Waterhouse
Brands

A custom WordPress build - bespoke theme, no page-builder - designed around the constraint that this is a firm whose product is judgment. The site has to read as considered rather than busy, so the work leans on generous whitespace, restrained typography, and full-bleed imagery that does the emotional lifting while the copy stays plain.

Under the hood it's a modular template system: the team composes case studies and service pages from a fixed set of blocks rather than filing a ticket for every layout change. Same principle we apply to email component libraries, applied to a marketing site.

The more interesting half of the engagement. Waterhouse runs a four-phase strategy methodology - Assess, Listen, Interpret, Galvanize - that produces long-form strategic documents: competitive audits, stakeholder synthesis, insights reports, and full brand positioning platforms.

Like every agency, they had started using LLMs somewhere in that drafting process. And for a brand agency specifically, generic LLM prose is not a cosmetic problem. It is an existential one. Clients pay for a point of view, and the failure mode of an LLM is confident text with no position in it.

ALI encodes the house voice as a machine-checkable system rather than a style guide, and it knows which phase of the methodology it is operating in - the register for a stakeholder synthesis is not the register for a tagline.

Rules the machine can check, judgment the human keeps

ALI is deliberately two-layered. The deterministic linter handles everything literal - a banned phrase either appears or it doesn't, an em-dash is or isn't there. That work is mechanical, it should never require a model, and running it as code makes it fast, free and impossible to argue with.

The model handles what the script cannot see. Is this three-item list substantive enumeration or a rhythmic tricolon? Is this the one earned antithesis in the section, or the fourth? Has the writer taken a position the reader could disagree with, or written something merely plausible?

Neither layer works alone. A linter with no judgment produces prose that passes every check and says nothing. A model with no linter drifts, because "avoid AI tells" is exactly the kind of instruction models are worst at holding across a long document.

ALI was built first for a terminal-based agent runtime with real file and shell access, which is where it works best. But most of the agency doesn't live in a terminal, so it's also packaged as a Custom GPT: the same rulesets as knowledge files, with the audit script translated into a structured manual self-check for a runtime that can't execute code.

The result is that a voice standard which used to live in a document - read once, half-remembered, applied unevenly - now runs as a step in the process. Drafts get audited before they reach a client, violations come back with line numbers, and the arguments that used to be about taste are now about evidence.