Results and Metrics¶
Numbers don't lie. Here's what the system produced over the course of writing Blueprint for An AI-First Company -- and what those numbers actually tell you about AI-assisted book writing.
Manuscript Metrics¶
| Metric | Value |
|---|---|
| Total words | 81,122 |
| Chapters | 12 |
| Parts | 4 |
| Sections | 81 |
| Target words per chapter | ~6,500 |
| Target words per section | ~1,200 |
| Inline citations | 775 |
| Citation density | ~1 per 105 words |
| Concept notes | 9 |
| Total vault links | 1,199 |
| Average links per section | 7.0 |
| Drafts produced | 3 complete |
System Components¶
The writing system wasn't one tool. It was layers of tools, each solving a specific problem.
| Component | Scale |
|---|---|
| Author voice files | 6 |
| Modular prompts | 27 (5 categories) |
| Claude Code skills | 14 |
| Claude Code agents | 3 (writer, reviewer, prompt writer) |
| Research prompts | 180+ |
| Python scripts | 17 |
| Book intelligence app modules | 70+ |
| Database migrations | 5 |
| Obsidian plugins used | 18 |
See Architecture Decisions for why each component exists.
Quality Pipeline Results¶
Every section passed through automated quality checks before human review. Here's what the pipeline covered:
| Quality Dimension | Tool | Coverage |
|---|---|---|
| Voice consistency | check-voice skill |
All 81 sections |
| Citation audit | check-citations skill |
All 81 sections |
| Opening variety | audit-openings skill |
All 81 sections |
| Link structure | audit-links skill |
Full vault |
| Term diversity | analyze-terms skill |
Full manuscript |
| Research coverage | map-research skill |
All 12 chapters |
Editorial Review Results¶
Three drafts, each tighter than the last. The editing pipeline ran in phases, each catching different classes of issues:
| Phase | Issues Found | Issues Resolved |
|---|---|---|
| Developmental editing | Structural issues across 12 chapters | All resolved |
| Line editing | 240 issues | All resolved |
| Copyediting | 188 issues | All resolved |
| Final verification | 15 critical issues | All resolved |
| Big themes review (10 dimensions) | 0 critical, 3 important, 2 minor | Publication ready |
The big themes review scored the manuscript across 10 dimensions -- voice consistency, argument coherence, audience calibration, practical density, and six others. Zero critical issues. That doesn't happen by accident; it happens because the quality pipeline caught problems early enough that they never compounded.
Research Pipeline Output¶
The research didn't start with writing. It started with programmatically generating 180+ prompts for Perplexity Pro, organized by chapter and section. A pre-research phase used web search to inform prompt design -- making the actual research prompts sharper than anything I'd write cold.
Playwright automation executed the Perplexity searches, collecting raw research into structured files. From there, synthesis scripts extracted the pieces a writer actually needs: statistics with source attribution, direct quotes from leaders, company examples with specifics, and analytical frameworks.
The result: by the time I sat down to write any section, I had citation-ready content waiting. Stats already formatted with footnote keys. Quotes already attributed. The writing session became about argument and voice, not hunting for evidence.
See End-to-End Flow for how research connects to writing connects to review.
Vault Health Metrics¶
The Obsidian vault started as flat files. It ended as a knowledge graph:
| Metric | Before Enhancement | After Enhancement |
|---|---|---|
| Section-to-concept links | 0% | 68% |
| Average links per section | ~1 | 7.0 |
| Total vault links | ~630 | 1,199 |
| Concept notes | 5 | 9 |
Links aren't decoration. They're how the manuscript maintains internal consistency across 81 sections. When Chapter 9 references a pattern from Chapter 4, the link makes that relationship explicit and auditable.
What the Numbers Mean¶
The citation density -- one citation per 105 words -- is higher than most business non-fiction. That's not because citations make writing better. It's because the research-first pipeline made citation cheap. When stats arrive pre-formatted with footnote keys, you use them. When you have to manually hunt down sources mid-sentence, you don't. The system changed the economics of evidence.
Voice consistency across 81 sections is the hardest metric to hit. It's easy to maintain voice for a chapter. Maintaining it across 12 chapters written over weeks, with AI generating first drafts? That's where the 6-file voice system earns its keep. The author voice guide, quick reference, gold standard, authenticity markers, audience empathy guide, and learnings file -- together they gave every writing session the same constraints. Constraints produce consistency.
Three complete drafts sounds like a lot of rework. It was. But here's the thing: each draft improved the system as much as the manuscript. Draft 1 revealed which prompts produced generic output. Draft 2 exposed where the review pipeline had gaps. Draft 3 was the draft where the system finally matched the ambition. The manuscript was the deliverable; the system was the real product.
The quality pipeline -- 14 skills running automated audits -- catches what human review misses. Repeated opening patterns across chapters. Inconsistent terminology. Orphaned concept references. But human review catches what automated audits miss: whether an argument actually lands, whether an analogy clarifies or confuses, whether the reader would keep going or put the book down. You need both. Neither is optional.