Writing a Memoir in 20 Days

55 scenes. One human. One AI. Neither gets here alone.

← All case studies

The problem

Bill's father died on December 23, 2025. Bill needed to write something true for his family — a memoir, scenes from his life, a record of who he was when he wasn't being someone's employee. He had a full-time job running data engineering across six countries. He had 20 days.

What the human did

Everything that matters.

• Wrote every word of prose. Every sentence in the book came from Bill's hand.
• Designed the structure: a present-tense first-person scene-opener ("I am [age] and..."), non-linear reading order, three acts, 55 scenes.
• Decided the reading order. The chronological order and the reading order are different. The non-linearity is the point — memory isn't sequential.
• Made every creative decision: what stays, what goes, what the ending means, who the book is for.

What the AI did

Everything that doesn't.

Scene management

55 markdown files in a scenes/ directory. Each scene is a standalone file with a numerical prefix for chronological order. But the book doesn't read in chronological order — it reads in emotional order. The AI maintained the mapping between the two:

READING_ORDER = [
    '13_rosebud.md',          # Opens the book — Gary's death
    '01_della.md',            # Grandmother
    '02_turandot_piano.md',   # Music and memory
    '03_turandot_festival.md',
    '04_the_dam.md',
    ...
    '50_the_mirror.md',
    '51_sadie_reviews.md',    # Daughter reads the manuscript
    '52_the_epilogue.md',     # Close
]

The reading order array is the spine of the book. Every assembly operation runs against it. When Bill moved a scene, the AI updated the array and rechecked every dependency.

Cross-scene tracking

Many scenes contain setups that pay off later — a line planted in scene 4 that lands in scene 38. Move a scene and you can break a setup. The AI tracked the cross-scene references so that restructuring the reading order didn't orphan a setup or land a punchline before its setup arrived.

Assembly pipeline

A Python script (assemble_codex.py) reads the 55 markdown files in reading order, converts them to HTML, inserts act breaks and part titles, bolts on the making-of appendix and the teacher's guide, and outputs a single HTML file ready for rendering.

The codex has three parts:

1. Part One: The Memoir — 55 scenes in three acts
2. Part Two: The Making Of — a production log documenting how the book was built
3. Part Three: Teacher's Guide — scene-by-scene references and notes

Front matter includes a half-title, title page, copyright page, and dedication.

Output formats

From a single HTML assembly, the AI generated:

• PDF — for digital distribution (via weasyprint)
• Print-ready PDF — 6×9 trade paperback, proper margins, page numbers, widow/orphan control
• EPUB — for e-readers
• DOCX — for people who insist
• HTML — the canonical format, viewable in any browser

The print layout handles gutter margins (wider on the spine side), alternating left/right page numbers, act break pages with no numbering, and front matter with no numbering. All in CSS @page rules — no LaTeX, no InDesign, no framework.

Distribution

A separate system (book_factory.py) handles distribution. It scans email for book requests, cross-references an identity ledger that tracks people by person (not by email address), generates fresh presigned S3 URLs with 24-hour expiration, sends personalized emails via SMTP, and updates the ledger. One command: python3 book_factory.py. Total infrastructure cost: $0/month.

The license model: WinRAR rules. The book is free. Copy it, share it, print it. The only ask: tell the author who it's for.

The timeline

The split

This is the part that matters. Read it twice.

Neither gets here alone. Bill without the AI has a stack of markdown files and no book. The AI without Bill has a perfectly formatted collection of nothing. The combination produced a 172-page codex in 20 days, distributed by request to friends and family.

An honest note on what this is and isn't

This is a real book by a real human who isn't a literary writer. It's a memoir Bill made for his family because his father died and left almost nothing personal behind, and Bill didn't want his kids to be in that position one day. The technique is the case study. The book itself is a strange beanie baby — built carefully, given freely, and as good as a non-novelist working in 20 days could make it. That's the honest claim. Anything more would be inflation.

What you can steal from this

1. Separate creation from production. The writing and the bookmaking are different jobs. Do them with different tools.
2. Use plain text as your canonical format. Markdown files are future-proof. Every output format is generated from them. If the tools change, the source doesn't.
3. Track your cross-scene setups. If scene B requires scene A to land, write that down. The AI can manage the bookkeeping. You manage the story.
4. Ship formats, not files. One source, five outputs. Don't hand someone a Google Doc. Hand them a PDF that looks like a book, because it is one.
5. The AI is infrastructure. Not a co-author. Not a ghostwriter. Infrastructure. Plumbing. It's the printing press, not the writer. Treat it accordingly.

← Back to case studies · Read the book