Claude Projects are the feature that justifies the Pro subscription. Not the extra messages. Not the priority access. Projects.
A Project gives Claude persistent memory for a specific domain. You upload your documentation, set custom instructions, and every conversation in that project starts with full context. No more re-explaining your tech stack, your writing style, or your company's terminology every time you open a new chat.
We run 6 active Projects at Like One — one for code review, one for blog writing, one for grant applications, one for legal research, one for financial analysis, and one for product strategy. This guide covers everything we have learned.
What Claude Projects Actually Are
A Claude Project is a workspace that combines three things:
- Custom Instructions — Persistent rules that tell Claude how to behave in this context. Think of them as a system prompt that never expires.
- Project Knowledge — Documents you upload that Claude can reference. Up to 200K tokens of PDFs, code files, docs, or any text.
- Conversation History — Past chats in the project that build on each other.
The combination is more powerful than any one piece alone. Custom instructions tell Claude how to think. Project knowledge tells Claude what to know. And conversation history gives Claude the thread of your ongoing work — which is a key part of building a personal AI assistant with real continuity.
Setting Up Your First Project
Step 1: Create the Project
Click "Projects" in the Claude sidebar, then "New Project." Give it a clear name that describes the domain: "API Documentation," "Q1 Marketing Plan," or "Thesis Research" — not "Work Stuff."
Step 2: Write Custom Instructions
This is the most important step. Your instructions should cover:
- Context: What this project is about and what role Claude should play
- Format: How you want responses structured (bullet points, code blocks, tables)
- Constraints: What Claude should NOT do in this context
- References: Which uploaded documents to prioritize
Example for a code review project:
You are reviewing code for our payments API (Node.js, TypeScript, PostgreSQL).
Focus on: security issues, error handling, type safety, and performance.
Reference: api-specs.md for endpoint contracts, style-guide.md for conventions.
Format: List issues by severity (critical, warning, suggestion).
Do NOT suggest refactoring code I did not ask about.
Do NOT add comments to obvious code.Step 3: Upload Project Knowledge
Upload the documents Claude needs to do its job. Good candidates:
- API specifications and documentation
- Style guides and brand guidelines
- Architecture decision records
- Sample outputs that represent your quality bar
- Glossaries of company-specific terminology
- Previous reports or analyses for format reference
Prioritize documents that contain information Claude cannot infer from training data. Your public-facing docs are probably already in the training set. Your internal specs, proprietary processes, and custom terminology are not.
Project Architectures That Work
The Domain Expert
One project, one domain. Upload everything Claude needs to be an expert in that area. Best for ongoing work in a single field: code review, legal research, medical writing, financial analysis.
Instructions focus: Domain rules, terminology, output format.
Knowledge focus: Reference materials, specifications, past examples.
The Content Pipeline
A project dedicated to content creation with your brand voice baked in. Upload your style guide, top-performing posts, and editorial calendar. Every new piece starts from your established voice.
Instructions focus: Tone, audience, format rules, SEO requirements.
Knowledge focus: Style guide, example content, keyword research.
The Research Hub
A project for ongoing research where each conversation builds on previous findings. Upload papers, data, and notes. Claude remembers what you have already explored and can connect new information to existing knowledge.
Instructions focus: Research methodology, citation format, synthesis approach.
Knowledge focus: Papers, datasets, literature review notes.
The Operations Center
A project that knows your business operations — your tools, processes, team structure, and KPIs. Use it for planning, troubleshooting, and decision support.
Instructions focus: Business context, decision-making framework, reporting format.
Knowledge focus: SOPs, org charts, tool documentation, metrics definitions.
Advanced Techniques
Layered Instructions
Your global Claude profile sets baseline behavior. Project instructions add domain-specific rules. Use this hierarchy intentionally: put universal preferences ("be concise," "show code examples") in your profile, and put domain rules ("use TypeScript," "cite sources in APA") in project instructions.
