He Who Fights With Monsters · a Young Adventurer Edition

'99 Napster,
but for books
& a literal child

How I built a system that adapts a lit-RPG for my eight-year-old — and the instrumentation to provide the proof, page by page.

Vincent Baskerville (the dad in question) · B&Co Labs · linkedin.com/in/vincebaskerville

Default title-cover · claim-first opener (ENG-2.15) — Variants keep source covers.

00 · The whole thing, one slide

Not a cleverer prompt. A pipeline.

layers built before a single chapter

8 (+2)

gauntlet steps every chapter runs

signals screening every word

The problem: Reading an adult lit-RPG to my eight-year-old meant live-translating every page — night after night, until I stopped.

The obvious move: “summarize this, make it simpler, bring it down to a fourth-grade reading level.” I didn't do that — that path fails in ways that look like success.

The build: A governed eight-layer adaptation pipeline. AI was labor inside it, not the product — my judgment stayed load-bearing the whole way.

The receipt: The gate missed on a real kid, on a real word. The system caught it — and turned the miss into a permanent test.

Form ws9-slides-open · catalog template copy (ENG-2.15) · Examples keep source verbatim

The heads-up

What's in here.

The bedtime problem

Reading an adult lit-RPG to an eight-year-old — and the obvious AI move I didn't make

The build

Eight layers, a ten-step gauntlet, and AI as labor inside the rails

The gate that missed

A real kid, a real word — and the miss that became a permanent test

The other half

Orchestration is the new skill. Verification is the half almost nobody builds

What you can take

The shape, the receipts — and yes, something you can download

Spec constraint-set · non-goals and protocol — not an options matrix.

The bedtime problem.

Reading an adult lit-RPG to an eight-year-old — and the obvious AI move I didn't make.

Talk-stage talk-instrument · the answer as a narrative job (T-1237) — one claim, one line of support.

01 · The bedtime problem

Reading to the offspring at night.

One of the things I really enjoy — like a lot of parents — is reading to the offspring at night. My youngest — offspring #2, my son, he's eight — can't take He Who Fights With Monsters at adult level yet, so I was translating it out loud in real time. Swapping the hard words & difficult adult concepts on the fly. Night after night, that's tough. I eventually stopped. His excitement didn't. That difficulty is what started this fun little side project.

Foundation L2 · title + content region (Keynote/PPT commodity job).

01 · The bedtime problem

“summarize this, make it simpler, bring it down to a fourth-grade reading level.”

The obvious move — one or two sentences. Done.

Talk-stage talk-instrument · the answer as a narrative job (T-1237) — one claim, one line of support.

01 · The bedtime problem

Set up to fail in ways that look like success.

Me, I didn't do that. Having produced & built many AI tools, apps, processes — I know that path. Some jobs can survive that loop. This one couldn't.

Wrong task

Summary? Simplification?

A one-line prompt doesn't know whether you wanted a summary, a simplification, or a faithful adaptation that keeps the same story.

Invented success

It fills gaps. It sounds confident.

What you get is a clean paragraph that isn't the desired output — the model invents the "success" criteria.

No check

You, reading, hoping

Unless you already built checks — evals, a canon, a way to score "did this pass" — hours burn while the context window quietly degrades.

Talk-stage talk-signals · three equal-weight signals (T-1237) — exactly three, one idea each.

The build.

So instead of a prompt, I built a governed adaptation pipeline — with verification baked in — before any adapted chapters ran through it for him.

Talk-stage talk-instrument · the answer as a narrative job (T-1237) — one claim, one line of support.

02 · The build

What I actually built

the-adaptation-system/

young-adventurer-edition/

├─guiding-principle/preserve function, not form — all-new sentences

├─rulebooks/ (6)vision · style · characters · world · education · workbook

├─architecture/story arc · chapter blueprint · vocabulary schedule

├─difficulty-checks/← 1 of 8 — per-word verdicts for this one reader

├─continuity/so book two can’t contradict book one

├─book-plan/a ~265,000-word source → two books, short-to-full ramp

├─workbook/the Field Log — expedition gear, not homework

└─prototype/live — he can actually open & use it

Form ws9-slides-open · catalog template copy (ENG-2.15) · Examples keep source verbatim

02 · The build

Eight layers, before a single chapter.

