How to Build a Context Library That Keeps Claude Useful

Claude power users write simpler prompts than you do. Not because they’re better at prompting — because they’ve moved the hard work somewhere else: a context library of markdown files, notes, and connected sources that the agent reads and updates on its behalf. That’s the core finding of a recent Nielsen Norman Group study of expert Claude users, and it reframes how designers should learn AI tooling.

The uncomfortable part: those same power users rarely edit their context files, often can’t find them, and suspect someone else is doing it better. NN/g’s research shows that even experts are improvising — and that the AI systems themselves do little to teach you how to curate context. Most participants learned from X posts, Reddit threads, and their own trial and error.

What a context library actually is

NN/g defines it plainly: a collection of institutional and procedural knowledge, stored in markdown files or external software, that AI agents may reference and edit. The agent does the typing. You decide what context to keep, where it lives, and when it gets updated.

It isn’t only markdown. MCP and API connections pull in Notion databases, Slack histories, and meeting transcripts. The study sorted context into three roles:

• Global — information that rarely changes and shapes the agent across all interactions (your brand voice, your design principles).
• Local — scoped to one project or task (a client’s component library, a sprint’s constraints).
• Ambient — raw, uncurated streams like email and transcripts the agent sifts through.

If you’re a product designer running Claude alongside Claude in a daily workflow, this distinction matters: global context is what you write once and trust for months — which is exactly where the rot sets in.

Lesson 1: Context needs ongoing maintenance

Files written once and used for months stop matching how you actually work. Participants had created pieces of context and never returned to them. Some kept running scheduled routines they’d quietly stopped trusting, or routines depending on brittle external connections that kept failing.

One participant’s daily task-summarizing automation degraded as his working style evolved. Even with sweeping access to his email, messages, and transcripts, the automation’s instructions were too insulated to adapt — and he’d started ignoring the useless parts without telling the system.

Lesson 2: Let the agent update context for you

NN/g’s fix is counterintuitive: ask the AI to audit your context against recent ambient context. If a meeting shifted a project’s priorities, have Claude compare the transcript with the relevant project context and propose updates. Verify the changes, then let it apply them.

They also recommend giving explicit feedback when your way of working has drifted from an established routine — the agent can’t sense that on its own.

Lesson 3: Structure for findability

If you’ve ever hunted for a context file you know you wrote, you’re in the study’s sample. Structure the library so both you and the agent can find things — a naming and folder convention you’d apply to any design system’s documentation. Think of it as information architecture for your AI, which is squarely a product designer’s skill.

Lesson 4: Give agents only what they need

Scope context to the task. Global context shapes everything; local context should attach only where it’s relevant. Dumping everything into one file makes the agent slower to find what matters and harder for you to audit.

What this means if you’re starting out

The study’s participants were experts working with Claude daily. NN/g notes that users in nontechnical domains or with lower AI literacy will hit the same problems amplified — fewer workarounds, less ability to notice when the library is failing them.

A practical starting point: one global file (who you are, how you write, non-negotiable design rules), one folder of local files per active project, and a monthly 15-minute audit where you ask the agent to flag stale instructions. That’s not a system you buy — it’s one you maintain, and the maintenance is the whole point.

For designers comparing where to put this effort: our catalogue of 607 design tools includes 52 in the AI & Vibe Coding discipline alone, but none of them will tell you when your own instructions have gone out of date. That part is still on you.

Sources

• NN/g — Context Libraries: How AI Power Users Manage Context Engineering