One approved source of truth the AI reads before it writes a word — what's true, how we sound, and what's approved — wrapped in rules it cannot break, doing the production work with the team.
When a company’s story changes — a rebrand, a launch, a new line — every letter, page, and phone script has to carry it accurately. The brain is the team's shared system for getting that done: it holds what's true, how we sound, and what's approved — and it does the production work alongside you. You talk to it like a colleague; it answers from the library, never from guesswork.
Every fact, price, claim, voice rule, and brand decision lives in one governed library. Ask anything — the answer comes back sourced. When something isn't known yet, it becomes a numbered question for the team instead of a guess in your copy.
You don't have to carry the rules in your head. The system loads them every session, scans every deliverable before handover, and refuses anything the rulebook forbids. Work at full speed; the guardrails catch the slips.
Ask for a letter, a page, a deck — it drafts from approved templates in our voice, saves to your Work Zone, and routes to approval. The big jobs, like the 1,200-file collateral rebrand, run as supervised agent fleets with a person approving every batch.
Every request runs the same seven-step line — no shortcuts for "quick" pieces. Position in the folder tree IS approval status, so nothing ships by accident.
Anyone in the team asks for a deliverable, a fact, a plan.
The prompt-enhancer scores the ask; thin requests get 2-3 sharp multiple-choice questions.
CLAUDE.md's routing table opens exactly the right rules and facts. Never crawls, never wanders.
Pricing from one file. Claims from one file. Trust order: Approved > Facts > Work Zone. Everything else is UNKNOWN.
Skills copy a gold-standard reference and adapt it. AIs diverge when designing, converge when adapting.
The pre-delivery scanner: banned terms, leak tripwires, dash policy, raw hex. Exit code, not opinion.
Work Zone → review queue → Jordan* → Avery*. Only a human promotes to Approved. Nothing is ever deleted.
Every layer, what it holds, and what it's waiting on. Cards link straight into the live documents; the burger menu opens the working views.
CLAUDE.md: 8 sections the AI loads every session — identity, load order, the routing table, trust order, ten hard rules, the output checklist, and the approval roster. A generated AGENTS.md mirror serves any other AI.
Guardrails (12 hard limits), master thinking-order, house style, the voice trio (brand voice, approved phrases, banned phrases with the internal/external message wall), 7 writing tones, 4 target audiences, 5 channel styles, and 6 output-type direction files that script entire builds.
Company overview, two authority files (pricing and claims — the only sources a number or credential may come from), the launch layer (Day-1 playbook, retention engine, media & PR plan), the decisions log, team + escalation tiers, competitors, and distilled meetings.
15 generation skills (decks, emails, microsites, price sheets, welcome kits, one-pagers…) each built to copy a gold-standard reference — activating at brand-lock. Plus the CEO presentation style, the prompt-enhancer front door, and the quality-control script suite.
Six named colours, each shipped with the contrast ratio it was measured at and the failures written down as laws — no red, no text-legal orange on light, no seventh value. One mark redrawn for every size it has to survive, five approved lockups, and four type faces with one duty each. Delivered as a standalone document the whole system re-renders from when a single value is repointed.
Every drafted launch piece filed verbatim from the GTM: CEO all-staff emails, customer letters (new accounts + existing accounts), the 5-email nurture sequence, crisis call scripts, the press release draft, rep talk tracks, vendor letters. All approval-gated, all slot-marked for open decisions.
Eight personal zones (Jordan*, Avery*, Sami*, Morgan*, Frankie*, Dana*, Gale*, Marley*), five simple folders each: My Space, Active Projects, Send for Approval, Save for Later, Archived. The review queue holds 6 launch drafts, every one fleet-verified line-by-line before human review. Approved/ holds canonical truth in 12 categories and fills only by promotion.
A mandatory changelog, the session work log, the setup-question ledger (57 tracked — the brain files questions instead of guessing, and the Questions page is answerable in place), the dated roadmap, and the rebrand migration plan with real counts.
The team Hub (11 pages + a branded tool suite, 10 people seeded in its data model) waits on launch day by design. Three strategic plans map Apex parity and the migration. A 5 GB Apex reference library feeds the rebrand — local-only, never uploaded.
The newest organ, installed Aug 8. The brain doesn't just write — it runs coordinated fleets of agents on work too wide for any single session, with independent verification wired into every run.
Wide jobs split into parallel workers — one per file, per draft, per partner kit. Every result is then attacked by independent verifier agents that never saw the worker's reasoning, across three lenses: rules, facts, and voice.
Every run counts its returns (a silent failure can't hide in a complete-looking report), anchors on signals that can't be argued with — the scanner's exit code, the authority files — and ends at a human gate. Fleets do the breadth; Jordan* and Avery* keep the judgment.
The 1,200-file Apex→Custom LLM rebrand, one partner kit per run: parallel piece-by-piece rebuild, triple-lens verify, Jordan* reviews one batch report.