Document Prioritization
When you upload multiple documents, tell Claude which ones take precedence. "When the style guide and the API spec conflict, follow the API spec" prevents confusion on edge cases.
Conversation Starters
Set up conversation starters for common tasks within the project. Instead of typing the same prompt every time, create starters like "Review this PR," "Draft a blog post about [topic]," or "Analyze this month’s metrics."
Knowledge Refresh
Project knowledge is static — it does not update automatically. Schedule regular uploads when your documentation changes. If your API spec updates quarterly, refresh the project knowledge quarterly. Stale documents are worse than no documents because they give Claude confidently wrong context.
Token Limits and What Counts Against Them
Claude Projects allow up to 200K tokens of project knowledge. Here is what that means in practice:
- 200K tokens is roughly 150,000 words — enough for several large documents, an entire codebase, or a book-length reference.
- Every uploaded document counts against this limit. PDFs, code files, markdown, plain text — all consume tokens proportional to their length.
- Custom instructions do NOT count against the 200K limit. They consume tokens from your conversation context window instead, alongside your messages.
- Images and binary files are not supported. Only text-based documents can be uploaded.
To optimize your token budget: prioritize internal documents over publicly available ones (Claude already knows public information), remove redundant files, and use concise document formats rather than verbose ones. A well-structured markdown file uses far fewer tokens than a formatted PDF with the same content.
Common Mistakes
1. Uploading Too Much
The 200K token limit sounds generous, but quality beats quantity. Ten focused documents outperform fifty loosely related ones. Claude's attention degrades with noise. Upload what Claude needs for THIS project, not everything you have.
2. Vague Instructions
"Help me with code" is not an instruction. "Review TypeScript code for our REST API. Check for missing error handling, inconsistent naming, and SQL injection vulnerabilities. Format issues as a numbered list with file:line references" is an instruction.
3. Never Iterating
Your first instruction set will not be perfect. Pay attention to when Claude's output misses the mark and add a rule to prevent it next time. Good instructions evolve through use.
4. One Giant Project
Do not put everything in one project. A project for code AND marketing AND legal AND finance will have conflicting instructions and irrelevant knowledge polluting every conversation. Use separate projects for separate domains.
5. Ignoring Conversation Context
Each conversation within a project builds its own context. Starting a new conversation resets to just the project instructions and knowledge. Use long conversations for multi-step work. Start fresh conversations for independent tasks.
Troubleshooting Claude Projects
Claude ignores my uploaded documents
Reference specific documents by name in your instructions: "Always check api-spec.md before answering questions about endpoints." Without explicit references, Claude may not prioritize your uploads over its training data.
Responses feel generic despite custom instructions
Your instructions may be too vague. Replace "be helpful with code review" with specific rules about what to check, what format to use, and what to skip. Specificity drives behavior change.
Claude contradicts information in my documents
Add an instruction: "When my uploaded documents and your training data conflict, always prefer my documents." By default, Claude may blend both sources, which creates contradictions when your internal processes differ from industry standards.
Projects vs. Custom GPTs vs. CLAUDE.md
If you have used ChatGPT’s Custom GPTs, Projects serve a similar purpose but with key differences:
- Projects are private. Custom GPTs can be published. Projects are for your personal or team use.
- Projects support document upload. Custom GPTs have knowledge files, but Projects integrate them more naturally into conversation.
- Projects keep conversation history. Each chat in a project builds on previous ones. Custom GPTs reset every conversation.
If you use Claude Code (the CLI), CLAUDE.md files serve a similar purpose to project instructions but are scoped to a repository. They are complementary — use Projects for non-code work, CLAUDE.md for code work.
Wondering which Claude model works best inside Projects? Read our Claude Sonnet vs Opus comparison for a practical breakdown of speed, cost, and capability trade-offs.
