Language & Structure
The Architecture Playbook · how things get arranged, before they're styled or written
Foundation
The Living Design System governs how things look. The Copy Playbook governs how things read. Neither one documents how things are arranged, and arrangement is the layer both of them have been standing on the entire time. Twelve styleguide pages, thirteen copy pages, dual spec layers, coined names, cross-references: that's information architecture, and so far it's been done on instinct. This page names the discipline and its rules.
Adapted from Abby Covert, How to Make Sense of Any Mess. Translated with your examples; nothing reproduced. Credited as inspiration. Note the deliberate exception to the renaming rule below; it's a decision, not an oversight.
00
Why three words on this page aren't renamed
The vocabulary rule normally says: rename it or don't use it. This page breaks that rule three times on purpose, so here's the reasoning before someone tidies it away.
Documented exception
Ontology, taxonomy, and controlled vocabulary stay as they are. The renaming rule exists to stop a borrowed framework arriving with someone else's branding welded on. These three aren't anyone's coinages; they're the working terms of an actual discipline with a century behind it. Inventing a private synonym for "taxonomy" wouldn't demonstrate ownership; it would demonstrate not knowing the word, and that costs more than it buys.
The rule applies in full to everything built on top of them. The Three Fogs, the Landing Words, the Say List, the Never List, the Nesting Test, Sorting Is Arguing, all yours, all named here. Use the field's words for the field's concepts. Use your own for your own.
01
The Three Fogs — every confusion is one of these
Confusing information has three causes. Only three. Naming which one you're looking at matters because they demand opposite fixes, and the standard mistake is treating the third as the second and adding more words to a page that already had too many.
Too much
Everything's present and nothing's findable. The reader can't tell what matters.
FixCut, or stage it across a sequence. Not smaller type.
Not enough
A gap the reader fills with a guess, and the guess is usually wrong and usually unflattering.
FixSay the missing thing. The Writer page's unstated post-trial price is exactly this.
Not the right
Volume is fine, content is wrong for this reader at this moment. The hardest one to see.
FixReplace, don't add. Adding here makes it Fog One as well.
The common error: diagnosing "not the right" as "not enough." The page feels thin, so more gets added, and now it's wrong and crowded. When something isn't landing, ask whether it's the wrong thing before assuming it's too little.
The Three Fogs
Too much · not enough · not the right
Three causes of confusion. Three different fixes.
Name the fog before touching the page. The fixes are opposites.
Yours: this is the audit lens. The Writer findings sorted cleanly into Fog Two.
02
The Landing Words — set them before you build
Pick the adjectives you want people reaching for about the work. Then, and this is the part that does the work, pick the ones you're content not to earn. Three rules make this useful instead of decorative, and all three are counterintuitive.
✓ What I want them to say
precisehonestusablecraftedopinionated
Fine if they never say
comprehensiveapproachablenovelfastaffordable
The three rules. The second list can't be opposites of the first; "imprecise, dishonest" teaches you nothing and takes ten seconds. It has to contain genuinely desirable things you're choosing not to chase. Nothing negative goes in either list; keep both neutral enough that a stranger couldn't tell which is which. And the two can't contradict each other. Done properly this is a decision rule: when "comprehensive" and "precise" pull against each other, you already know which one wins.
The Landing Words
Two lists, no opposites
Set before building. Resolves the judgement calls later.
List two must be desirable things you're declining, not the negation of list one.
Nothing negative. A stranger shouldn't be able to tell which list is which.
Yours: sits upstream of the six-question brief. Words first, then the brief, then the writing.
03
Ontology — the thing you already do
Deciding that a word carries a specific meaning inside your context is called ontology, and it's what you've been doing since the first coinage. Naming it matters because it converts an instinct into a practice with an obligation attached.
Ontological decisions you've already made
memory ladder — not a metaphor for memory in general. Four specific storage stages, in order, with a rule about when each earns its place.
earned complexity — complexity that survived a justification, as opposed to complexity that was adopted because it was fashionable.
slop — reframed. Not a model failure. A prompt failure.
the Floor — not "the setup." The specific part of a story where stakes get installed, weighted heavier than the event.