STAGED · FIRES AT BRAND-LOCKThree fresh verifiers per draft plus the scanner anchor. Ran live Aug 8 — now a saved one-line command.
LIVE · 18/18 FIRST RUNNew strategy doc arrives → parallel checkers diff it against every fact area; conflicts become numbered questions, never silent overwrites.
READY · NEXT KNOWLEDGE DUMPThe brain's health check run cold: parallel probes on the four proof questions plus refusal tests that must refuse.
READYFour parallel certifiers before anything ships: twin freshness, scanner sweep, link audit, the rules grep gates.
READYFresh eyes diff every status surface against every other — including the rules files against the constitution itself.
READY · PERIODICEighteen independent verifiers swept the six launch drafts and found what serial review hadn't: a customer-facing "48-hour executive response" commitment that exists in no approved source, production notes sitting in shippable copy, and unapproved superlatives inherited from strategy docs. Same day, the engine audited the brain itself — and caught two stale rules and a broken check in the scanner, all fixed within hours. The system finds its own faults before the market can.
Every engagement runs the same dated line: ten working days from kickoff to finalization. This is the build sequence The Process describes, laid out on one calendar.
The narrative version of these ten days lives on Build & Training, and the full dated plan is the project plan document.
A skill is one packaged ability — a recipe the brain follows every time so the output is consistent. You don't type commands for these: open a session at the brain root, ask for the thing, and the constitution routes to the right skill. The eighteen generation skills are built and waiting — the day leadership locks the brand, their gold standards get built from the approved Brand OS and they all switch on at once.
Runs work too big for one sitting as coordinated agent fleets — fan out, verify with fresh eyes, synthesize one answer. Powers the QA fleet, the coming migration batches, and the brain's own audits.
"Run the qa-fleet workflow" · "Run this ingestion as a graph"Runs first and silently on every request; scores completeness and asks only real-option questions before any work starts.
Automatic — every request passes through it.The last gate before any handover: banned words, the rules tripwires, pricing sources, phone formats, dash police. Half deterministic script, half judgment. Exit code, not opinion.
Automatic on every deliverable · "Scan this file"The generators that build the machinery itself: the team site, the styled twins, the Asset Directory, AGENTS.md, the brain audit. Operator surface.
"Rebuild the guide" · "Regenerate the asset directory"Keeps every markdown file's styled html twin current: staleness scan, regeneration with the day's stamp, and the never-hand-edit-a-twin rule.
"Run twin-sync"Turns a team member's answer into filed fact with attribution, cleared placeholders, a checked-off register question, and a refreshed site. Also lives as the /file agent.
"File this answer: Q-24 — our locations are..."Builds a gate-verified preview bundle of the site when push-to-deploy is not available, so a change can be reviewed before it ships.
"Build a preview bundle"The standard library-cleanup process, ported from Apex's proven pattern: version ranking, archive-never-delete, card refreshes, promotion mechanics. Run on any Approved folder that has grown messy.
"Clean up the BioSummit folder"Cass Louis Kai*'s presentation language — density doctrine, photo treatments, badges, stat cards, stair-steps — overlaid on any deck bound for his desk. Ported from the proven Apex build.
"Build this deck in the Kai* style"Slide decks on the locked brand system, from the Brand OS deck theme.
"Build a deck about..." (after brand-lock)Intro, follow-up, delivery, and sequence emails in brand voice and skin.
"Draft the welcome email" (after brand-lock)One-page executive summaries, print-true, in the brand system.
"Summarize this for leadership" (after brand-lock)Standalone landing pages on brand tokens, ready for hosting.
"Build a landing page for..." (after brand-lock)A partner's full microsite — the migration's per-partner web presence.
"Build the Trellis microsite" (after brand-lock)Partner price sheets from the filed price lists, in the brand system, as HTML templates the brain can regenerate.
"Build the Fairline price sheet" (after brand-lock)Print-ready one-pagers: company, product, or program.
"One-pager on digital dentures" (after brand-lock)Step-by-step quick-start guides for partners and products.
"Quick-start for the scanner program" (after brand-lock)Partner RX lab slips, the working forms of the lab relationship.
"Build the lab slip for..." (after brand-lock)Social graphics plus captions across platform formats.
"Social set for the announcement" (after brand-lock)The partner welcome flyer — first-touch print for every onboarding.
"Welcome flyer for..." (after brand-lock)The checklist guide that tells a practice exactly what to send for each case type.
"What's-needed guide for full arch" (after brand-lock)Branded deck straight from the Brand OS deck theme tokens.
Part of the Brand OS treatment set (after brand-lock)Branded email set — intro, follow-up, delivery, newsletter, sequence.
Part of the Brand OS treatment set (after brand-lock)Print leave-behind or flyer, two pages max, on the brand system.
Part of the Brand OS treatment set (after brand-lock)Company or topic one-pager, print-true, from the Brand OS.
