AI / STANDARDS
AI agent systems, built standards-first
An AI agent system is only as good as the standards written down before it exists. That ordering is the whole point, and it is the system I build.
The problem
Most teams adopt AI tool-first. Buy the licence, open the blank prompt box, hope. What comes out is fluent, generic and slightly wrong everywhere: the voice drifts, claims appear without receipts, and every person on the team gets a different quality of output because everyone prompts from their own head. The tool isn't the problem. The missing layer underneath it is. An AI with no standards to follow produces content with no standards, just faster.
What I build
I build it in the opposite order. First, the standards: written rules for voice and register, what vocabulary is banned, how claims must carry evidence, how templates are structured, what good looks like for each content type. The kind of documentation most marketing teams keep in one senior person's head, captured and made explicit, with a worked example attached to every rule. A specimen page from the standards pack is published on this site, so anyone can inspect the format before paying for anything.
Then the agents, built on those documents. A drafting agent works inside the rules rather than from a blank page. A review agent checks work against the same rules, flagging where a claim has no receipt or the voice has slipped. And nothing external-facing publishes without a named human signing it off: the agents enforce the bar, people keep setting it.
The outcome
Quality holds because it doesn't depend on the model having taste. It depends on rules that exist in writing, and writing can be versioned, checked and improved. When an output is wrong, the fix is not just the output but the standard that allowed it, and the whole system gets better in one move. That is the difference between renting AI and operating it.
The honest framing: the agents are not the differentiator. The standards are. Anyone can rent the same model. The judgement encoded in the documents it follows is rather harder to rent, which is why the documentation is captured first and owned outright by the organisation it describes.
The stack
- Claude, running agents against a documented standards library
- The standards themselves: markdown documents covering voice, claims, templates and review checklists
- A named human owner signing off anything external-facing
- HubSpot as the execution surface the output ships into
Where next
Next build
A design system for HubSpot: 21 modules, 30 templates, five web properties
Component thinking applied to CRM content, so a multi-brand portfolio ships consistent email and web without rebuilding anything twice.
The thinking behind it
AI agents are the easy part. The standards are the work.
When I build an agent system for a company, the agents come last. What a standards pack contains, why the documentation is the moat, and where governance and sign-off sit.