The obligation that comes with itIf a word means something specific here, the definition has to be findable. A coined term with no definition anywhere isn't ownership; it's Fog Two, and the reader fills the gap with a guess.
Ontology
This word means this, here
The formal name for what every coinage is doing.
Coin a term, owe a definition. Findable, or it's a fog.
Test it on someone outside before you commit. You'll think it's clear and it usually isn't yet.
Yours: the vocabulary-fight trigger is an ontological decision made under pressure. This is the calm version.
04
The Say List and the Never List
Two halves of one artifact. The Say List is the words you use and how you use them; the Never List is the words you refuse and why. The second half is the powerful one and almost nobody builds it. The rule that makes it hold: every entry carries a reason, because a ban with no reason gets quietly reversed by the next person who has a good use for the word.
✓ The Say List
agentthe thing that acts. Not "AI" generically, not "assistant."
specthe written, enforceable version of a rule. Not "guidelines."
playbooka living reference that is also machine-readable. Not "guide," not "doc."
slopkept, and redefined. A prompt failure, not a model failure.
earnedthe load-bearing modifier. Complexity, links, attention, all earned, never assumed.
✕ The Never List — with reasons
journeya passenger word. Names something passive that should be active. Also an Impostor.
leverage (as a verb) consultant filler. "Use" is the word and it's shorter.
seamlessunfalsifiable. Says nothing a reader can check.
verythe modifier that admits the adjective was weak. Pick a better adjective.
delve, tapestry, landscapemachine tells. Detailed treatment lives in the Humanizer; the ban lives here.
limited time onlyrefused on principle, not on style. See the Levers.
One artifact, not four. These rules currently live scattered across
the Humanizer's tells,
the voice rules, and
the Levers refusal. Consolidating them here doesn't replace those pages; the Humanizer stays the detailed source on machine tells. It gives the rules a single enforceable index with a reason per line, which is the difference between a standard and a preference.
Say · Never
The controlled vocabulary
Words used, words refused, reasons attached.
Every Never entry needs a reason. Reasonless bans get reversed.
Living document. Add the word the moment you catch yourself deleting it twice.
Yours: refusals are structural here. Same commitment, applied to vocabulary.
05
The Nesting Test — jargon hiding inside jargon
Write the definition. Underline every word inside it that itself needs defining for this reader. Define those. Repeat until you land on words the reader already owns. It's tedious for about four minutes and it catches the most common failure in technical writing: a definition that's correct and useless.
Round 1 — the definition as first written
A
memory ladder is a staged
persistence architecture where each
rung adds
retrieval capability at the cost of
operational overhead.
Round 2 — four underlined terms, all of which need definingEvery one of those is doing work the reader can't check. This definition is accurate and it explains nothing to anyone who didn't already know.
Round 3 — rewritten to words the reader already ownsA memory ladder is four ways to store what your agent knows, from simplest to most powerful: files, a database, search over documents, then a graph. Each step remembers more and costs more to run. Start at the bottom. Climb only when the one below stops working.
The stopping condition is their vocabulary, not yours. You stop when the words in the definition are ones the reader already had before they met you. Anything else is a definition written for someone who didn't need it.
The Nesting Test
Define, underline, recurse
Until every word is one the reader already owns.
Catches coined terms resting on other coined terms that were never defined.
Yours: run it on every new coinage before it ships. It's the ontology obligation, made checkable.
06
Nouns → Verbs → Specs
The bridge from naming a thing to enforcing it, and the formal description of what your embedded comment specs have been doing all along. List the nouns. List what can be done to each. Combine them, and every sentence you get is a requirement an agent can act on and a human can check.
Nouns — what exists
playbook
page
job
spec panel
coinage
refusal
Verbs — what can be done
a page can be added to a playbook
a job can be numbered
a coinage can be defined
a refusal can be reasoned
a spec panel can be paired with a demo
Specs — the enforceable line
Every job pairs a demo with a spec panel, or is a full-width calibration block.
