What's the difference between the Knowledge base and System Prompt? What should go where?
System Prompt holds 'rules for how to work': role setting (who you are, how to think), behavioral rules (response format, tone requirements, prohibitions), task boundaries (what this Project mainly handles). System Prompt is the rulebook for 'how you should behave.'
Knowledge base holds 'reference material needed while working': company background, product descriptions, terminology guides, brand handbooks, quality past output examples. The knowledge base is 'reference material to consult when making decisions.'
Simple decision rule: if you're adding 'a rule or requirement' ('responses should not exceed 200 words,' 'tone should be formal') → System Prompt. If you're adding 'a document or file' (company overview, format example from last quarter's report) → Knowledge base.
Best used together: System Prompt says 'you are a professional editor helping [Company] write client communication documents'; Knowledge base holds company overview, brand voice guide, and three quality past email examples. Claude knows its role (System Prompt) and knows what to reference (Knowledge base) — output consistency and quality both improve.
What's most useful to put in the Knowledge base? Any recommended content checklist?
First priority: company and client background 200–500 word company overview (core business, target customers, key value propositions), plus for multi-client projects, 100–200 words per client. After this, Claude no longer needs you to explain 'what we do' when drafting documents or communications.
Second priority: terminology and format specifications Company or industry-specific terminology (especially terms that external people frequently get wrong), required format conventions (words to avoid, heading formats, number formats). This solves 'Claude uses the wrong term' without you explaining in every prompt.
Third priority: your best past outputs as examples 2–3 of your best reports, emails, or documents as format and tone learning examples. Claude learns your definition of 'good' from these, making future output quality closer to your expectations.
What not to include:
Does the Knowledge base have file size or count limits? What if I have many documents?
Core principle for capacity management: curated > stuffed Many people assume the Knowledge base should hold as much as possible, but too many documents can actually reduce response precision — Claude has to filter through more material to find what's relevant. 'Only what Claude genuinely needs' beats 'as much as possible.'
When a document is too long: If you have a long reference document (50-page brand handbook), don't upload the whole thing. Extract the most important parts: a few key brand voice principles, the most critical format conventions, the most commonly misused terms — use this 2–3 page essence instead of the full 50 pages. Usually more effective because Claude doesn't need to search through 50 pages to find the 2–3 you actually need it to reference.
When you have too many documents: If you have multiple clients or work contexts, create a separate Project per context rather than putting all clients' materials in one Project's Knowledge base. Each Project's Knowledge base stays precise and context-specific — Claude won't see Client B's materials when answering Client A's questions.
Regular maintenance: Knowledge base materials don't manage themselves. Monthly 5-minute review recommended: what's outdated (company overview is old, product features updated), what new important material should be added.
Does Claude read all Knowledge base content every conversation? Or does it only look when needed?
Knowledge base content is loaded into Context Window when the conversation starts. Each time you open a new conversation in a Project, Claude can 'see' all your Knowledge base content — not accessed on-demand, but read in before the conversation begins.
Practical impact: every Knowledge base document consumes Context Window space. A large knowledge base (e.g., a complete 50-page document) consumes tokens that reduce space for other content you can input in the conversation. This is another reason for 'only put what you genuinely need' — not a feature limitation, but an efficiency consideration.
Potential impact of a very large Knowledge base: with a very large knowledge base, Claude may be less precise integrating all reference material than with a small one — similar to giving someone 10 documents to 'refer to' versus giving them the 2 most relevant ones; the latter usually works better.
Most practical advice: Knowledge base size and precision have an inverse relationship — smaller, more precise knowledge bases produce more consistent Claude output quality. Use 'does Claude genuinely need to know this when answering?' as the standard for what goes in — not 'whatever I have.'
Mr. Ko is the Creative Director at a digital marketing agency managing three major client accounts. Each client's brand voice, target audience, and landmines to avoid (e.g., one client absolutely cannot mention competitor names) are completely different.
Before building a Knowledge base: every time he generated content for a client, he spent the start of every prompt explaining client context: 'This client is a skincare brand with an eco-conscious core positioning, target audience is eco-aware women aged 25–40, tone should be warm but professional, don't use the word "natural" (regulatory issue), and...' This explanation alone took 1–2 minutes, and slight wording variations each time caused Claude's output to occasionally be inconsistent.
After building Knowledge bases: he created a separate Claude Project for each of three clients. Each Project's Knowledge base held:
After setup, he opens the relevant client Project and asks questions directly — no background pasting needed. More importantly, Claude's compliance rate with prohibited terms approached 100% (because the Knowledge base listed them clearly), and tone matched the quality examples consistently.
He estimates saving 30–40 minutes per week on 'pasting client background,' with significantly reduced cognitive load from not needing to mentally reconstruct each client's context every session.
The core trade-off: consistency and accuracy vs. flexibility and currency.
The Knowledge base gives Claude a stable, consistent reference framework for your work context — substantially improving output consistency. But once material is in the Knowledge base, it becomes a 'fixed framework.' Tasks that need to break out of it (e.g., needing a different tone, targeting a different audience) require explicit indication in your prompt.
The other trade-off is currency: Knowledge base material is static, requiring manual updates. If your work context changes rapidly (e.g., new brand strategy every month), maintenance cost increases accordingly.
Best Knowledge base content: relatively stable background knowledge (company background, core terminology, brand principles) — not frequently-changing data.