1Guiding principle

Preserve Function, Not Form — same story, same arc, all-new sentences. Never a summary.

2Rulebooks

Six, loaded before any adaptation runs — Vision · Style · Characters · World · Education · Workbook.

3Architecture

Structure before sentences — story-arc outline, chapter blueprint, vocabulary & focus-skill schedule.

4Difficulty checks

The reader-bank + grade floor — a per-word verdict, with a personal override loop when the tools miss.

5Continuity tracking

A ledger tracking every invented detail — so book two can't contradict book one.

6Book plan

How the ~265,000-word source navigates into two books — a deliberate short-to-full length ramp.

7Companion workbook

The Field Log — the same difficulty judgment, pointed the other direction.

8Live prototype

Working — he can actually open & use it.

Form ws9-slides-open · catalog template copy (ENG-2.15) · Examples keep source verbatim

02 · The build

Every chapter runs the gauntlet — and it loops.

□

Blueprint → original scene

From the architecture, not the source page — function preserved, form new; the chapter's planned stretch words seeded on purpose; in-world panels rendered.

steps 01–04

The gate — Difficulty checks

The difficulty screen: zero walls; swaps re-screened.

05 · the gate

✓

Presence & continuity checks

Grep the chapter — did the planned words actually land? Reconcile against the Ledger.

steps 06–07

✎

Editorial pass, at each arc close

Ten fixed lenses: readability · pacing · voice · read-aloud · coverage …

step 08

↻

Read-test → bank

The built-in (+2) — there just in case, by design: the in-person read-test banks the rare real miss as a permanent override, and the chapter re-screens. Designed-in iteration, not a patch.

09–10 · the (+2)

10 · re-screen — until it converges toward the mechanical

Form ws9-slides-open · catalog template copy (ENG-2.15) · Examples keep source verbatim

02 · The build

AI was labor inside that pipeline — not the product.

That distinction is the difference between the due diligence of what I built vs a cheap 1 or 2 prompt.

Default solution-frame · locked answer after problem-frame — same narrative job, not a product claim.

The gate that missed.

A real kid, a real word — and the miss that became a permanent test.

Talk-stage talk-instrument · the answer as a narrative job (T-1237) — one claim, one line of support.

03 · The gate that missed

I did the homework here.

An interesting failure lived in the vocabulary difficulty checks — layer four of eight. Not "the whole pipeline was wrong," but a data-source gap inside a gate that already had checks. I built a coverage map against Georgia's Grade-4 English Language Arts standards — narrative crosswalk, stretch-word selection rules, how the story and the Field Log teach against the state's expectations. That is an alignment layer, not a word filter.

Foundation L2 · title + content region (Keynote/PPT commodity job).

03 · The gate that missed

Length isn't difficulty.

Frequency

how common in print

wordfreq Zipf scores — how common a word is in print.

Concreteness

how picturable

Brysbaert concreteness ratings — a thing you can see beats an idea you can't.

Acquisition age

how early kids learn it

Kuperman age-of-acquisition norms — where a value exists.

fused into one reader-bank — with a Flesch–Kincaid Grade-4 floor riding along

Form ws9-slides-open · catalog template copy (ENG-2.15) · Examples keep source verbatim

03 · The gate that missed

Two measures. Two completely different failures.

GRADE-4 LINE

← EASIERHARDER →

The frequency screen

chandelier

rated easy — really a wall

The 1948 familiar-word list

rated too hard — really easy

hamster · maze

the same scale  ·  missed two opposite ways

Form ws9-slides-open · catalog template copy (ENG-2.15) · Examples keep source verbatim

03 · The gate that missed

Never an afterthought.

ONE DIFFICULTY JUDGMENT

the reader-bank

INTO THE BOOK

The too-hard words get handled, so the story stays readable for him.

the adaptation

INTO THE FIELD LOG

Those same words become collectible loot in his Word Hoard.

the companion

not a checkpoint  ·  the same call feeds the book and the workbook

Form ws9-slides-open · catalog template copy (ENG-2.15) · Examples keep source verbatim

03 · The gate that missed

He hit the word chandelier — and stopped cold.

chandelier

And that was not the only word — there were a number he struggled with that were not even set as stretch words.

