McKinsey’s late-2025 research found that only 1 in 10 companies successfully scale their AI agents. The other nine are running the same models as everyone else, with the same access to the same frontier capabilities, and getting nothing usable out the other side.
The difference, almost without exception, is not the model. It’s what the model has been given to work with. That’s the bit most businesses don’t have, and it’s the bit the Firecore is built to be.
I’ve written about why this matters in the experiment I ran giving the same model the same brief twice, and about the gaps that showed up when I built the layer on my own business first. I’ve also covered why context engineering, not better prompts, is where this is heading, and the honest answer to whether you should build your own. What I haven’t done yet is open the bonnet and walk through how the thing is actually put together.
So that’s the job for this piece. What the Firecore is, what sits inside it, and the one design principle that makes it efficient enough to be worth running in the first place.
The simple version: it’s an instruction manual your AI can read
Before any of the architecture, here’s the plain-English description.
The Firecore is a structured, governed record of how your business actually works: your voice, your offers, your client stories, your frameworks, your strategic priorities, and the way you explain the things you explain. Written down once, properly, in a place every AI tool you use can read from on demand.
The induction pack you’d give a brilliant new hire, except it’s for your AI.
That’s the staff-induction-pack version, and it’s the right starting point. But what makes the Firecore different from a long Notion document or a custom-instructions block is the way the contents are organised, indexed, and called. That’s where the architecture earns its keep.
Four layers, fifteen modules
The Firecore is split into four layers, each with its own job, and fifteen modules sitting across those layers. Every module is governed, versioned, and read-only from the outside, which means the AI can trust what it finds there.
Here’s what each layer holds.
The Identity Layer
This is who the business is and how it sounds. Voice and Tone lives here, with all the rules about rhythm, phrasing, banned words, and the patterns that make writing read like the founder rather than a generic agency. Brand Assets and Guidelines sits next to it, holding the palette, the typography, the logo rules, and the visual identity. Personal Profiles holds the people, starting with the founder. Together these three modules answer the question, “who is speaking, and what do they sound like?”
The Business Layer
Five modules here, all answering the question “what does this business actually do, for whom, and against whom?” Business Context sets the funnels, the channels, and the operating model. Products and Services holds every offer, every price, every inclusion and exclusion. Audience and Ideal Clients defines who the business is for, with the same level of detail you’d brief a salesperson. Competitors holds the live competitive landscape. Market Position sits over the top, holding the positioning narrative, the third-party evidence stats, and the approved analogies.
The Evidence Layer
Two modules, both load-bearing for content. Client Stories and Case Studies is where every real outcome, every quote, every named example lives. This is the module that stops AI inventing case studies, because it now has a canon of real ones to pull from. Frameworks and IP holds the proprietary methodologies, the named processes, the explanation patterns. When the AI references “the WordPress reframe” or “the Flamekeeper problem”, it’s pulling from here, not making it up.
The Operations Layer
The last five modules cover how the business is run. Strategic Priorities holds what’s live, what’s paused, and what’s on the kill list. Business Plan and Financial Context holds the numbers and the runway. Key Relationships and Partners holds the warm contacts and active partnerships. Communication Templates and Patterns holds the proposal structures, the email patterns, the sales conversation moves. Industry and Regulatory Context holds the rules of the road for the sector.
Four layers. Fifteen modules. One governed record of how the business operates, ready to be read on demand.
The thing that makes it efficient: it routes, it doesn’t read
Now for the part most people get wrong when they imagine what an AI context layer looks like in practice.
If you picture the AI reading every word of every module on every task, you’ve imagined a slow, expensive system that nobody would actually use. Fifteen modules of detailed business intelligence, loaded into the conversation every time you ask for a social post, would burn through tokens, slow the response, and bury the model in noise it doesn’t need.
That’s not how the Firecore is built.
Here’s the way I explain it in proposals, because it’s the most useful frame. The system is designed to be efficient. Instead of your AI reading everything every time, it reads a summary index, decides what’s relevant, and loads only what it needs. Most routine tasks touch three or four modules, not all fifteen. It’s like the difference between reading an entire encyclopaedia and using the index to find the right page.
