When I first wrote this, a shared brain was a set of files every writer and AI tool read from. Six months of building the full version inside a regulated marketing team turned it into the bottom layer of a seven-layer stack: context, rules, skills, agents, controls, governance and evidence. Here is the stack, what each layer adds, the order that matters, and how the rollout changed.
TL;DR: Most marketing teams have their ICP in a Notion doc, their value prop in a deck, their tone of voice in a Google Doc and their messaging matrix in someone’s head. A shared brain fixes that by putting the foundational truth of the brand into one versioned set of files that every person and every AI tool reads from. That was the whole of this article when I first published it in May. Since then I have built the full version inside a regulated B2B marketing team, and the shared brain turned out to be the bottom layer of something larger: context, rules, skills, agents, controls, governance and evidence, in that order. This is the rewritten version, with the stack, what each layer adds, the order that matters, and how the rollout changed.
First published 20 May 2026. Rewritten 6 October 2026.
The problem: every team has a brain, no team has a shared one
Walk into any marketing team and ask four people to define the ICP. You will get four answers. Ask them to write the value prop in one sentence and you will get four versions. Ask where the messaging matrix lives and you will get four locations.
This is the natural state of a function whose foundational documents live in a founder’s head, a sales deck from eighteen months ago, a brand workshop output from the last agency, and a “messaging v3 FINAL FINAL.docx” in someone’s Drive. When the team was four people sitting together it worked, because the brain was shared by osmosis. It stops working the moment the team has freelancers, an SEO agency, a PR agency, a designer, a demand-gen contractor and a stack of AI tools all generating content from whatever brief the human in front of them pasted in.
Every one of those touchpoints is a place where the brand drifts. Multiply by the volume a modern team ships and you get the complaint every CMO makes: “It all sounds slightly off-brand and I can’t put my finger on why.” There is no single source of truth, or there are seven, which is the same thing.
What a shared brain is
A shared brain is a versioned, structured set of plain-text files holding the foundational truth of the brand: who you sell to, why they buy, how you talk, what you sell, who you compete with, and what good looks like. It lives in a Git repository, so it is reviewable and diffable. It is loaded into your AI tools as context, so every prompt is grounded in it. And it is the brief every human writer works from too.
The point is not the file format. The point is that there is exactly one canonical version of every foundational fact, and everything else reads from it. When the ICP changes you change one file, and every piece of content created after that inherits the change.
| Layer of the brain | Files | Changes |
|---|---|---|
| Foundation | ICP, personas, value prop, competitors, tone of voice | Quarterly at most |
| Application | Messaging matrix, objection handling, content pillars, writing rules, style guide | When the foundation does |
| Examples | Approved samples per channel, case studies in the house structure, before-and-after fixes | As good work ships |
Three of the foundation files matter more than the rest. The ICP answers “who is this for?”, and a real one is specific enough that a list-builder can pull a target list in ten minutes. Personas answer “who inside that account am I writing for?”: the champion and the economic buyer read different things and respond to different proof. The value prop answers “why should they care?” in one sentence, backed by the differentiators that make it credible. Get those three right and every piece of content has a built-in answer to the three questions that decide whether it converts. Get them wrong and the rest of the stack produces mediocrity at scale.
What I learned building the full version: context is the floor
In May I thought the shared brain was the system and the skills were the productive surface on top of it. Six months of building one inside a marketing team of around thirty, in a regulated category where a wrong claim is a legal problem, changed that view. The shared brain is the floor. Six more layers sit on it, and each one exists because something went wrong without it.
| Layer | What it holds | A concrete example | Who owns it |
|---|---|---|---|
| 1 · Context | Verbatim copies of approved sources, each with its owner and migration date in the frontmatter, plus a register of what does not exist | The tone of voice file is owned by one director. Any attempt to write to it stops, names the owner and asks “are you sure?” | The named owner of each file, by job title |
| 2 · Rules | 52 standards in nine groups, approved on a recorded date | “Evidence-bound outputs”: every material claim carries a source and a date, and regulatory statements stay claims until a subject-matter expert clears them | The AI programme lead |
| 3 · Skills | Around forty repeatable workflows, each with an owner, an agent role, a lifecycle state and a line saying where its output goes | The Content ROI Auditor reads the ICP, value prop and competitor files, and never writes to context; a proposed change goes by pull request | The job title in the skill’s frontmatter |
| 4 · Agents | Eleven specialist configurations; three more deferred until their data connectors are approved | The brand and editorial reviewer’s stop rule: “cannot approve regulatory claims or its own drafts” | The role that owns the agent’s skills |
| 5 · Controls | Deterministic hooks wired into the session, a register of 29 controls with their state, and the designs for the services not yet built | A handover that does not start with “Skill run:” is sent back to the agent by the runtime | Marketing ops, with IT and security |
| 6 · Governance | Committee charter, approval matrix by risk tier, five release gates, data policy, roles mapped to job titles | No agent runs unattended before Gate 4; scheduled jobs exist in the catalogue with every one disabled | The AI committee |
| 7 · Evidence | Golden adversarial cases, a context benchmark, logged agent runs before and after each change, pilot baselines | Three adversarial cases every evidence skill must pass: a stale law, an unsupported claim, a contradictory source | Analytics, with the domain expert |
Layer by layer
Context, with owners and a sell-by date
Two things changed from the May version. First, context files are verbatim copies of approved sources, never summaries, and each carries the owning job title and the date it was pulled. A session-start check lists anything older than ninety days so it is re-confirmed with the owner before being cited. Second, the repo keeps a register of what does not exist. If nobody has approved an expert point of view library, the folder is empty and says so. Nothing is drafted to fill a gap, because an invented fact in the brain propagates to every tool that reads it.