Part of the Brand OS treatment set (after brand-lock)Client-specific proposal from the Brand OS proposal treatment.
Part of the Brand OS treatment set (after brand-lock)Social graphics and captions in eight platform formats from the Brand OS.
Part of the Brand OS treatment set (after brand-lock)Every folder and file in the brain, mirrored on the Apex structure and coded by the signal pair: Landfall is built and ready, Meridian has an md explanation or sits in the Work Zone, and planned but not started carries no colour at all — nothing else gets one. Search it, browse it, open anything — every document opens its styled page.
Four phases from first contact to a trained team. The two-week calendar covers phase 3 day by day; this page is the whole line, in order.
Eleven sections; complete it incrementally — blanks become tracked questions, not blockers.
Every answer keyed to its destination file; an upload manifest; a completion percentage.
Price lists, decks, SOPs land in source-docs/ untouched — the brain quotes them, never rewrites them.
Every unanswered field becomes a numbered item with an owner on the Team Questions page.
Twenty collateral treatments, five chapters, zero hardcoded values — the starting point every time.
The client's palette, type, marks, and name replace the token table; everything re-renders.
Contrast measured, exclusions registered, leadership signs. PROPOSED until locked.
Charter, the six-zone tree, the registers, the constitution with its ten hard rules.
Fact ingestion, the two authority files (pricing, claims), the voice trio, audiences.
The scanner wired into every handover — banned terms, policy tripwires, slot checks.
Work Zone lifecycle, the Brand OS tokens wired, generation skills on gold standards.
The agent team staged on real gates; the QA fleet's first live run.
The brain answers its proof questions cold — and refuses the things it must refuse.
Three sessions: the tour, first drafts through the approval loop, the machinery.
The team owns the flow; the gates that pace everything after (brand-lock, launch, phase 2) take over.
The day-by-day version of phase 3 is the calendar; the narrative is Build & Training.
Tested against this brain. Every prompt below does what it says, including the three that prove the guardrails by refusing.
Everything that explains the system, runs the launch, and keeps the brain honest.
The cascade with dates and owners, the hour-by-hour run of show, crisis protocols, rumor control.
Open ↗ The Retention EngineZero attrition: coverage tiers, the four phases, field sales, the top 20%, win-back.
Open ↗ Media & PR PlanThe four phases, the publication and podcast targets with per-outlet talking points, the agencies.
Open ↗ Launch Copy LibraryThe drafted, send-ready letters, emails, scripts, and the press release: verbatim, approval-gated.
Open ↗The zones, the rules, the workflow: how to work with the brain day to day.
Open ↗ The ConstitutionCLAUDE.md: the contract the AI reads first. Hard rules, routing, trust order, governance.
Open ↗ Project Plan & StatusPrepared for Avery*: where we are, what's done, what the team owes, milestones ahead.
Open ↗The structure the AI navigates:
Every deliverable passes the automated gate: banned words, brand rules, the rules tripwires, dash police. Exit code, not opinion — hardened Aug 8 after the engine audited its own scanner.
Four questions prove the brain in 3 minutes, including two deliberate refusals. Last run Aug 8: PASS 4/4.
21 fresh-context verifiers sweep the review queue on demand: "run the qa-fleet workflow." First live run Aug 8.
Every change to the brain, dated and explained. The audit trail — now a tab of its own.
Open →The design system this demo is built in. Six colours, four faces, one mark redrawn at every size it has to survive. Every ratio below was measured on 2026-08-13, and the ones that fail are written down as laws rather than nudged away.
The complete standalone document: the design system demonstrating itself, component by component, with the usage rule written under each one.
Open ↗ How the brain uses the Brand OSThe plain-English explainer: what a token is, why the AI never hardcodes a colour, and how collateral inherits the system instead of copying it.
Open ↗ The explored directionsThe earlier drafts, kept for the record. In a real sprint each direction ships as a complete working system, not a mood board — which is what makes them comparable.
Open ↗ Material-concept prompts StagedAdditional directions, written and ready to run if a client wants range before locking.
Open ↗Three do the work: Midnight Green grounds everything, Paper White carries the type, True North is the only warm thing in the system. Two are reserved for signal and are never decoration. The sixth is what the material falls back to when it cannot render.
Primary · ground · ink. The one dark. Heroes, headers, footers, dark bands, and every line of body type on a light page. There is no second dark. White on it 15.8:1 ✓ · on white 15.8:1 ✓
Secondary · the needle. Named for the point it colours. Primary buttons, the eyebrow on dark, the live rule. Used once per view, twice at most. On Midnight 5.6:1 ✓ · on white 2.8:1 ✗
Base · card · knockout. True white, never off-white. Cards, sheets, knockout marks, and all type on dark grounds. On Midnight 15.8:1 ✓
Signal — arrived, shipped, passing. Land sighted: builds that pass, work that shipped. Never a decorative green. On Midnight 9.6:1 ✓ · on white 1.7:1 ✗
Signal — in progress, information. A line of longitude: the thing you are still travelling along. On Midnight 5.3:1 ✓ · on white 3.0:1, large only
Material fallback · link · label. What the material collapses to with no WebGL. Doubles as the link and eyebrow colour on light, and draws the bezel and grid inside the mark. On white 7.4:1 ✓ · on Midnight 2.1:1 ✗
5% Steel on white. The page beneath the cards, so white sheets have something to sit on.
