Writing a Memoir in 20 Days

55 scenes. 183 jokes. 4 voices. ? meanings of gross.

← All case studies

The problem

Bill's father died on December 23, 2025. Bill needed to write something true. Not a eulogy. Not a blog post. A literary memoir — three voice registers, a non-linear timeline, a joke dependency graph, and an undecidable sentence constructed so that no count of its meanings could be correct. 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 three voice registers — Cormac McCarthy for the lyrical passages, Charlie Murphy for the comedy, David Mamet for the machine-gun dialogue — and decided which voice belonged to which scene.
• Invented the "I am [age]" temporal device, borrowed from Dr. Manhattan in Watchmen: the narrator exists at all ages simultaneously. Time isn't linear. Memory isn't sequential. The structure mirrors how a mind actually works.
• Constructed an undecidable sentence — "gross" — where the number of valid meanings cannot be resolved. This isn't a pun. It's a formal property. His professor Alonzo Church proved in 1936 that some functions have no computable output. Bill built one into a joke.
• Decided the reading order. Scene 13, "Rosebud," opens the book. The chronological order and the reading order are completely different. The non-linearity is the point.
• 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.

The joke dependency graph

183 jokes. Many of them are Chekhov's guns — a line planted in scene 4 that pays off in scene 38. Move a scene and you break a joke. The AI tracked 35 cross-scene dependencies so that restructuring the reading order never orphaned a setup or detonated a punchline too early.

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, analysis, and the Church method

Front matter includes a half-title, title page with lambda symbol, copyright page, and dedication. The copyright page contains the formal expression for the undecidable sentence:

λgross.(¬∃n : |gross| = n)

"Given the word 'gross,' there is no number that equals the count of its meanings." That's the thesis of the book in one line of symbolic logic. If those symbols look like hieroglyphics, the false termination case study walks through every one of them in plain English.

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, submitted to two awards, distributed to 30+ people, and currently under professional review.

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 dependencies. If scene B requires scene A to land, write that down. The AI can manage the graph. 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