One rule made the whole thing work socially: owners are job titles, never names. People leave. The Content & Communications Director still owns the voice file after the person in the role changes.
Rules the skills apply but never restate
Standards that live inside a prompt get copied, drift and contradict each other. So they live once, in nine files, and every skill points at them. These are the nine groups.
| Rule group | What it governs |
|---|---|
| Output tenets | What every deliverable must contain and how it is structured |
| Ontology and taxonomy | The controlled vocabulary: product names, segments, personas, never-say terms |
| Planning doctrine | How briefs, calendars and campaigns are planned before anything is produced |
| Orchestration patterns | When work is chained, parallelised or handed between skills, and the limits on each |
| Evidence-bound outputs | Sources and dates on every material claim; regulatory statements are claims until cleared |
| AI-speak and editorial anti-patterns | The phrasing and structures that mark text as machine-written, and are refused |
| Marketing psychology and ethics | No manipulation, no false urgency, no misleading comparisons |
| Cost and credit gates | Spend ceilings and when a workflow must stop and ask |
| PII and access protection | No contact-level data in the repo; what may and may not be read into a session |
Skills that know where their output goes
The skills are still the productive surface: research, positioning, messaging, content, campaign activation, web QA, and the chains that run eight of them in sequence. What changed is the frontmatter. Each skill now states its owner, the agent role that runs it, its lifecycle state, and the system its output lands in. That last line solved more arguments than anything else, because the team runs on three systems and each has one job.
A skill whose output is a candidate update to a locked fact, say a refreshed persona, names the context file a maintainer updates by pull request once it is approved. The skill never writes to context itself. That one boundary keeps the brain clean while forty workflows run against it.
Agents: a role, its skills, and a stop rule
An agent in this model is a thin thing on purpose. It is a specialist role from the operating model, the skills that role owns, a stop rule quoted from the catalogue and never softened, and a standing block of rules identical across every agent. It holds no facts of its own and takes no external action.
| Agent | Owns | Stop rule |
|---|---|---|
| Research specialist | ICP research, company context, competitor research, win-loss analysis | Produces findings with sources; never positions or recommends messaging |
| Product marketing specialist | Positioning, messaging, vertical messaging, one-pagers, launch chain | Cannot lock positioning; proposes it for the owning director |
| Brand and editorial reviewer | Tone of voice audit, brand kit | Cannot approve regulatory claims or its own drafts |
| Evidence review specialist | Content audit, Content ROI audit, evidence check | Flags and routes claims; never clears a regulatory statement |
| Campaign operator | Activation checklist, content ops, webinar brief, pre-mortem | Never sends, publishes, enrols or changes spend |
| Lifecycle and nurture specialist | Nurture sequence plans, email copy | Plans and drafts; the send is a human action against an approval record |
The standing rules are where the first round of testing paid for itself. Early runs produced good work in the middle of a session and then lost it, so “only your last message is delivered” became a rule. Some runs answered a question from memory, so “read the skill and its context files fresh, every run” became a rule. Some invented a rule to fill a gap in the brand file, so “no rules of your own; route the gap to the owner” became a rule. And a hook now checks that every handover starts by naming the skill it ran, or sends it back.
Controls that run whether or not the prompt asks
Rules are soft. A prompt can be argued with. Controls are the deterministic layer: small scripts that run on every tool call, regardless of what the session is trying to do. The register lists 29 of them with their state. These are the ones enforced in-session today.
| Control | What it does |
|---|---|
| Kill switch | One file at the repo root blocks every tool call until it is removed |
| Protected files | Any write to context or rules pauses, names the owning role and asks for confirmation |
| PII and secret gate | Refuses any write containing personal contact data or credentials |
| Shell guard | Refuses force-push, recursive deletes, hard resets and outbound POST requests |
| Context freshness | At session start, lists context files older than the review window |
| Handover check | Returns any agent handover that does not name the skill it ran |
| Local telemetry | Logs which files changed, when and by how much; never the content |
| Approval record | An external action needs an approval bound to the exact payload hash and destination, by an authorised role, before it expires |
| Scheduled jobs | Eleven are catalogued; every one is disabled until the governance gate that allows unattended runs |
The approval record is the one I would carry into any team tomorrow. It binds a yes to a specific file, a specific destination and a date. Change the copy after approval and the approval is void. The producer of an asset never approves it. Nothing is an approval until the approver has written their own name on it.