8% Steel on white. Chips, code fills, quiet washes.
72% Midnight on white · 6.4:1 ✓. Secondary text.
65% Midnight on white · 5.1:1 ✓. Labels and metadata — the quietest legal level.
66% white on Midnight · 7.6:1 ✓. Dimmed type on dark — shown here on its own ground, because that is the only ground it has.
Every one of these is a mix against a named token, never a pasted hex. Repoint Midnight Green and the whole page re-renders correctly — that is the test this table exists to pass.
Never type on white at any size. On a light ground it is a fill, a rule, or a shape — nothing else. At 5.6:1 on Midnight Green it is legal for body text on dark, and that is the only place it gained. One True North button per view; a second primary means the first one was not primary.
A light-ground colour only. 7.4:1 on white makes it the link, the eyebrow and the label; 2.1:1 on Midnight Green makes it illegal as type on dark, where Paper White carries the reading instead. Links get a 1px underline, 3px clear of the baseline.
Landfall and Meridian only ever mean something: status, and fills in data — bars, nodes, stat tiles, never the axis and never the label. Neither appears in a lockup, a header, a button, or anything adjacent to the mark. On light the dot carries the colour; on dark the label may. In greyscale they separate at 0.59 and 0.30 against Midnight’s 0.02, which is why the pair can carry status without changing shape.
The palette has no error colour, because True North means action, not failure. An error is carried by a 2px Midnight Green border, the glyph, and a line of Plex Mono under the field. Focus is a 2px Surface Steel ring, never a glow. Depth is a hairline and a 1–2px lift; shadows exist only for genuinely floating surfaces.
Black 900, width axis pinned to 125 — Archivo’s blockier sibling from the same foundry, so the wordmark and the headlines share a skeleton without repeating each other. Never a headline, never a heading, never live text in production. Outline it and it becomes artwork.
500–900. Headlines, section heads, card titles, buttons. 800–900 for display, 700 for anything under 24px. Headlines set at -0.025em; sub-heads are Archivo 800, never bold body.
400–700. Everything a person reads for longer than a glance. Around sixteen pixels over a 1.55 leading, holding a 66-character measure, and it never competes with the figures beside it.
400–600. Eyebrows, figures, IDs, file names, code. Tabular figures always, so a column of numbers lines up and a paragraph full of percentages does not ripple.
A compass rose set over a gridded globe: the globe is the territory, the rose is the read on it. One point of that rose, north, is True North orange, and it is the only warm thing in the mark. The handoff between drawings is a measurement, not a preference. Clear space is 0.5 × icon height on all four sides, measured from the outer bezel, and drops to 0.3 in dense chrome — never below.
Never: rotate the mark · recolour the north needle · set the wordmark in Archivo, or in Saira at any width but 125 · put the icon to the right of the wordmark · scale the primary icon below 48 px instead of switching drawings · add a stroke, glow or shadow the mark did not ship with. Below 240 px the tagline is dropped, not shrunk; below 168 px the wordmark is dropped and the icon stands alone.
A skill does one thing; an agent chains skills into a finished outcome, and each one has exactly one thing it refuses to do. That last part is the product. Checkers check, makers make, librarians remember, operators run the machinery — and every one of them obeys the constitution: nothing external, nothing invented, nothing into Approved without a human.
Custom LLM Brain folder, type the slash command (it autocompletes), and describe what you need. The specialist runs with fresh eyes and reports back.Reads every register and returns the honest one-page where-are-we: done, in review, blocked and on whom, plus the one decision that unblocks the most.
/statusWill never: Round progress up to sound better. Blocked is blocked.
Generates the operating guide and the prompt gallery from the brain itself, then runs the four-question health gate at handoff. It writes the page from the confusion it actually watched happen.
/guide for the retention teamWill never: Call a handoff done on a failing acceptance test.
Owns the open-questions register: who owes what, since when, and what it holds up.
BLOCKED: The register stores owners as roles rather than people and carries no dates. That schema change comes first.
/chase everything owed by Avery*Will never: Answer on the owner’s behalf just to clear the queue.
Wakes standing sweeps on schedule and caps what a fleet may spend before it fans out.
/cadence weekly register sweepWill never: Raise its own cap.
Scores every incoming ask before any work starts. A thin one gets two or three sharp multiple-choice questions rather than a confidently wrong draft.
make me a one-pagerWill never: Let a vague ask through as though it were specific.
