How to Use the Projects Feature: Build a Specialized AI with a Knowledge Base

What Is the Projects Feature? Turning One-Off Chats into a “Business AI”

Projects in Claude Pro is not merely a feature for saving conversation history. By packaging goals, tone of voice, reference materials, and work rules into each project, you can create a “specialized AI” that excels at a specific task—without needing to explain everything from scratch every time. That is the biggest value of Projects.

For example, in a typical chat, each time you must preface with something like, “You are the in-house help desk representative. Based on the attached employment regulations and expense reimbursement manual…” On the other hand, with Projects, you can lock this into the project as Custom Instructions and a Knowledge Base. As a result, answer variation decreases, and you get much higher reproducibility even for internal use.

Projects are well suited for use cases like the following.

• Help desk AI that reads internal manuals
• Sales support AI built from product specification documents
• Article-writing AI that learns your company blog tone
• Inquiry AI by department, such as legal, HR, and accounting

Basic Structure of Projects: Focus on Three Elements

Projects is easiest to understand if you think of it as three main components: the project itself, Custom Instructions, and a Knowledge Base.

A common stumbling point for beginners is assuming that once you add documents, the system will automatically become perfect. In reality, it becomes more stable when you also instruct which documents to prioritize and how to respond when something is unclear.

How to Create a Project

1. Decide the purpose first

Start with one project, one job, rather than trying to build “an AI that can do anything.” Here are example tasks.

• Internal policy Q&A
• Create reply drafts for customer support
• Review sales proposals
• Search specifications for the development team

If you expand the scope too much, both the instructions and the documents get mixed, and answer quality drops.

2. Create a new project

From the Projects screen in Claude, choose “Create new,” and enter the project name. Pick something that won’t confuse you later. A good recommendation is department name + purpose + version.

Example: HR_Internal Policy Q&A_2025Q1
Example: Sales_Product Proposal Support_Japanese

3. Set up Custom Instructions

This is the core of Projects. Simply saying “Answer politely” is weak; for real-world usability, describe the role, information sources, output rules, and escalation conditions.

Here’s an example of Custom Instructions for an HR AI that reads internal manuals.

You are an internal assistant for the Human Resources department of a Japanese company.
Your main role is to answer employee questions in Japanese clearly and accurately, based on employment regulations, leave policies, the expense reimbursement manual, and business trip rules.

Rules:
- Start by stating the conclusion in 1–2 sentences.
- Then cite the relevant regulation name(s) and specific sections.
- If the documents do not explicitly state it, do not guess—answer: “Not confirmed in the provided materials.”
- Do not make legal judgments, HR approvals, or finalize individual exceptions; if needed, encourage the user to confirm with an HR representative.
- Structure the output with three headings: “Conclusion,” “Supporting Evidence,” and “Items to Confirm.”
- Because the audience is employees, explain specialized terms as simply as possible.

The key is to standardize the “answer format”. Doing this alone significantly reduces quality variation from one response to another.

How to Upload a Knowledge Base and Tips for Organizing It

Next, add the documents you want the AI to reference to the Knowledge Base. Common formats include PDF, Word, plain text, internal procedures, and collections of FAQs. As of 2025, it’s generally better to choose and upload only the latest official versions rather than dumping large quantities of mixed materials.

Pre-upload organization checklist

• Are old versions and revised versions mixed together?
• Have you registered duplicate files (e.g., the same content in both PDF and Word)?
• Are scanned PDFs stable in terms of text recognition?
• Can you identify the content from the file name?

Here are recommended naming examples.

Employment Regulations_2025-01.pdf
Expense Reimbursement Manual_Domestic Business Trips_2025-02.docx
FAQ_Childcare Leave and Maternity/Paternity Leave_2025Q1.md

Also, in some situations, splitting by theme can make reference accuracy more stable than stuffing everything into one massive document.

How to Build a Specialized AI That Reads Internal Manuals [Case Study]

Here, we’ll concretely outline the process for building a “HR department internal inquiry AI.”

Prepare the documents

• Employment regulations
• Paid leave policy
• Work-from-home guidelines
• Expense reimbursement manual
• Frequently asked questions (FAQ)

Setup steps

1. Create a project in Projects called “HR_Internal Inquiry AI.”
2. Register the Custom Instructions described above.
3. Upload the documents above to the Knowledge Base.
4. Send around 10–20 test questions.
5. If incorrect answers appear, replace the documents or adjust the instructions.

Example questions you can use for testing

Can I take paid leave in half-day units?
Are business class/Green Car seats on the Shinkansen available to everyone during trips?
Are communication expenses eligible for reimbursement while working from home?
Can I use bereavement/marital leave even during the probation period?

What you want is not merely “a plausible-sounding response,” but an answer grounded in regulation names. If you get too many ambiguous responses, adding to the FAQ or strengthening the “cite evidence / do not guess” instruction is effective.

Operational Tips for Using It as a Team

Projects is valuable not only for individual use, but also for teams as a shared business AI. However, when you share it, operational rules matter more than convenience.

• Limit edit permissions: Letting anyone freely change instructions or documents is risky.
• Assign an update owner: Clearly define who is responsible for updating HR policies and product specifications.
• Keep version management: Record what was replaced and when.
• Check for confidential information: Be careful with personal data or documents that cannot be uploaded under contract terms.

Especially for internal use, it helps to explicitly state that the AI’s responses are not confirmed facts as-is, and that the final decision-maker is a human—this can prevent accidents.

When to Use vs. Normal Chat

If you’re unsure, consider converting tasks where you’ve had to explain the same context three or more times into Projects.

Tips and Cautions for Using Projects Effectively

• Don’t aim for perfection from the start: Start small with one department and one use case.
• Fix the answer format: Include items like “Conclusion,” “Supporting Evidence,” and “Next Actions.”
• Explicitly state “no guessing”: This is especially important for internal policy topics.
• Update documents regularly: Reliability drops when outdated materials remain.
• Evaluate with real questions: Validate not only with sample prompts, but also with actual inquiries from the field.

Projects is not a “high-performance notepad.” It is an AI environment for task units that comes with context and knowledge. With Custom Instructions defining roles, a Knowledge Base providing evidence, and shared operating rules to make it consistent, Claude shifts from being just a conversational AI to becoming a specialized assistant you can use in real work.