Every coinage carries a findable definition.
Every refusal carries a stated reason.
Every page carries a source note naming what it adapted.
Every page carries one honest caveat.
This is why the dual layer works. The visible panel carries the noun so a human can read it. The embedded comment carries noun plus verb plus constraint so an agent can act on it. They aren't two formats of the same content; they're two levels of the same specification, and this is the discipline that produces them.
Nouns → Verbs → Specs
Naming becomes enforcement
A noun plus a verb plus a constraint is a requirement.
If you can't write the verb, the noun isn't defined well enough yet.
Yours: the five specs on the right are now the standing rules for every page in every playbook.
07
Sorting Is Arguing
Grouping isn't neutral housekeeping. Every arrangement is a claim about how things relate, and readers absorb the claim whether or not you meant to make one. Worth running on your own system, because you've been making arguments for months.
The arrangement
- Twelve styleguide pages under one Living Brand Playbook index
- Copy split out from design as its own playbook
- Architecture added as a third peer rather than a section
- Every page carrying its own honest caveat
The argument it makes
- Design is governed, not expressive. Taste is written down and checkable.
- Words and visuals are separable disciplines with their own rules.
- Arrangement is a peer of both, not a byproduct of either.
- A system that states its limits is more trustworthy than one that doesn't.
All four are defensible. The point is that they're arguments, and they should be made on purpose.
The practical version: before you add a page, ask what its placement claims. A page filed under "Copy" argues it's about words. The Conversation page tested that; it's about replies rather than writing, and filing it in Copy argues that everything reaching a reader is copy's business. That's a real claim and it's the one intended.
Sorting Is Arguing
Taxonomy is rhetoric
Every grouping is a claim. Readers absorb it either way.
Before filing something, ask what the filing claims.
Exactness costs flexibility. A very precise taxonomy is harder to grow into.
Yours: this is the page that governs the index. The index is an argument, so build it like one.
08
The Query Layer — addressable without being read
Nouns → Verbs → Specs named two layers already living on every job here: a visible panel for a human, an embedded comment for whoever edits next. This is the third layer, added system-wide across all four playbooks: a fixed set of data-* attributes that let an agent find and pull exactly the section — or, one level down, exactly the single card — it needs without opening the file at all.
Page level — every <html> tag
data-playbook="living-copy" — which of the four playbooks
data-page="honest-line" — which page inside it
data-doc-type="guide" — hub, guide, or styleguide. Closed vocabulary.
data-summary="…" — one sentence, what the page is for
data-last-updated-at="2026-08-05" — bumped by hand at edit time
Section level — every top-level content containerdata-role="section" — identical string everywhere, the query hook that finds every section in any playbook in one pass. Plus its own data-summary — one sentence, what's specifically in that section, not the page.
Component level — one sub-item inside a multi-item sectiondata-role="component" — same hook, one level down. Only used where a section genuinely holds more than one distinct card: a pattern in
the Honest Line's 15, a nav tile on a hub page, one column of a compare grid. Plus its own
id and
data-summary.
The failure mode to avoid: tagging a heading div that doesn't contain its section's body, or a content div that doesn't contain its heading — at either level. Both pass a casual glance. Both return an incomplete chunk to anything that queries them. The id has to sit on whatever element actually wraps the whole thing, and a component tag only earns its place when there's more than one sibling to tell apart.
The Query Layer
Addressable without reading
A third layer under the dual spec, built for a reader neither panel nor comment was written for: one that never opens the page.
Grep data-doc-type="styleguide" across the repo and get every styleguide page's path — zero reads.
Query [data-role="section"] in a browser and get every top-level section in the document, whatever visual class it happens to use.
Query [data-role="component"] and get every individually addressable card across the whole corpus — 103 of them as of 2026-08-05 — without walking into a single section first.
Yours: the ontology rule from job 03 applies here too — coin an attribute, owe it a definition. This job is that definition.
09
The other playbooks
This page arranges the other three. Until now it never linked to them — the one link this whole document's own argument says shouldn't be missing.