There’s a friction point that most professionals hit within a few weeks of using AI tools seriously. Every session starts the same way: re-explaining who you are, what you do, what you’re working on, how you like things structured. The AI gives good output, but the setup cost is real. You spend ten minutes providing context before you can get to the actual work.
The problem isn’t the AI. It’s the architecture. A conversation-based interface has no memory between sessions by default — every session is a blank slate. The teams and individuals who have moved past this friction didn’t find a better way to explain themselves. They built a system that means they don’t have to explain themselves at all.
This is how to build that system using Claude Projects — the same approach I use for this newsletter.
The Architecture: Two Layers
A persistent AI intelligence system has two components. A master context document — a persistent description of who you are, what you do, and how you work — that the AI carries into every session. And a knowledge base — a library of specific, curated information relevant to your work that the AI can reference as needed.
The master context handles the setup cost problem. Instead of re-explaining yourself every session, that information is permanently available. The knowledge base handles the depth problem — instead of the AI working from general training data, it works from your specific documents, research, and reference material.
Claude Projects implements both. Project Instructions are your master context — they’re included in every conversation in the project automatically. Files added to the project are your knowledge base — the AI can reference them throughout the session without you explicitly providing them each time.
Building Your Master Context Document
The master context document is the most important piece. Done well, it permanently changes the baseline quality of AI output for your work. Done poorly, it creates noise that degrades output. Here’s what it needs to include:
Professional identity and role
Who you are, what your actual job is, and — critically — what you’re responsible for rather than what your title says. A VP of Engineering who’s currently leading a platform migration needs the AI to understand that context, not just the title. Write this as you’d explain it to a new colleague in two minutes: precise, specific, no jargon that requires insider knowledge.
Current focus and projects
What you’re actively working on right now, at whatever grain of detail is useful. For me: the specific topics EchoNerve is covering in the current cycle, which pieces are in progress, what I’ve already published that I shouldn’t repeat. Update this monthly or whenever your focus shifts significantly.
Audience
Who you’re communicating with and what they know. This is the most underspecified element in most people’s context documents. “Technical audience” is insufficient. “Technical practitioners who understand AI concepts at a working level but don’t need academic citation conventions; they read for practical application, not theoretical completeness” — that’s a useful audience description that changes what the AI writes.
Voice and format standards
How you write and how you want output formatted. For writing work: examples of your actual writing are more effective than descriptions of your style. Include two or three paragraphs you’re proud of. The AI will calibrate to them more accurately than to any adjective you use to describe your tone.
For format: specific, negative constraints work better than positive ones. “No bullet points for analytical writing. No ‘In conclusion’ or ‘To summarize’ transitions. Headings should be specific statements, not category labels.” These specifics prevent the defaults that make AI output feel generic.
What to avoid
An explicit list of patterns you don’t want. This is worth taking time on because AI models have strong defaults that are hard to override without explicit instruction. Common ones worth suppressing: excessive hedging language (“it’s worth noting that,” “it’s important to consider”), list-heavy structure for analytical content, vague superlatives, and any phrase that signals the model is performing comprehensiveness rather than thinking clearly.
Building Your Knowledge Base
The knowledge base is a curated library of documents the AI can reference. Curated is the operative word — everything you add increases the context the AI has to process, and quality matters more than quantity. A knowledge base of ten precisely chosen documents outperforms one of a hundred unfocused ones.
What’s worth adding:
Reference documents you consult repeatedly. Style guides, brand guidelines, technical specifications, regulatory requirements, frequently referenced data. Any document you currently find yourself attaching to AI sessions regularly belongs in the permanent knowledge base.
Prior work as examples. Your best previous outputs — reports, analyses, presentations — that represent the standard you’re aiming for. The AI calibrates to them rather than to whatever “good” means in the abstract.
Domain reference material. Primary sources in your domain that you want the AI to work from rather than from general training data. For technical work: documentation, specifications, research papers. For market analysis: company filings, industry reports you’ve verified. For legal work: relevant statutes and cases.
What doesn’t belong: anything you’ve added hoping it might be useful someday. If you can’t articulate the specific task the document will improve, it’s noise. Keep the knowledge base lean enough that you know exactly what’s in it.
Operating Protocols: How to Work With the System
The system is only as good as how you use it. Three operating principles that determine whether this compounds over time or gradually degrades:
Maintain the context document actively. When your focus shifts significantly, update the context document before your next session. A context document that’s six months out of date creates subtle friction — the AI makes reasonable assumptions based on outdated information and you can’t always tell when that’s happening. Monthly review, fifteen minutes: update the current focus section, check whether the voice examples still represent how you want to write, remove anything that’s no longer relevant.
Use the project for all related work. The value of the persistent context compounds when you do all related work in the same project rather than opening separate conversations. Each conversation in a Claude Project has access to the project instructions and files. Using separate conversations fragments the context and loses the continuity benefits.
Add what you reference repeatedly, not what you reference once. When you find yourself adding the same document to multiple sessions, it belongs in the knowledge base. When you find yourself adding a document once and never referencing it again, it shouldn’t have been added. The knowledge base evolves through this discipline.
The Compounding Effect
The reason to invest in this system rather than treating it as overhead: it compounds. A well-maintained context document means the AI’s baseline output quality is consistently higher than it would be with ad-hoc context. You catch errors faster because the AI understands your standards. You spend less time reformatting output because the format standards are embedded. You get better first drafts because the AI understands your audience as well as you do.
The practitioners who’ve built this kind of system consistently report that it changes their relationship to AI tools from “useful occasionally” to “genuinely indispensable.” Not because the AI became smarter, but because the architecture finally gave it what it needed to be consistently useful: a persistent, accurate understanding of who it’s working for and what that work requires.
The setup takes a few hours. The maintenance is fifteen minutes a month. The return is permanent.