Context Libraries: What NN/g Found in Its Claude Study

Context libraries are now as important as prompts — and the people who use Claude every day are quietly failing at them. That’s the finding from a fresh NN/g study of expert Claude users, published 9 October 2026. The researchers watched power users spend more effort curating context files than writing prompts, then discovered something uncomfortable: those files went stale, went missing, or stopped matching how the person actually worked.

If you’ve built a context library for your AI agent — a folder of markdown files describing your brand voice, your code conventions, your client preferences — this study is a mirror. Here’s what NN/g found, and what it changes about how you should maintain yours.

What a context library actually is

NN/g defines it plainly: a collection of institutional and procedural knowledge, stored in markdown files or external tools, that AI agents can read and edit. The point is that you write simpler prompts because the agent pulls what it needs from the library itself.

It’s not just files on disk. The study’s participants connected Notion databases, Slack histories, and meeting transcripts through MCP and API links. Context falls into three roles:

• Global — rarely-changing info that shapes the agent across all tasks.
• Local — scoped to one project or client.
• Ambient — raw streams (email, transcripts) the agent sifts through unprompted.

Most designers start with global files: a voice guide, a design-system cheat sheet, a list of tools you use. The study suggests that’s where the rot begins.

Four lessons for your context libraries

1. Context needs ongoing maintenance

A context file written once and used for months will drift out of sync with how you work. NN/g found participants running scheduled routines they’d stopped trusting, or context pieces they’d written and never revisited. One participant’s daily task-summary automation became less useful as his working style evolved — and even with sweeping access to his email and transcripts, the automation’s instructions were too insulated to notice.

The fix NN/g suggests: have the agent audit your global and local context against recent ambient context. If a meeting changed a project’s priorities, ask the agent to compare the transcript with your project context and propose updates. Then verify and approve.

For a product designer, the equivalent check is your design-system context file: does it still name the components, tokens, and spacing rules you actually ship, or the ones you shipped two quarters ago?

2. Let the agent update context for you

You shouldn’t be the only one editing the library. The study’s participants dictated most of their prompts casually and leaned on the agent to reference and update information on their behalf. The maintenance burden shifts from manual rewriting to review: the agent proposes, you approve.

3. Structure for findability

Participants often couldn’t find their own context files. That’s a structural problem, not a memory problem. If your library lives across a local folder, a Notion page, and a few MCP connections, you need a naming and location convention you can actually navigate — and the agent needs one too.

4. Give agents only the context they need

More context isn’t better context. NN/g’s framing: global, local, and ambient context play different roles, and dumping everything into every request makes the agent’s job harder, not easier. Scope local context to the task at hand.

Why this matters if you’re learning the skill

One line in the study should concern anyone training designers: AI systems do little to help users learn context curation. Participants learned mainly from X and Reddit posts and their own experimentation. NN/g argues figuring this out shouldn’t require reading Anthropic’s developer docs.

That gap is the actual opportunity. In our catalogue of design tools, 52 are AI and vibe-coding tools — a category growing faster than the guidance around it. The designers who’ll use these tools well aren’t the ones with the cleverest prompts; they’re the ones who treat their context library like a design system: versioned, reviewed, and pruned.

A practical starting point

If you haven’t built one yet, start small and structured:

1. One global file — your voice, your non-negotiables, your standard tools. Keep it under a page.
2. One folder per project — local context scoped to that client or brief.
3. A monthly audit — ask the agent to flag anything in your global file that no longer matches your workflow.
4. A findability test — can you locate the right file in under ten seconds? If not, restructure.

NN/g’s core warning is simple: an AI agent is only as powerful as its context library, and libraries decay without attention. The experts in this study already learned that the hard way. You can schedule the review instead.

Sources

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