Runs the eleven-section intake. Every blank becomes a numbered question with an owner instead of a gap somebody fills in later from memory.
/intake for the Trellis accountWill never: Fill a blank itself.
Answers anything from the brain’s filed knowledge, with a file citation for every fact. It would rather hand you a visible blank than a plausible invention, and it drafts the question that would close the gap.
/facts what are our written retention commitments, exactly?Will never: Fill a gap with general industry knowledge. The brain’s files or UNKNOWN, nothing else.
Files a meeting the brain’s way: verbatim source stored, decisions and action items distilled, conflicts with the decisions log flagged rather than silently resolved.
/scribe [paste today’s call transcript]Will never: Record a decision nobody actually made.
Takes an arriving answer and lands it everywhere it belongs in one pass: the fact recorded with attribution, every placeholder it unlocks filled, the question checked off, the changelog line written.
/file Q-13 is answered: Avery* confirmed the type pairing todayWill never: File an answer whose source is unnamed.
Diffs an arriving document against every filed fact. Conflicts become numbered questions, never silent overwrites.
/ingest the updated services sheetWill never: Resolve a conflict by picking whichever document is newer.
Keeps the competitor and category picture current, sourced and dated.
/market who else sells into this category?Will never: Present an inference as a finding.
Checks anything visual or branded against the token system, the surface law, the mark rules and the name rule.
/brandcheck 05 Work Zone/review-queue/website/info-page-DRAFT.mdWill never: Redesign the piece or pick new colors. It reports.
Clones the master system, runs the swap protocol, and produces the full client system across all twenty treatments.
/brandos build the Custom LLM system from the approved directionWill never: Publish a system leadership has not signed.
Audits a finished brand system against the five-part contract before a brain is assembled on top of it.
BUILDABLE: It exists today as a document somebody runs by hand rather than a repeatable command. The packaging is the whole build.
/brandaudit the current systemWill never: Pass a system with unmeasured contrast.
Re-renders the collateral library into the new brand, in reviewed batches. Runs after the Migration Lead, never before it.
STAGED: Gated on brand-lock.
/migrate Trellis (after brand-lock)Will never: Run a batch nobody approved.
Routes a production ask to the right skill and the right gold standard.
/produce something for the Trellis welcome packWill never: Invent a format that has no reference.
Builds a new piece the brain’s way: right template, right tone, facts from the authority files, scanned clean, saved to your Work Zone.
/draft a one-page internal update on where the brand sprint standsWill never: Invent a fact, price, name or testimonial to fill a gap.
Rewrites any copy into an approved tone, showing before and after. Fixes banned words, hype and off-voice phrasing without touching the facts.
/revoice 05 Work Zone/Jordan/01-scratch-pad/partner-email.md into reassuring-advisorWill never: Alter a fact, price or commitment while rewording.
Builds presentations on the configurable executive overlay.
STAGED: Gated on brand-lock.
/deck the quarterly partner review (after brand-lock)Will never: Use a person’s presentation style outside its declared audience.
One-pagers, flyers, trifolds, price sheets, mailers and certificates.
STAGED: Gated on brand-lock.
/print the Trellis one-pager (after brand-lock)Will never: Ship a piece with an unresolved slot still in it.
Landing pages, microsites, blog treatments and share cards.
STAGED: Gated on brand-lock.
/page a landing page for the launch (after brand-lock)Will never: Publish outside the approval chain.
Frames, stat cards, quote cards and announcement sets.
STAGED: Gated on brand-lock.
/social an announcement set for the rebrand (after brand-lock)Will never: Post anything, anywhere. It produces files.
Owns the launch sequence and who hears what, in what order.
/comms plan the rebrand announcementWill never: Let an external piece go out before its internal one.
Multi-touch email sequences built from the approved library.
STAGED: Gated on brand-lock.
/nurture a five-touch sequence for lapsed accounts (after brand-lock)Will never: Send. It drafts only.
Launch kits: press draft, fact sheet, talk tracks and FAQ.
STAGED: Gated on launch.
/announce the Custom LLM launch (after the date is set)Will never: Publish before the date.
The pre-approved scripts for the day it goes wrong, one per scenario.
/crisis a shipment defect reaches a partnerWill never: Improvise a statement.
The one-page brief on any account: history, contacts, terms, collateral, open items, and who owns it.
STAGED: Gated on partner facts.
/brief Trellis (after partner facts land)Will never: Brief from memory rather than from the file, or pad with industry generalities.
Scoped proposals with every figure traced to the pricing authority.
STAGED: Gated on brand-lock.
/proposal BioSummit (after brand-lock)Will never: Quote a number that is not in the authority file.
Answers any pricing question from the single legal source, with the citation attached.
/price what is quotable on the standard tier?Will never: Quote a price from a meeting note.
Per-partner onboarding kits cut from one master architecture.
