Skip to content

Prompt Architecture

One giant "write me a chapter about AI infrastructure" prompt produces garbage. Not sometimes -- reliably. The output sounds confident, covers the topic, and reads like every other AI-generated business book you've never finished. No voice. No edge. No consistency between sections.

The fix: modular prompts that compose at runtime.

Why Modular

Think of it like CSS inheritance. You have base styles (the master system prompt) that apply everywhere. Then you layer on component-specific styles (the task prompt) that handle the particular job. The master system prompt carries voice, audience, constraints, and density rules. The task prompt carries structure, word targets, and content-specific instructions.

This gives you three things monolithic prompts can't:

  1. Testability. When chapter openings sound formulaic, you fix one prompt. You don't debug 2,000 words of tangled instructions to figure out which sentence caused the drift.
  2. Composability. The same master prompt works with a case study prompt, a framework prompt, or an editing prompt. Mix and match without rewriting.
  3. Independent improvement. The De-AI editing prompt went through 4 iterations. None of those changes touched the writing prompts. No cascading side effects.

The 27 Prompts in 5 Categories

Category Count Purpose
Writing 8 + master Create new content (openings, sections, frameworks, case studies, analogies, summaries, closings, technical deep dives)
Editing 5 Refine existing content (voice, audience, de-AI, balance, simplify)
Review 4 Quality gates (chapter complete, authenticity, audience fit, final quality)
Linking 3 Connect content (concepts, cross-references, broken link fixes)
Fix 5 Targeted improvements (generic language, examples, action items, two-audience, tighten prose)

How Prompts Compose

flowchart TD
    MSP["Master System Prompt\n(voice + audience + rules + density test)\nAlways present"]

    MSP --> W["+ Writing Prompt\n(opening, section, framework, etc.)"]
    MSP --> E["+ Editing Prompt\n(voice, de-AI, balance, etc.)"]
    MSP --> R["+ Review Prompt\n(chapter complete, authenticity, etc.)"]
    MSP --> L["+ Linking Prompt\n(concepts, cross-references, etc.)"]
    MSP --> F["+ Fix Prompt\n(generic language, tighten, etc.)"]

    W --> WO["Content Output\n(draft sections)"]
    E --> EO["Refined Content\n(with change notes)"]
    R --> RO["Quality Report\n(grade + issues)"]
    L --> LO["Connected Content\n(linked sections)"]
    F --> FO["Targeted Fixes\n(specific improvements)"]

    style MSP fill:#1a1a2e,stroke:#e94560,color:#fff
    style W fill:#16213e,stroke:#0f3460,color:#fff
    style E fill:#16213e,stroke:#0f3460,color:#fff
    style R fill:#16213e,stroke:#0f3460,color:#fff
    style L fill:#16213e,stroke:#0f3460,color:#fff
    style F fill:#16213e,stroke:#0f3460,color:#fff

The master system prompt gets prepended as system context to every AI interaction. The orchestration skill (the chapter-writer) handles this automatically. Individual prompts assume it's already present -- they don't repeat voice rules or audience profiles.

3 Workflow Patterns

1. New Chapter (Writing Flow)

Prompts fire in sequence because each builds on the previous output:

master + chapter-opening   --> Hook + context + preview
master + section           --> Core content (repeat per section)
master + framework         --> Numbered frameworks woven in
master + case-study        --> Company narratives where needed
master + chapter-closing   --> Synthesis + takeaways + bridge
master + tighten-prose     --> Density pass (cut 20%)

The section prompt is the workhorse -- it runs 4-6 times per chapter. The framework and case-study prompts get inserted where the content calls for them, not mechanically in every section.

2. Edit Pass (Refinement Flow)

After a draft exists, editing prompts refine it:

master + voice-consistency     --> Align to authenticity markers
master + de-ai-patterns        --> Remove AI cliches and tells
master + audience-calibration  --> Balance Sarah and Marcus
master + simplify              --> Clarity pass
master + tighten-prose         --> Density pass (final cut)

The de-AI prompt is the single most valuable editing prompt in the system. It catches patterns that are invisible to the writer -- the "leverage" and "important to note" phrases that slip through even with voice rules in place.

3. Final Review (Quality Flow)

Review prompts produce reports, not rewrites:

master + chapter-complete  --> Full structural checklist (A-D grade)
master + authenticity      --> Voice fingerprint verification
master + audience-fit      --> Both audiences served?
master + final-quality     --> Go/no-go with remaining issues

These are quality gates. They don't change the content -- they tell you what to change. The output is a graded report with specific line numbers and recommendations.

Building Your Own

Start with three prompts: a master system prompt, a section writing prompt, and a de-AI editing prompt. That's enough to produce consistent, voice-accurate sections. Add prompts as you discover specific quality problems -- each new prompt should solve a real failure mode, not a theoretical one.

The full 27-prompt system wasn't designed upfront. It grew across three drafts as specific problems surfaced. The framework prompt exists because early frameworks had inconsistent point lengths. The tighten-prose prompt exists because first drafts were reliably 20-30% too long. Every prompt earned its place by fixing something that broke.

Deep dives: Master System Prompt | Writing Prompts | Editing and Review Prompts | Prompt Evolution