AI writing agent
A brand-governed writing agent producing on-brand work in English and native Arabic, with a human sign-off that's enforced in the database, not requested in a prompt.
The problem
A marketing team was producing content in two languages by hand. Brand voice drifted between writers. Arabic was being translated from English rather than written in Arabic, so it read like a translation. And there was no reliable way to stop off-brand or unchecked work going out. The controls were guidelines in a document, which is to say they were suggestions.
What we built
A writing agent with three modes (copywriting, structured documents, and video scripts) that takes a brief and returns on-brand drafts in English, Arabic, or both. Nothing it produces can leave the system without a named person approving it.
The important part isn't the drafting. It's that the rules are enforced where they can't be talked around.
How it works
- 01
Brief
a structured intake per mode, which fails closed: no platform, no copy; no source material, no document; no duration, no script.
- 02
Draft
generation runs against a versioned brand record loaded at runtime, so voice, approved claims and house rules come from one place that can be updated without touching a prompt.
- 03
Review
every draft goes to a named reviewer. Arabic output additionally requires a native-Arabic reviewer. Documents additionally require a subject-matter expert.
- 04
Approve
approval is reachable only through a guarded database function, and a sign-off is valid only against the draft it was given for. Edit the draft, and the sign-off no longer counts.
What it delivers
On-brand bilingual drafts, and a paper trail. Every output carries provenance, every approval is attributable to a person against a specific draft, and every human edit at the review gate is captured as training data the agent improves from.
No usage figures yet. The agent moved onto real model calls recently, so any time-saved or volume claim would be invented, and we'd rather show you the build. What is measured: 33 of 33 gate tests and 73 of 73 confirmation tests passing on every change, and accessibility taken from 84 violations to zero across all fifteen screens, on desktop and mobile.
One known gap: there's no in-app way to answer the agent's clarifying question yet. The workaround is to submit a complete brief and regenerate. It's on the list.
Who else this benefits
Any team producing regular content in more than one language, or in a regulated space where someone must be accountable for what goes out: marketing teams, agencies, clinics, financial services, and anyone whose brand voice currently lives in a PDF nobody opens.
Key features
- Three modes (copy, structured documents, video scripts) on one shared spine
- English and native Arabic, composed in-language rather than translated, with correct RTL and numeral handling
- Five safety rules enforced in code, not in prompts: no send path exists anywhere in the system · approval only via a guarded function · brand isolation · provenance stamped on every output · generation fails closed
- Brand memory as a versioned record with history and restore, so no brand rule is hard-coded in any prompt
- A feedback loop that writes a correction record inside the same transaction as each approval, so it cannot be skipped or forged
- Facts in documents come only from supplied source material; anything unverifiable is flagged rather than invented
Tools
- Next.js
- TypeScript
- Supabase
- Claude
- Vercel
- Vitest
- Playwright + axe-core
Want an agent that can't go rogue?
We build AI systems where the guardrails are enforced in the build, not requested in a prompt. Tell us what your team is doing by hand.