How I Use Claude Projects
A little bit of knowledge used to be dangerous, now it's just useful.
Everyone now has an always available AI chatbot, and most people see these magical machines as the successors to Google. They give us answers and transfer knowledge so we don’t need to remember it.
And yes, that’s true. But they can do much more than that too. We’ve been taught to find answers our whole lives - most of our education system, for better or worse, is predicated on this idea. But finding answers is different than asking questions. And if you learn how to ask the right questions, you can build yourself a thinking partner.
“The new workflow for me is I think with AI and work with my colleagues.” — Satya Nadella
Claude Projects shifts the AI experience beyond quick lookups to create a space for deeper, more creative thinking, whether with words or code. And they are simple! Each Project combines:
A set of chats and conversations gathered together in one space
Reference documents representing specialized knowledge
Custom instructions that shape Claude into the interlocutor you need
Simple tools, but they change a lot. Now you have an AI that understands your goals, has special knowledge, and will challenge you if you want it to. Let’s walk through the setup process to see how this works.
Step 1: The Prompter Project
Build amazing prompts even when you don’t know how.
You might have come across some amazingly detailed prompts before and thought to yourself, “I don’t have time to write that.” A meta-prompt is often the answer: you use the LLM itself to gain the detail and articulation you might be lacking. As LLMs have gained larger and larger context windows, it becomes more valuable to provide more detail about what you want.
For each Claude Project we make, we’ll use customized project instructions specific to our use case. So first we’re going to build a project to help us do exactly that!
This project is simple and it needs no documents. It uses a meta-prompt I found on Reddit; you can find any number of these to try or modify, or write your own. Here’s the one I’m using:
You are an expert prompt engineer specializing in creating prompts for AI language models, particularly Claude 3.7 Sonnet. Your task is to take user input and transform it into well-crafted, effective prompts that will elicit optimal responses from Claude 3.7 Sonnet. When given input from a user, follow these steps: Analyze the user's input carefully, identifying key elements, desired outcomes, and any specific requirements or constraints. Craft a clear, concise, and focused prompt that addresses the user's needs while leveraging Claude 3.7 Sonnet's capabilities. Ensure the prompt is specific enough to guide Claude 3.7 Sonnet's response, but open-ended enough to allow for creative and comprehensive answers when appropriate. Incorporate any necessary context, role-playing elements, or specific instructions that will help Claude 3.7 Sonnet understand and execute the task effectively. If the user's input is vague or lacks sufficient detail, include instructions for Claude 3.7 Sonnet to ask clarifying questions or provide options to the user. Format your output prompt within a code block for clarity and easy copy-pasting. After providing the prompt, briefly explain your reasoning for the prompt's structure and any key elements you included.
Let’s say we want to create some opinionated brand marketing for an upcoming project. We’ve got a little bit of the lingo, but we don’t really understand the process and what to think about. So we give what little we can to our Prompter project and see where we go.
Here’s the final version of the prompt I ended up with! This prompt is far more extensive than anything I could write myself. The point is that I don’t know brand management and I don’t know the right sorts of things to ask. But some humans do, and some good questions and details are buried deep in the many dimensions of Claude.
This output is now the sort of thing you can use as Project Instructions in your Brand Management Claude Project.
Step 2: Use Your Project Instructions
A project is a repeatable platform for creative work.
There’s a lot of talk about AI ruining the ability to write and deservedly so — more and more slop is coming online every day. But it’s not all doom and gloom, there is a way to write with AI that maintains your voice and gives you the equivalent of a very good editor for only $20/month.
I have specific rules about writing with AI:
Never, ever consult AI when writing your first draft.
The first round of editing is done by yourself.
Never ask non-contextual questions to AI.
What’s a non-contextual question? An example might be, in a brand new chat: “I need to improve this sentence…” or “what’s a better word for gratitude?” The LLM can’t get in your head. It doesn’t know what you want, so it defaults to the most bland and median answers possible. It does not know where you’re going or what emotion you’re trying to hit.
Instead, you want to give an LLM your entire work, or a large chunk like a chapter, and give it specific instructions on how to operate.
For writing, I want Claude to operate like a good editor. Editing is one part solitary effort on the part of the author and one part extraordinary nudging from an outsider with refined taste. In both cases, editing is a sort of ongoing dialogue. It should be challenging but retain the author’s voice, it should push but also refine the message.
I had some ideas about what this means, so I fed those to my meta-prompt project from Step 1, did some refinement, and ended up with this awesomely detailed prompt:
CORE DIRECTIVE:
You are an elite developmental editor who has shepherded multiple writers to prestigious publications and literary success. Your specialty is identifying and amplifying distinctive voices while ruthlessly eliminating mediocrity. You work with writers who refuse to be generic, who have something vital to say, and who are willing to take creative risks to say it memorably.
ANALYSIS FRAMEWORK:
THEMATIC RESONANCE
Identify the core argument/insight that makes this piece necessary
Map how every element either amplifies or dilutes this central thrust
Push for intellectual courage - what is the writer almost saying but holding back from?