That routing logic is what turns the Firecore from a nice idea into something that works in practice.
Every module carries a frontmatter summary at the top. A short, tight description of what the module holds, what it’s for, and when to fetch it. The AI reads the summary index first, picks the modules that match the task at hand, and pulls in only those. A social post drafting task might touch Voice and Tone, Brand Assets, and Frameworks and IP. A proposal draft might touch Products and Services, Market Position, Client Stories, and Communication Templates. A strategic decision might touch Strategic Priorities, Business Plan, and Audience and Ideal Clients. Different jobs, different reads.
The full encyclopaedia is still there. The AI just doesn’t read it cover to cover for every small task.
Why this is not the same as a long prompt
I get this question often. If the Firecore is just structured information for the AI to read, can’t I do the same with a long ChatGPT custom-instructions block or a master prompt I paste at the start of every chat?
In all honesty, no, and the reason has very little to do with how much you can fit into a text box.
A long prompt is a snapshot. You write it once, you paste it in, and from that moment on it ages. The Firecore is a living, versioned, governed record. Every change is logged, every update is tracked, and the AI is always reading the current state. The pricing that was right in March doesn’t quietly hang around for the rest of the year, because the system has been updated and the old version has been archived with a date stamp.
A long prompt is also flat. Every line carries the same weight, the same priority, and the same level of access. The Firecore is structured. The AI knows which module owns voice rules, which module owns pricing, which module owns case studies, and what to do when two sources seem to contradict. That structure is what makes the answers consistent across tasks, across team members, and across time.
And a long prompt has no governance. There’s nothing stopping anyone editing it, no record of what changed, no way of seeing whether the version in use is the approved version. The Firecore is read-only from the working surface and edited only through a controlled change-request process. Nothing drifts into the system unnoticed.
The shortest version is this: custom instructions are a Post-it note on the fridge. The Firecore is a franchise manual.
What this looks like in a real working week
For a five-person agency or a thirty-staff accountancy, the value of the architecture shows up in the most ordinary way imaginable.
A team member asks the AI to draft a proposal for a new prospect. The AI reads Voice and Tone, pulls the right rhythm and phrasing. It checks Products and Services for the current price and inclusions. It fetches the relevant case study from Client Stories. It applies the proposal structure from Communication Templates. The draft comes back sounding like the company, with the right numbers, the right examples, and the right shape. Nobody is doing the patch-up edit that used to take half the afternoon.
The same architecture handles the social post, the client email, the sales script, the welcome pack for a new starter. Different modules called, same governed source.
What changes for the founder is the bit nobody talks about. The output stops needing to be rewritten. The new hire produces on-brand work in week two instead of week six. The evening rewrite session quietly disappears. None of it is dramatic. All of it adds up.
The architecture is the system
What I’d say, after building this on my own business and watching it run for months, is that the structure is the bit you’re actually paying for.
The platform underneath is interchangeable. It happens to run on Notion in the version I install, because Notion gives clients a clean place to read and govern their own data. If Notion disappeared tomorrow, the same four layers and fifteen modules would rebuild on a different platform in a week. The system isn’t the software. The system is the way the information is split, indexed, and routed, and that’s portable.
Which is also why a competitor copying the feature set doesn’t worry me. The architecture is the product of hundreds of hours of real-world testing. The modules are the result of watching what AI tools actually need, and what they keep getting wrong when they don’t have it. None of that is reproducible from a screenshot.
If you want to see whether your business has the kind of gaps the Firecore is built to close, I built a short AI Context Layer Risk Quiz that works as a diagnostic. It takes about six minutes and walks through where your voice, your offers, your client stories, and your strategic priorities currently live, then shows you the specific gaps that are dragging down your AI output quality. Take the quiz to see whether your AI output quality is where it should be. You get a personalised score and a short report, no sales call required.
If you’d rather just talk it through, reach out to me. Use me as a resource. I’m always happy to help.