Form ws9-slides-open · catalog template copy (ENG-2.15) · Examples keep source verbatim

03 · The gate that missed

Why did my gate pass this?

My first thought was not aww. My first thought was the operator question.

This gate already had checks — and I’d done the homework: an alignment layer mapped to Georgia’s grade-4 standards, and per-word difficulty screens over the reader-bank. A chapter cleared all of it — with a wall still in it.

what the instruments said

✓ The standards crosswalk

alignment with Georgia’s grade-4 ELA expectations — in place

✓ The difficulty screens

every word checked against the reader-bank + grade floor — in place

✓ The chapter

cleared

↓

what reality said

× He stopped cold — struggling through words every screen had cleared.

the instruments said one thing — reality said another

Form ws9-slides-open · catalog template copy (ENG-2.15) · Examples keep source verbatim

03 · The gate that missed

The gap was in the instruments.

I went and looked — and the answer was almost funny: the databases had no read. The tools I’d trusted had no data on the exact words that were hardest for a real eight-year-old. And the gap wasn’t specific to my kid. It was in the instruments — a common failure mode when you lean on external corpora: coverage holes, wrong signal for the job, confidence without a verdict.

word	in print	to decode	age learned
chandelier	common	brutal	— no data —
peculiar	common	hard	— no data —
aching	common	hard	— no data —

Form ws9-slides-open · catalog template copy (ENG-2.15) · Examples keep source verbatim

03 · The gate that missed

So of course I went & built an add-on.

Frequency

how common in print

Concreteness

how picturable

Acquisition age

how early kids learn it

+ the floor — Flesch–Kincaid

The classic readability measure rides along as a ~grade-4 floor beneath all three signals.

+ the add-on — a personal override loop · the layer I had to build

Every real miss from reading together goes back in as a permanent verdict:

never-slipconfirmed-easydeep-read

all of it fuses into ↓

the reader-bank≈1,580 words on one screen — every word carries a verdict before a chapter reaches him.

easystretchwall

Form ws9-slides-open · catalog template copy (ENG-2.15) · Examples keep source verbatim

03 · The gate that missed

The reader-bank — and the add-on I had to build.

every word carries a verdict before a chapter reaches him.

Form ws9-slides-open · catalog template copy (ENG-2.15) · Examples keep source verbatim

03 · The gate that missed

“Looks right” isn’t the finish line.

Form ws9-slides-open · catalog template copy (ENG-2.15) · Examples keep source verbatim

03 · The gate that missed

That miss was a good catch.

It was also MVP iteration: find the hole in the instruments, tighten the eval, make the next miss cheaper. Important. Not the whole story of why this build worked.

Talk-stage talk-instrument · the answer as a narrative job (T-1237) — one claim, one line of support.

The other half.

Orchestration is the new skill. Verification is the half almost nobody builds.

Talk-stage talk-instrument · the answer as a narrative job (T-1237) — one claim, one line of support.

04 · The other half

Orchestration was only half the job.

There's a line going around that prompt engineering is dead and orchestration is the new skill. I've been writing in that lane myself — a few LinkedIn posts on why the prompt is just a step, and why the real work is the system around it.
Wiring models into a governed, multi-stage workflow beats firing one clever prompt.
But orchestration is only half the job.

Default exec-summary · job-shaped spine compress (ENG-2.15) — claim + stacked supports, not an agenda.

04 · The other half

The other half is verification.

Building the check the model won't hand you on its own. (In shop language: an evaluation layer — automated scoring where you can, human-labeled ground truth where you can't, and a loop that turns a miss into a permanent test.)
Ask a model and it'll happily tell you the book is now grade-appropriate; it won't volunteer that it screened chandelier as easy because it had no data on the word.
You find that out only if you built the thing that goes looking — and then trusted your own read over the tool's when they disagreed.

Default exec-summary · job-shaped spine compress (ENG-2.15) — claim + stacked supports, not an agenda.

04 · The other half

That's not a bigger prompt. That's a different discipline.

Talk-stage talk-instrument · the answer as a narrative job (T-1237) — one claim, one line of support.

04 · The other half

I know what this sounds like. Human-in-the-loop.

That is not the claim I’m making.

Operator judgment isn’t a safety net under the system. It’s part of how the system gets built and steered.

