Claude Projects Tutorial: From Creation to Knowledge Base and Pinned Instruction Settings

If you want Claude to “remember” your work background and materials over the long term, Projects is the most hassle-free approach. It puts persistent instructions, reference files, and related conversations into a single workspace, reducing repeated explanations. Below, following the actual operating sequence, we’ll walk you through setting up Claude Projects and get you using it smoothly.

Enter Projects and create a workspace

After opening the Claude web app and logging in, find the Projects entry in the left sidebar; if you don’t see it yet, first check whether your account has switched to a workspace interface that supports this feature. Click Create new project, fill in the project name and a description of its purpose. It’s recommended to use “Client name/Topic name + Scenario” to make later searching easier. After creation, enter the project page—subsequent conversations will all be grouped under this project.

If you handle multiple business lines at the same time, you can create one Claude project for each line, instead of piling all files into a single conversation. This makes the context Claude references more stable and reduces the chance of mixing topics.

Set Project Instructions so Claude consistently outputs in your style

In the project settings, find Project Instructions and write in the rules you want Claude to follow long term—for example: output format, tone, whether to give the conclusion first and then expand, default terminology, and so on. This is better suited for “long-term, stable” requirements; temporary needs are more flexible to put into each individual conversation. After saving, new project chats will automatically include these instructions.

To avoid Claude misunderstanding, make the instructions as specific and verifiable as possible, such as “Keep the conclusion within 3 sentences, then provide steps,” and “When citing sources, include the file name/chapter.” When you notice Claude starting to drift, prioritize making small adjustments in Project Instructions—this is usually more effective than repeatedly correcting it in the chat.

Upload materials to the project knowledge base to build reusable background information

In the project’s Knowledge/Materials area, upload files or paste text materials. Most common document and text formats can be recognized; if you exceed platform limits, Claude will directly prompt that it can’t upload or that parsing failed. After uploading, you can use a few test questions to confirm it read the key content—for example, ask Claude to restate the main points of a specific paragraph or extract a glossary. When you have lots of materials, it’s recommended to split files by topic rather than cramming dozens of pages into one large document.

Note that Claude will try to find support from the project materials, but that doesn’t mean it will fully cite every file each time. If you explicitly specify in your question, “Please use XX file in the project knowledge base as the source of truth,” the hit rate will improve significantly.

Start conversations in a project, switch scenarios, and common small issues

After entering a project, you can start a new conversation directly—Claude will automatically include the project instructions and materials background. If you want to temporarily step outside the project rules, starting a new chat back on the regular chat page is the cleanest option. Project chats are suited for continuous iteration: repeatedly revising the same materials, ongoing Q&A with the same messaging, and multi-round breakdown of the same requirements. You can also use project names to distinguish different clients, reducing the time spent copy-pasting background information.

If you run into “materials have been uploaded but Claude answers as if it didn’t see them,” first confirm whether your question is specific enough, then try having Claude produce a “materials-based summary/index” first. If you encounter upload failures or parsing anomalies, first switch to a more standard/clean file version (for example, export as a clear PDF or plain text), then re-upload it to the Claude project—this usually restores normal behavior.