STAGED: Gated on brand-lock.
/kit BioSummit (after brand-lock)Will never: Ship a kit carrying a partner’s mark it has not been given.
The first-ninety-days pack promised at signing, built in one consistent run.
/onboard TrellisWill never: Promise a date operations has not agreed to, or leave another partner’s details in via template residue.
Banned terms, policy tripwires, unresolved slots, raw hex, house style. It does not have opinions. It has an exit code.
runs on every draft, before anything leavesWill never: Pass anything on judgment.
One piece, three fresh-context lenses: rules, facts, voice. Returns pass, flag or fail with line numbers. It cannot be talked into being the approver.
/proof 05 Work Zone/review-queue/retention/welcome-letter-DRAFT.mdWill never: Approve, or rewrite the piece. It reports and a human decides.
The same three lenses across the whole review queue at once, counting returns before reading them, then filing the ranked report.
/qaWill never: File a report with a missing verdict as complete.
Traces every number and every claim in a piece back to its authority file.
/verify the partner deckWill never: Accept a figure whose source is somebody’s recollection.
Diffs every status surface against every other and returns the contradictions with file and line on both sides.
/auditWill never: Reconcile by trusting whichever surface is newest.
Probes the brain cold, including refusal traps it is supposed to fail to answer.
/accept run the acceptance testWill never: Judge on self-report rather than on behavior.
The loop that closes after the mistake: what went out, who it reached, what gets corrected, and the new filed fact plus rule so the scanner catches that whole class forever after.
BUILDABLE: The honest gap in this roster. Today the loop ends at approval and never at outcome.
/incident a wrong figure reached a partnerWill never: Close an incident without a rule that prevents the recurrence.
Routes money, legal and external-send decisions to the named human who cannot be bypassed.
BLOCKED: Who holds which approval is itself an open question in the register. It has to be answered before this can exist.
/escalate this needs a signatureWill never: Be the bypass.
Operator command. The whole ship-it pipeline as one careful sequence: rebuild, audit, deploy, sync, push. A failing gate kills the deploy.
/publishWill never: Deploy past a red gate.
Operator command. Moves an approved piece into the official library the safe way: copy, archive the old one, log the change.
/promote the welcome letter, Avery* approved it this morningWill never: Promote without a named human approval.
Regenerates every derived surface: directories, indexes, twins, constitution mirrors.
/regenWill never: Let a generated file be hand-edited.
Runs wide jobs as coordinated fleets with independent verification wired in, one worker per file and a fresh-eyes checker on every return.
/workflow rebrand the collateral libraryWill never: Merge results it has not counted.
Sequences the collateral migration, one reviewed batch at a time, with Jordan* reading the batch report before the next one starts.
/batch start the migrationWill never: Start batch two before a human approved batch one.
Owns permissions and what any connector may reach. It has to exist before any integration does.
/access what can the brain reach today?Will never: Grant an integration a scope nobody asked for.
Every question the intake asks, on one answerable page. The brain never guesses: anything unanswered stays here, numbered, with a destination — and in production, answers typed here file themselves into the right document.
Every decision leadership has made, dated and sourced. Superseded rows stay, marked, because history matters. All rows below are demo examples.
| When | Decision | Approver | Status |
|---|---|---|---|
| Week 1 | New masterbrand selected + trademark filed; name locks at brand-lock; ™ until registration | Dana* + counsel | LOCKED |
| Week 1 | Existing product lines keep their names through the transition window | Leadership | LOCKED |
| Week 1 | Public posture: one brand, told through customer continuity first | Leadership | LOCKED |
| Week 2 | Brand sprint scope: directions → leadership top-2 → one revision round | Avery* | ACTIVE |
| Week 2 | New brain built fresh for Custom LLM (not a legacy-brain override); rules carried, branding rebuilt | Jordan* + Avery* | ACTIVE |
| Week 2 | Atlas adopted as the identity; red retired outright; True North licensed as the one accent, and never as type on white | Avery* + Jordan* | ACTIVE |
| Week 2 | The six colour names locked into the Brand OS (Midnight Green, True North, Paper White, Landfall, Meridian, Surface Steel), each with the ratio it was measured at | Leadership | LOCKED |
| Week 3 | Written retention terms adopted (rate lock + service guarantee); verbal figures superseded | Leadership | LOCKED |
| Week 3 | Launch cascade locked: T-7 sign-off · T-4 manager briefings · go/no-go · launch day hour-by-hour | Launch plan | LOCKED |
| Week 3 | PR agency shortlist to three; selection within the week | Comms | ACTIVE |
Open items live on the Q register — nothing here ships while its question is open.
The mandatory record: every change to the brain, what it was, why it happened, and who made it. Newest first. Entries below are demo examples of the real format.
Everything a teammate needs on one page: how access works, how to connect the files and your AI, how to use the agents and skills, which model and effort to pick, and the habits that keep the brain honest. This demo shows Claude — the same pattern works with any capable AI.