Form ws9-slides-open · catalog template copy (ENG-2.15) · Examples keep source verbatim

04 · The other half

Integral to the design — brick by brick.

The eight layers. The six rulebooks. The architecture before any chapter ran. The ten-step gauntlet. The Field Log as a second output of the same judgment. Strategy, constraints, done-means — me in the loop the whole way, on purpose.
I'm not building myself out of the process. I'm building a process where my judgment is load-bearing — continuously applied while the machine does labor inside the rails. (See: the judgment premium.)

Default exec-summary · job-shaped spine compress (ENG-2.15) — claim + stacked supports, not an agenda.

04 · The other half

And here's the delicate part.

The chandelier part of the story is a vivid example of an instrument miss and an override loop. It's not the strongest Judgment Premium proof. That miss is what good teams do in MVP. The premium in this piece is the holistic build: the layers, the gauntlet, the companion, the decisions, the actual quality of the product & how fast it was built the whole way through.

Talk-stage talk-instrument · the answer as a narrative job (T-1237) — one claim, one line of support.

04 · The other half

Two things, one judgment.

If you need proof this wasn't a one-prompt job, look at the second deliverable. The same difficulty judgment that decides what can stay in the adaptation also feeds the companion workbook — the Field Log — as collectible vocabulary. Loot. A number of different activity types.

Foundation L2 · title + content region (Keynote/PPT commodity job).

04 · The other half

The same judgment, pointed in both directions.

The Field Log is more than vocabulary — a purpose-built companion, never an afterthought. One judgment shapes the adaptation and builds the workbook against Georgia’s full ELA span: comprehension, language structure, communication — not just words.

the adaptation

Five stretch words locked into every chapter — bold and italicized, on purpose — chosen, scheduled, and scaling up; walls are handled. Thirteen source books in the long plan.

the Field Log

A standards-aligned adventure log, not a vocab list: comprehension, language structure, scene recreation, discussion-only sections, the Word Hoard — experience points, not grades.

Georgia standards — the execution model underneath the fun

grade 4grade 6

deliberately, across the series arc

Form ws9-slides-open · catalog template copy (ENG-2.15) · Examples keep source verbatim

What you can take.

The shape, the receipts — and yes, something you can download.

Talk-stage talk-instrument · the answer as a narrative job (T-1237) — one claim, one line of support.

05 · What you can take

Both are real and both are live.

The adaptation — He Who Fights With Monsters: Young Adventurer Edition — a personal-use adaptation of a series I love, built for exactly one reader in my house, nothing for sale. The Field Log is my original companion. Read / download at labs.baskervilleandco.com/do/hwfwm/book_1 · Field Log at /do/hwfwm/field-log.

Spec closing-cta · exactly one ask, stated as an action.

05 · What you can take

~1,580words in the bank — roughly 1,328 easy, 138 stretch, 87 walls

14chapters in the first arc

56planned stretch slots — zero walls in the target set

90+corrections banked — each one continues to hold

For the record — a real machine, not a gloss.

Spec stat-kpi · up to four captioned numbers — numerals are the largest type on the slide; labels share equal columns (no crushed max-width).

05 · What you can take

What you can take from the shape.

what mine looks like
what you take

VestaOS

→

an orchestration layer — rails already in place before the problem shows up

my virtual team (the Hounds)

→

a team shape around the work — AI agents doing labor inside the rails

the gauntlet + the reader-bank

→

a governed pipeline with a verification loop — misses become permanent checks

my private repo

not applicable

Mine is focused on my use case — this is the universal cut.
A similar shape, malleable — point it at your problem.

Form ws9-slides-open · catalog template copy (ENG-2.15) · Examples keep source verbatim

05 · What you can take

The Governed Feature Brief.

What I built for the book was tuned to my use case — so I cut a universal version. The Governed Feature Brief makes you name what done means, how it can look right but be wrong, and the check that can catch it — before the clever prompt, before the demo gets trusted.

SKILL.md /governed-brief

name: governed-feature-brief description: the brief before the prompt — done-means, failure modes, a check that can fail, and where overrides get saved.

type /governed-brief in any chat — it takes a paste, or interviews you

runs where you already work — one, or all three

✓ Cursor a skill in your editor

✓ Claude a skill in your project

✓ ChatGPT instructions in your GPT

governed_brief.md

One page. Four guts. No hiding.

