An AI agent that knows nothing about your client will produce work that looks fine and fits nobody. It doesn't know the brand avoids the color red because of a competitor, that the app speaks Egyptian Arabic and not Modern Standard Arabic, or that the client killed the bottom-sheet checkout two sprints ago. Prompting harder doesn't fix that. Context does.
NN/g recently wrote about "context libraries": the files power users feed their AI tools, and how those files decide how useful the tools are. Figma's agent is now generally available and supports team guidelines, so there's finally an obvious place to keep that context. For an agency the problem is bigger, because we don't manage context for one product. We manage it for every client at once.
Key takeaways
- For client work, the quality of an agent's output depends more on the context you give it than on the prompt.
- Give every client a context pack: brand rules, language and voice, tokens, past decisions and things to avoid.
- Language context (dialect, formality, transliteration) is the part most teams skip and the part Arabic products need most.
- Context goes stale. Give each pack an owner and update it at the end of every sprint.
- Keep each client's context separate, both for confidentiality and so one brand doesn't leak into another's work.
Why "just prompt it" fails on client work
A prompt describes one task. Client work depends on everything around the task: what the brand allows, what was already tried, who the users are and what the stakeholders care about. When that's missing, the agent fills the gaps with the most generic answer available. The result passes a quick look and fails the client review.
Retyping that context into every prompt doesn't scale either. Different people on the team will describe the same client differently, and the agent will get a slightly different brief each time.
Anatomy of a client context pack
A context pack is a small set of files per client, stored in one place and loaded before an agent does any work for that client. Here's what goes in it:
1. Brand rules
Colors, type, imagery, tone, and the things the brand never does. Keep it short and specific. "We never use stock photos of people" helps more than a 40-page brand book the agent will only skim.
2. Design tokens and components
The client's token file and component list, or a link to the Figma library. If you use Code Connect, the mappings belong here too, so the agent builds with real components instead of lookalikes.
3. Product and user context
Who the users are, the main flows, the markets, and the key metrics the client cares about. One page is enough.
4. Decision log
What was decided, when, and why. "Removed guest checkout in March because of fraud" saves you from an agent cheerfully suggesting guest checkout again.
5. Things to avoid
Competitor patterns, rejected directions, words the client dislikes, legal constraints. Agents are good at following explicit "don't" lists.
[Waleed: add how Breem stores packs today, e.g. folder structure or Figma team guidelines, plus a screenshot.]
Language and voice context
This is where most context packs are thin, and where Arabic products suffer most. Write down:
- Dialect: Modern Standard Arabic, Egyptian, Gulf or Levantine, and whether it changes by screen. Marketing copy might be colloquial while legal text is MSA.
- Formality: how the product addresses users, and whether it uses gendered forms or avoids them.
- Transliteration rules: how brand and product names appear in Arabic, and which terms stay in English.
- Numerals and formats: Western or Eastern Arabic digits, currency format, date format.
- Example strings: ten real, approved strings in each language. Agents copy examples far better than they follow descriptions.
Keeping context current
NN/g's point that heavy users let their context files go stale matches what I've seen. A context pack written at kickoff is out of date by the third sprint. Two rules keep it alive:
- One owner per client. Usually the lead designer on the account. Anyone can suggest changes; one person merges them.
- Update at the end of every sprint. Add new decisions to the log, remove rules that no longer apply, and refresh tokens. It takes fifteen minutes.
Date the pack. If an agent's output feels off, the first thing to check is when its context was last updated.
Confidentiality and separating clients
An agency has a problem a product team doesn't: context from one client must never show up in another client's work. That's a confidentiality issue, and also a quality issue. You don't want a fintech client's tone of voice appearing in a food delivery app.
- Keep one workspace, project or folder per client, and load only that client's pack.
- Don't keep a shared "everything" context that mixes clients.
- Check the client's contract and NDA for rules about AI tools before you upload anything.
- Remove a client's pack from tools when the engagement ends.
A template to start from
client-name/ 00-readme.md owner, last updated, how to use this pack 01-brand.md brand rules and things the brand never does 02-voice.md dialect, formality, transliteration, sample strings 03-product.md users, markets, key flows, metrics 04-decisions.md dated decision log 05-avoid.md rejected directions, competitor patterns tokens.json design tokens (or link to the library)
Start with the readme and the voice file. Those two fix most of the generic output.
See it in real projects
The products in my portfolio were designed for clients across MENA, often in Arabic and English. If you're setting up AI workflows for client work, my Arabic RTL design checklist is a good companion for the voice and language parts of the pack.
Sources: Context Libraries: How AI Power Users Manage Context Engineering (NN/g), Figma's AI agent for product design is now generally available (The Verge), The benefits of Code Connect in MCP (Figma).
FAQs
What is a client context pack?
A small set of files per client (brand rules, voice and language rules, tokens, a decision log and a list of things to avoid) that an AI agent loads before it does any design work for that client.
How often should AI context files be updated?
At the end of every sprint, by one named owner. Add new decisions, remove outdated rules, and date the file so everyone knows how current it is.
How do you keep client data separate when using AI tools?
Use a separate workspace or project per client, load only that client's context, never keep a shared context that mixes clients, and check the client's contract for rules on AI tools.