Best Practices for Project Instructions
The quality of your project instructions determines the quality of every conversation. Here is what works after months of daily use:
- Be specific about format. "Respond in bullet points with no more than 3 sentences per point" produces better results than "be concise." Claude follows formatting instructions reliably when they are explicit.
- Define the role precisely. "You are a senior tax advisor specializing in US freelancer tax obligations for single-member LLCs" outperforms "You are a tax expert." Specificity reduces hallucination and increases domain accuracy.
- Include negative constraints. Tell Claude what NOT to do. "Never suggest incorporating in Delaware without first asking about the user's state of residence" prevents common bad advice patterns. Negative constraints are as important as positive instructions.
- Use examples. Include 2-3 examples of ideal responses in your project instructions. Claude calibrates tone, length, and detail level from examples more effectively than from abstract descriptions. Show, do not tell.
- Version your instructions. Keep a local copy of your project instructions in a text file. When you update them, note what changed and why. This prevents instruction drift where small edits accumulate into incoherent guidance over time.
For a deep dive into instruction writing, read our complete Claude custom instructions guide.
Project Templates for Common Use Cases
Here are project configurations we use daily that you can adapt:
Content Creation Project
Instructions define brand voice, target audience, formatting rules, and SEO requirements. Knowledge files include style guide, brand glossary, and top-performing content examples. Every conversation produces on-brand content without re-explaining context.
Code Review Project
Instructions specify coding standards, security requirements, and review checklist. Knowledge files include the project's architecture documentation, style guide, and common anti-patterns to flag. Claude reviews pull requests with full codebase awareness.
Research Analysis Project
Instructions define methodology, citation format, and analytical framework. Knowledge files include source papers, data sets, and previous analyses. Claude synthesizes findings across multiple sources with consistent methodology.
Client Communication Project
Instructions set tone, confidentiality boundaries, and response templates. Knowledge files include client history, project status, and communication guidelines. Every email draft matches your professional standards automatically.
The common pattern: project instructions handle the "how" (tone, format, rules) while knowledge files handle the "what" (domain content, reference material, examples). Keep them separate for clean maintenance.
Our Setup
Here is how we structure Projects at Like One:
Enterprise Implementation
Setting up Claude Projects for your organization? Our consulting services help teams design project architectures, custom instructions, and knowledge management workflows that scale.
- Code Review: Stack docs, style guide, architecture decisions. Instructions focus on security and type safety.
- Blog Writing: Brand voice guide, SEO checklist, top 5 performing posts. Instructions enforce tone, length, and keyword usage.
- Grant Applications: 501(c)(3) determination letter, program descriptions, budget templates. Instructions follow funder guidelines.
- Legal: Lease agreements, correspondence history, relevant statutes. Instructions cite specific sections.
- Finance: Revenue reports, subscription data, tax filings. Instructions use our accounting terminology.
- Strategy: Competitive analysis, product roadmap, user feedback. Instructions challenge assumptions.
Each project has 3-5 uploaded documents and 10-20 lines of custom instructions. That is it. Focused, specific, and immediately useful.
Security and Privacy Considerations
Claude Projects offer strong privacy defaults. Anthropic does not train on your conversations by default, and project data stays within your workspace. For enterprise users, the Team plan adds admin controls and data governance.
Key security practices for Claude Projects: never upload credentials, API keys, or passwords as knowledge files. Use placeholder values and reference external secret managers. Review uploaded documents for personally identifiable information before adding them to shared team projects. And remember that project instructions are visible to all team members on shared plans — do not include sensitive business logic that should be restricted.
Getting Started
Create one project today. Pick the domain where you spend the most time with Claude. Upload your two most important reference documents. Write five lines of custom instructions. Use it for a week.
You will notice two things: Claude’s outputs get dramatically better because they are grounded in your actual context, and your conversations get shorter because you stop re-explaining things Claude already knows. That is the Projects advantage. For a direct platform comparison, read our Custom GPTs vs Claude Projects comparison.