Demo note. Everything on this page is demo content, and names marked * are demo placeholders, not real people.
The company brain is the team’s shared system: one library that holds what is true, how we sound, and what is approved, plus an AI that works inside it like a colleague. It lives in three connected places that always agree: the file library (plain files and folders — host them wherever the team already lives: SharePoint, OneDrive, Google Drive, Dropbox, Box, or a GitHub repo), the team site (the readable, browsable version of everything), and the AI (connected to the folder). You save work into it, the brain checks and files it, and the site updates so everyone sees the same truth.
This one document replaces the usual stack of guides. If something is not answered here, ask on the Team Questions page and the answer becomes part of the brain.
Access is deliberately simple: everyone can read everything, you can only write in your own space, and one admin runs the machinery.
| Level | Who | Can see | Can edit |
|---|---|---|---|
| Admin | Jordan* (operator) | Everything | Everything, site settings, memberships, promotions into the brain |
| Approver | Avery*, Morgan* | Everything | Their own Work Zone, plus approve/return anything in any Send for Approval folder |
| Team member | Sami*, Frankie*, Dana*, Gale*, Marley* | Everything (view only) | ONLY their own Work Zone folder |
Folder-level permissions enforce this on whichever platform hosts the library. Until every level is wired, the working rule is the same: edit only inside your own Work Zone; everything else is read-only by convention.
| You want | Go to |
|---|---|
| Orientation, this guide | 00 Start Here/ |
| The AI’s knowledge: facts, rules, approved library | 01 The Brain/ |
| Brand system and design assets | 02 Brand/ |
| Source documents, transcripts, price files | 03 Knowledge and Sources/ |
| Plans, the strategic plans | 04 Plans/ |
| YOUR working space | 05 Work Zone/<your name>/ |
| The browsable everything | The team site: Assets tab (every file, color-coded), Agents, Skills, Plan, the Batch Board |
Five folders, and that is the entire system: 00 My Space (yours, no rules) · 01 Active Projects · 02 Send for Approval · 03 Save for Later · 04 Archived. Drop finished work into Send for Approval; Jordan*, Avery*, or Morgan* reviews it; approved work gets built into the brain and appears on the site, and anything needing changes comes back to your Active Projects with a note. Nothing is ever deleted, it retires to Archived.
.md is the source of truth; the styled page beside it regenerates automatically. Edit the .md, never the .html.Jordan Vale* — operator, AI owner, access, anything technical. Avery Kellen* — marketing lead, approvals, priorities. Everything else: the Team Questions page on the site.
The AI side of the brain. Connected to the library, the AI answers from filed facts with citations, drafts in the house voice, checks work before approval, and runs the machinery. This demo runs on Claude; the library also generates AGENTS.md, a universal mirror of the constitution, so ChatGPT, Gemini, Copilot, or any other tool the team prefers inherits the identical rules.
cd into the brain folder, then claude. This is where the agents and skills live.AGENTS.md first — the rules travel with the library, not with any one tool.| Surface | What it is | Use it for |
|---|---|---|
| Chat | The browser conversation | Quick thinking, wording help, anything that needs no files |
| Cowork | The AI working inside a folder with you | Day-to-day brain work: drafting, reviewing, filing, “find me X” |
| Code | The terminal operator surface | Agents, skills, batches, site builds — the machinery. Team members mostly will not need it; Jordan* lives here |
Type the slash command, add your request, press enter. Each specialist runs with fresh eyes and reports back. The full roster with examples is the site’s Agents tab, and every specialist ships real work — browse example outputs ↗. The daily four:
/facts — ask anything; sourced answers with file citations, never guesses./proof — verdict on any draft before it goes to Jordan* or Avery* (facts, voice, rules, the scanner)./draft — a new piece the brain’s way: right template, right tone, saved to your Work Zone./revoice — rewrites existing copy into the approved voice.Skills are packaged abilities: you do not type commands, you just ask for the thing (“scan this file”, “build the deck in the Kai* style”, “clean up the BioSummit folder”) and the brain routes to the right skill. The full library with what each does is the site’s Skills tab. Fourteen generation skills are built and switch on automatically at brand-lock.
The table shows Claude’s lineup, the demo’s stack — pick the equivalent tier in whichever AI the team runs.
| Model | Character | Use for |
|---|---|---|
| Haiku | Fastest, lightest | Quick lookups, simple summaries, high-volume small tasks |
| Sonnet | The balanced daily driver | Everyday drafting and questions when speed matters |
| Opus | Deep, careful, thorough | The standard for brain work — anything touching facts, approvals, or deliverables. All brain agents are pinned to Opus |
| Fable | The frontier tier above Opus | The hardest builds and largest campaigns; use deliberately, it costs the most |
Some surfaces let you set how hard the model thinks: Low answers fast with minimal deliberation (lookups, formatting); Medium is the default balance; High reasons deeply before answering (analysis, plans, anything with edge cases). Rule of thumb: raise the effort when being wrong is expensive, lower it when speed matters more than polish.
