success
  • approve architecture and boundaries
  • decide which distinctions are non-negotiable
  • judge whether a stage is ready to advance
  • evaluate whether outputs remain faithful to intent
  • change the workflow when recurring failure patterns appear

AI contribution

  • expand approved structures rapidly
  • generate candidate arguments, examples, and transitions
  • draft from canonical inputs
  • perform structured revision passes
  • translate core material into derivative formats
  • maintain production throughput across many artifacts

The transferable method

The original project was long-form research and publishing, but the process generalizes to any AI-assisted workflow in which output quality depends on preserving state and preventing premature execution.

Business / technical analogueEquivalent workflow control
Product requirements → implementationDo not generate/build before acceptance criteria and boundaries are stable.
Research brief → final reportLock questions, evidence classes, and claim boundaries before synthesis.
Knowledge-base migrationTreat approved source documents as canonical inputs, not conversational recollection.
AI content pipelineSeparate architecture, generation, QA, and derivative-format stages.
Agentic workflowAssign explicit approval points where downstream action depends on unresolved judgment.
AI QAUse named failure categories and completion criteria instead of generic quality impressions.
AI orchestration workflow design quality gates specification writing canonical inputs human-in-the-loop process QA

Result

The system converted an open-ended, multi-stage intellectual project into a sequence of bounded AI tasks with explicit inputs, outputs, review criteria, and human decision points. The key capability demonstrated is not “writing good prompts.” It is designing the process through which AI is allowed to contribute.