Done-means

what done is — and what it is not

Failure modes

three ways it can look right but be wrong

A check that can fail

one test that can return FAIL — before trust

The override slot

where a human miss becomes a permanent rule

green only when all four guts exist · yellow lists exactly what’s missing · red means you can’t govern this yet — don’t build.

Form ws9-slides-open · catalog template copy (ENG-2.15) · Examples keep source verbatim

Don't start with the clever prompt.

Start with how you'll know it actually did the thing.

If you're an IC

PMs and UXers doing the work included

Pick one feature that sounds easy to prompt. Write the failure modes first: wrong task level, no eval, hallucinated attributes. Land a mini-gauntlet: inputs → constraints → scorecard → override table. A demo isn't evidence until that scorecard exists.

If you're a leader with a team

one feature or a small set of AI bets

Name what's in flight; pre-mortem what bites you first. For each: a source of truth, a check that can fail, and a named override owner — not the team. Demo ≠ evidence.

If you own the business

one customer-facing promise

Pick one promise AI is supposed to keep. Map what's required before real customers — canon, checks, a second output that proves the judgment is reusable — not a single chat thread.

Talk-stage talk-signals · three equal-weight signals (T-1237) — exactly three, one idea each.

05 · What you can take

So — what actually landed?

A governed eight-layer pipeline — and an eight-step gauntlet (+2) every chapter runs

A live adaptation for one reader in my house

A companion Field Log built as an adventure log

Named instruments, a coverage hole found on purpose, and an override loop that made the next miss cheaper

AI doing labor inside the rails — not standing in for the build

A few days on the side — not a flex about hustle, a signal about leverage: the operating system was already in place

Spec constraint-set · non-goals and protocol — not an options matrix.

You don't need a cleverer prompt. You need a system you can move fast in — and a way to know the work actually held.

The models are getting better. The datasets are getting better. The tooling is getting better. Good. Use them. Wire them. Score them. Build the second half — validation — and stay integral where judgment has to be a person.

Talk-stage talk-close · the one ask (T-1237) — the button, stated once.

/fin

Default title-cover · claim-first opener (ENG-2.15) — Variants keep source covers.

1 / 42

'99 Napster,but for books& a literal child

Not a cleverer prompt. A pipeline.

What's in here.

The bedtime problem

The build

The gate that missed

The other half

What you can take

The bedtime problem.

Reading to the offspring at night.

“summarize this, make it simpler, bring it down to a fourth-grade reading level.”

Set up to fail in ways that look like success.

Summary? Simplification?

It fills gaps. It sounds confident.

You, reading, hoping

The build.

What I actually built

Eight layers, before a single chapter.

Every chapter runs the gauntlet — and it loops.

AI was labor inside that pipeline — not the product.

The gate that missed.

I did the homework here.

Length isn't difficulty.

Two measures. Two completely different failures.

Never an afterthought.

He hit the word chandelier — and stopped cold.

Why did my gate pass this?

The gap was in the instruments.

So of course I went & built an add-on.

The reader-bank — and the add-on I had to build.

“Looks right” isn’t the finish line.

That miss was a good catch.

The other half.

Orchestration was only half the job.

The other half is verification.

That's not a bigger prompt. That's a different discipline.

I know what this sounds like. Human-in-the-loop.

Integral to the design — brick by brick.

And here's the delicate part.

Two things, one judgment.

The same judgment, pointed in both directions.

What you can take.

Both are real and both are live.

For the record — a real machine, not a gloss.

What you can take from the shape.

The Governed Feature Brief.

Don't start with the clever prompt.

PMs and UXers doing the work included

one feature or a small set of AI bets

one customer-facing promise

So — what actually landed?

A governed eight-layer pipeline — and an eight-step gauntlet (+2) every chapter runs

A live adaptation for one reader in my house

A companion Field Log built as an adventure log

Named instruments, a coverage hole found on purpose, and an override loop that made the next miss cheaper

AI doing labor inside the rails — not standing in for the build

A few days on the side — not a flex about hustle, a signal about leverage: the operating system was already in place

You don't need a cleverer prompt. You need a system you can move fast in — and a way to know the work actually held.

/fin

'99 Napster,
but for books
& a literal child