/proof before anything goes up the chain costs one minute and catches what tired eyes miss.Four stages, about two weeks, one system at the end: a governed library your AI reads before it writes a word — plus the site, the agents, the skills, and a team that knows how to use it.
One structured form. Every answer maps to a file in the brain; every blank becomes a numbered question instead of a guess.
Open the form → STAGE 02The bottleneck-breaker. A master design system with 20 collateral treatments gets swapped to your brand — so every output lands on-brand, every time.
See the full master → STAGE 03About two weeks: constitution, fact layer, voice, scanner, work-zone lifecycle, agents, skills — QA-fleet verified.
The build plan → STAGE 04The first gold-standard pieces ship, the team gets onboarded, and the brain starts doing production work alongside people.
Team & training →One governed source of truth: facts with authority files, voice rules, approved templates, and a scanner that gates every deliverable.
A team of specialist agents and packaged skills — drafting, checking, filing, and wide jobs run as verified fleets.
This site pattern: browsable assets, plans, questions, decisions, changelog — the whole brain visible to the whole team.
A trained team, an approval flow that moves work by moving folders, and a register that turns unknowns into answers.
This whole system — the intake, the Brand OS, the brain, the training — is something I build for companies as an engagement. Email me and we’ll set up a time to talk through yours.
Everything needed to set up the brain, in one pass. Every field maps to a real file in the library — and any field left blank ships as a numbered question in the new brain’s register, because the brain never guesses.
A complete Intake Packet: every answer keyed to its destination file · every unanswered field numbered into the questions register · an upload manifest with a home for every document · a completion percentage. The packet plus the signed-off Brand OS is everything the two-week build needs.
This whole system — the intake, the Brand OS, the brain, the training — is something I build for companies as an engagement. Email me and we’ll set up a time to talk through yours.
Creating this is the first build step of the LLM setup — and it’s the step most AI rollouts skip, which is why they stall.
This is the actual master template every engagement starts from — five chapters (Token Schema · UI Pattern Skeleton · Collateral Layer · Swap Protocol · Exclusions Register), 20 collateral treatments, 395 fill-slots, zero hardcoded colors, rendered live below on the neutral “Placeholder & Co.” identity.
The master is brand-agnostic by construction: every color, font, and name lives in one token table. Nothing is hardcoded anywhere else.
The client’s palette, type, marks, and name replace the tokens. All twenty treatments re-render in the new brand in one pass.
Contrast is measured, exclusions are registered, leadership signs. The treatments become the gold standards every generation skill copies from.
This whole system — the intake, the Brand OS, the brain, the training — is something I build for companies as an engagement. Email me and we’ll set up a time to talk through yours.
About two weeks from intake packet to a working brain — then the part most rollouts skip: making sure the team actually uses it.
| Days | Ships |
|---|---|
| 1–2 | Charter · the six-zone tree · the registers · the constitution v1 with the ten hard rules and the routing table |
| 3–4 | Fact ingestion from the intake packet · the two authority files (pricing, claims) · the voice trio and audiences |
| 5 | The scanner: banned terms, policy tripwires, slot checks — wired into every handover |
| Days | Ships |
|---|---|
| 6–7 | Work-zone lifecycle (position is status) · Brand OS swap locked · tokens wired |
| 8–9 | Generation skills on gold-standard references · the agent team, staged on real gates |
| 10 | The QA fleet’s first live run · the acceptance test · team walkthrough |
The full dated version lives in the project plan, and the gates that pace everything after (brand-lock, launch, phase 2) are on the calendar.
Read-only walkthrough: the site, the zones, where truth lives, how position-is-status works. Everyone leaves able to find anything.
Each person ships one real piece through the full loop: ask → draft → proof → Send for Approval → approved. The flow becomes muscle memory.
Agents and skills for the daily drivers; the questions register as the team’s habit; operators learn the deeper surface.
The How To page and Guides & Tools stay with the team — onboarding that outlives the onboarder.
This whole system — the intake, the Brand OS, the brain, the training — is something I build for companies as an engagement. Email me and we’ll set up a time to talk through yours.
This site is a working rebuild of a real production system — de-branded page by page with the same discipline the brain itself teaches.
{{slot}}.| Real | Demo |
|---|---|
| The architecture, the code, the tree shape, the operating loop, the agents-and-skills design, the QA-fleet method | Every company, person, partner, number, price, date, and logo |
Built and de-branded in Claude Code sessions — the same operator surface the brain teaches — with scripted map passes for the sweeps and hand editing for everything with judgment in it. The full change log is on the changelog, demo-formatted like everything else.
This whole system — the intake, the Brand OS, the brain, the training — is something I build for companies as an engagement. Email me and we’ll set up a time to talk through yours.