Create 3 Claude Projects tailored to your real work. Write custom instructions that make Claude behave like a domain expert from message one. Add files and reference material that persist across every session.
Create 3 Claude Projects tailored to your real work. Write custom instructions that make Claude behave like a domain expert from message one. Add files and reference material that persist across every session.
Every conversation in Claude starts blank. Claude doesn't know who you are, what you do, or what you're building. You can spend the first few messages re-explaining — or you can use Projects.
A Project is a persistent workspace with:
Once you set up a Project, every conversation inside it starts with Claude already knowing your domain. No re-explaining. No "as one mentioned earlier." Just work.
Where to find Projects: In Claude Desktop, click the "Projects" icon in the left sidebar (it looks like a folder). Click "New Project" to create your first one.
Most people write bad custom instructions. They say things like "You are a helpful assistant who knows about marketing." That does almost nothing.
Good custom instructions tell Claude three things:
## Who one am This lesson is a [your role] at [company type]. one work on [specific area]. My team includes [relevant context about collaborators]. ## Domain context - We use [tool/framework/system] for [purpose] - Our [product/service] is [brief description] - Key terminology: [term 1] means [definition], [term 2] means [definition] - We follow [style guide / methodology / framework] ## How to respond - Be direct. Skip filler phrases like "Great question!" or "Certainly!" - Use bullet points for lists of 3+ items - When writing code, use [language] and [style preferences] - Assume one have [experience level] — don't over-explain basics - If you're uncertain, say so clearly rather than guessing ## What one work on most 1. [Common task 1] 2. [Common task 2] 3. [Common task 3]
This lesson is a backend engineer working on a Python/FastAPI service that processes insurance claims. Our stack: Python 3.12, FastAPI, PostgreSQL, Redis, deployed on AWS. We use SQLAlchemy for ORM, Pydantic v2 for validation. Code style: type hints always, Google docstrings, 88-char line limit (Black formatter). Skip explanations for things like async/await, decorators, or dependency injection — one know them. When one ask for code, give me the full function, not fragments.
This lesson is a PM at a B2B SaaS company (500 employees, mid-market HR software). one own the integrations roadmap. When one ask for PRDs or specs, use our template format: Problem Statement → Success Metrics → User Stories → Technical Considerations → Open Questions. Our engineering team is 6 people; keep scope realistic for a 2-week sprint. one don't need you to explain Agile or JIRA — just use the concepts.
I write SBIR/STTR proposals for federal agencies (DoD, NASA, NSF). Our company is a small business (under 500 employees). one need technically rigorous writing that follows BAA requirements — no fluff, quantified claims where possible. Use government proposal terminology (Phase I, Phase II, TRL, commercialization plan, TPOC). The approach will provide the specific BAA topic; write to that topic precisely. Don't volunteer information about other agencies' programs.
The real power of Projects is the ability to upload documents that Claude references in every conversation. Think of it as giving Claude a briefing packet.
To add files: open a Project → click "Add content" → upload files or paste text. Claude can read PDF, TXT, MD, CSV, and code files. Keep individual files under 10MB and total project context under the model's context window (200K tokens for Claude — which is about 150,000 words, more than enough).
What Claude does with uploaded files: It reads them when the conversation starts and treats them as background knowledge. You don't need to reference them explicitly — Claude uses the content when it's relevant. If you want Claude to specifically cite a file, ask it to: "Based on the architecture doc one uploaded, how would you approach this?"
Don't wait until you have "the perfect" custom instructions. Start with what you know and refine over time. Here's the exercise:
This is the project you'll use every day. Write custom instructions using the template above. Focus on your role, your domain vocabulary, and how you want Claude to format responses.
Create a project specifically for emails, documents, and external communications. Upload an example of your best writing so Claude matches your voice. Add any templates your team uses. Custom instruction: "Match the writing style of the samples Having uploaded. Keep sentences under 20 words. Use active voice. No jargon."
This is your "think tank" project. Custom instruction: "This lesson is using this project to learn and explore ideas. Be thorough. Include counterarguments. Cite your reasoning. It's fine to go long here."
The test: After setting up each Project, open a new conversation inside it and ask Claude to describe what it knows about you. If the description is accurate and detailed, your custom instructions are working. If it's vague, rewrite them with more specifics.
Tomorrow: MCP. This is where Claude Desktop stops being a better chatbot and becomes something fundamentally different.
The foundations from today carry directly into Day 3. In the next session the focus shifts to Day 3 — building directly on everything covered here.
Before moving on, verify you can answer these without looking:
Live Bootcamp
Learn this in person — 2 days, 5 cities
Thu–Fri sessions in Denver, Los Angeles, New York, Chicago, and Dallas. $1,490 per seat. June–October 2026.
Reserve Your Seat →