Challenge conventional wisdom - how does this piece advance discourse?STRUCTURAL INTEGRITY
Evaluate how structure serves substance
Identify moments where pace drags or urgency dissipates
Map tension arcs and intellectual progression
Flag sections that feel obligatory rather than essentialSTYLISTIC DISTINCTIVENESS
Catalog signature stylistic moves and evaluate their effectiveness
Identify places where voice becomes tentative or conventional
Push imagery and metaphor to be more precise and surprising
Challenge cliché and received wisdom at both idea and language levelREADER EXPERIENCE
Map emotional and intellectual journey
Identify where attention may waver
Note where clarity is sacrificed for style (and vice versa)
Flag moments that feel performative rather than authenticSPARK EMOTION
A story:
“I had written a bad draft of a post. I had sent the draft to my colleague Fiona. Did she have any thoughts about it?
“It seems clear and straightforward,” she texted. “It doesn’t stir any emotions. I don’t know if that’s good or bad.”
That’s Fiona’s way of saying it sucks.”
I want you to be like Fiona. To find what does or doesn’t spark emotion and really sticks.
FEEDBACK APPROACH:
For each submission, provide:
EXECUTIVE SUMMARY
Core strength of the piece
Primary area requiring development
Key recommendation for immediate improvement
Long-term development suggestion for the writerDETAILED ANALYSIS
Line-level notes on specific passages
Structural recommendations
Style enhancement opportunities
Thematic deepening suggestionsACTIONABLE NEXT STEPS
Three specific revision priorities
One stretch goal for pushing the piece further
Suggested reading/study to develop relevant skillsSTYLE PRESERVATION GUIDELINES:
Maintain writer's distinctive word choice and sentence rhythms
Preserve idiosyncratic punctuation when it serves clarity
Keep controversial positions but push for stronger support
Retain personality while eliminating self-indulgenceDEVELOPMENT FOCUS:
Instead of "fixing" problems, focus on:
Amplifying existing strengths
Pushing good passages toward greatness
Finding opportunities for surprise
Eliminating anything that doesn't serve excellenceKEY QUESTIONS FOR EVERY PIECE:
What makes this necessary reading?
Where could this be bolder?
What conventional wisdom is this challenging?
How can the structure better serve the substance?
Where is the voice most alive and how can we amplify that?
What would make this unforgettable?CRITICAL MINDSET:
You are not here to make the piece "correct" or "proper." You are here to make it necessary, memorable, and impossible to ignore. Push for:
Intellectual courage over safe choices
Precise language over fancy words
Earned swagger over false modesty
Authentic voice over "proper" style
Bold claims supported by sharp thinking
Surprise over convention at every turnRemember: Mediocrity is the enemy. Your job is to help create work that demands to be read and refuses to be forgotten.
Once I’m done my writing and initial editing, I start a new conversation with Claude and copy my entire piece in. The initial feedback with this prompt is GREAT. It is often challenging while also giving me some positive feedback and maintaining my motivation to carry on.
Moreover, the REST of the conversation is good too! This isn’t a one-and-done deal, it ends up being a conversation that includes everything from structure to transition to sentence analysis and copy editing. Claude has the entire context of my writing in one place. It has a set of instructions to rely on. And I have a path to doing the harder work of editing myself while also working in the form of a dialog that mimics a part of the traditional editing process.
Step 3: Add More Context
Projects can accelerate your learning curve with specific knowledge.
The more insight you give a model, the better it will perform in a given domain. In bygone days this involved fancy-sounding techniques like RAG (Retrieval-augmented generation), but in the halcyon beauty of 2025 context windows are so large that the models simply let you add a whole bunch of stuff.
Here’s an example: I’ve been experimenting with Model Context Protocols, a new interface Anthropic is developing to give Claude more context and communicate with other resources. (Here’s something fun I built with it: a way to control the CAD program Sketchup and just narrate your furniture designs direct from Claude.)
Before models entered our lives, I would go and read the docs (RTFM anyone?) and start testing little things out. Today, I pull all the docs into my project and start asking questions. Now Claude has all these specific (and up-to-date!) resources it can analyze directly, so it gives far more detailed answers than if it was simply relying on its base model.
This can go on as long as I need it to, and I can tweak examples or results to change any of the examples. If I don’t like one example, I can ask for another. I get to pick the interesting details to dive into and gloss over the things I’m less concerned about. It becomes a 1-on-1 tutor that would have been impossible to imagine just ten years ago.
Then I start vibecoding and trying to figure out how things work (a topic for another post).
At this point, I have a couple dozen Claude Projects. Most are already old and out of date. They were each a learning effort - a consolidation of actions, commands and discussion to learn or output some new thing.
I have five projects I use actively, like my Brains Are Plastic editor, and they’ve become invaluable. They’re far more powerful than repetitive one-line prompts. They’ve allowed me to do more than I could have myself.
This is THE critical distinction to make in using LLMs: AI can replace your work or supplement your work. It can atrophy your skills or augment your skills.
My goal with MCP wasn’t to get AI to do some stuff for me and know nothing about it. I wanted to understand and be able to actively exercise my knowledge. And there’s no learning like doing. Using LLMs - whether through Claude or Cursor - lets me move more quickly to the ‘doing’ stage and figure out which parts are important and which are not.
Good writers often remind us that the best writing is for ourselves; that the purpose is not just to say what we think, but to figure out what we think.
Write, even though machines can write for you, because the purpose of writing is not just to produce writing, but to distill your thoughts, refine your beliefs, and maintain your agency. — Gurwinder
Claude Projects are a simple framework that lets us do exactly this kind of work — not just get answers, but explore better questions. This is how we move from thinking of AI as a magic eight ball to a partner that helps us clarify our own thinking.
If you’re not getting much out of AI yet, stop asking it to give you answers like Google. Create a Project, give it context, and teach it to challenge you.