Governance that lives in the repo
The committee charter, the approval matrix by risk tier and job title, the five release gates with their evidence tables, and the data policy all live next to the code, and CI checks them on every pull request along with a secrets scan and a frontmatter check. Owners are job titles. CODEOWNERS routes any change to the voice and style files to the person in that role. None of this is exotic engineering. It is the ordinary discipline software teams have used for twenty years, applied to marketing’s source of truth for the first time.
Evidence that it works
Every agent run during configuration was logged, before and after each change, so the standing rules above are traceable to the run that caused them. A golden set of adversarial cases checks that the evidence skills catch a stale regulation, an unsupported statistic and two sources that disagree. A twenty-question benchmark checks that the context answers the questions the team actually asks. And three pilots each have a baseline, an owner, and an acceptance bar before anything scales: a minimum number of evaluated runs, zero critical claim errors, a reviewer acceptance rate, and a measured time saving net of review.
The order matters more than the parts
If I had to compress six months into one sentence: build context before skills, skills before agents, and controls before anyone runs an agent unattended. Every shortcut I watched a team take ran the other way. Agents first produce confident, well-formatted output grounded in nothing. Skills before context produce the same output in a nicer template. Automation before controls produces a send nobody can trace to an approval.
The same ordering runs through the pod model I wrote about last week. The pod is the people. This stack is what the pod runs on. The platform layer in that post is layers one, two, five and six of this one.
How the rollout changed
In May the rollout was three phases: a shared Claude project, then a GitHub repo, then Claude Code on everyone’s machine. That still works for a small team. For a team of thirty in a regulated category it was too thin, because it had no gates and no evidence. The rollout I run now has five phases and a decision at the end of each.
| Phase | What lands | Gate at the end |
|---|---|---|
| Mobilise | Inventory the stack and the existing wiki; stand up the tracking board; agree who approves what | Sponsor sign-off on scope and approvers |
| Foundations | The repo, the context layer migrated verbatim with owners, the rules, the first controls | IT and security release the repo for use |
| Specialists | Skills repointed at the context; agents configured and tested, not yet in production | The committee releases specialists for supervised use |
| Prove and connect | Three pilots run as tests against baselines; read-only data connectors wired in under the controls | Pilot results clear the acceptance bar; connectors approved |
| Scale and sustain | Team trained, approved jobs switched on, quarterly refresh of context, the committee’s standing cadence | Acceptance by the sponsor; the system becomes business as usual |
The first two phases are still mostly writing. The technology is the easy bit. Migrating a hundred wiki pages verbatim, finding the owner for each, and deciding what is approved fact versus raw material is the slow, valuable work, and no tool does it for you.
Why none of this works without the floor
The clearest proof is still the Content ROI Auditor. It scores any article across six commercial dimensions, pipeline alignment, commercial intent, CTA strength, specificity, search intent match and content type effectiveness, and tells you whether the piece has a path to pipeline or is vanity content burning budget. Its required inputs are non-negotiable: the ICP, the value proposition, and the named competitors. Without those it can comment on writing quality and nothing else. That is the difference between a copyediting tool and a strategic audit.
Every high-leverage skill has the same shape. It looks like it is about auditing, drafting or briefing. It is actually about the foundation it reads from. Teams that evaluate AI tools in isolation, see uneven output and conclude the tools are “not quite there yet for marketing” are usually looking at a missing floor.
| Skill | Reads from | Produces |
|---|---|---|
| Content ROI Auditor | ICP, value prop, competitors, evidence rules | Per-piece commercial scorecard and prioritised rewrites |
| Thought leadership writer | ICP, personas, tone of voice, content pillars, approved samples | Long-form draft against a brief, with distribution excerpts |
| Messaging builder | Locked positioning, personas, objection handling | The ten-component messaging library per product line |
| Nurture sequence plan | Personas, messaging, lifecycle rules, eligibility rules | Trigger, segment, touch-by-touch plan and exit conditions |
| Evidence check | Evidence rules, claims register, approved proof points | Flagged claims with the source each one needs |
How I do this for clients
For a team that can maintain a repository, I run the five phases above over roughly fourteen weeks, with the first deliverables in the first week: the inventory, the repo skeleton and the first context pack migrated with owners. For a smaller team without that appetite, I run the compressed version from May, a discovery sprint, an application sprint and an examples sprint, and hand over the repo with the controls already wired in, so the floor is sound even if the upper layers come later.
Two of the foundational skills, free
Two of the skills I package into every shared brain are available to download. They are standalone, so you can drop them into Claude Code today without the rest of the stack.
- ICP Research skill: triangulates the founder view, sales win patterns and the competitive landscape into a usable ICP document with disqualification criteria and intent signals.
- Competitor Research skill: produces a structured competitive landscape, direct and adjacent, with the wedge analysis and the messaging implications.
If you want the full stack stood up inside your team, with the context migrated, the rules approved, the skills and agents configured and the controls running before anyone presses send, book a strategy call.

Jessica Redman
GEO, SEO and AI enablement consultant. Ten years across insurance, SaaS, eCommerce and regulated B2B.
Keep reading