When I came back from my break a few weeks ago, I was trying to get back up to speed on projects. I had a a bunch of Notion pages open, Claude open and struggling to respond, and then — the dreaded notification: "This conversation has reached its maximum length…" (if you use Claude, you know the rest)

That's when it hit me: we now have three completely different ways to give Claude information — Projects, MCPs, and the newly released Skills. I work in an environment where we use AI everyday and have our own AI product, and even I get confused about when to use which one.

If I'm struggling with this (and I spend way too much time thinking about AI), then small teams and business owners are definitely drowning in it.

Imagine this scenario, I am sure something similar has caused you trauma recently…

You just spent 20 minutes uploading 50 documents to a Claude Project. Marketing guidelines, past campaign reports, customer research, competitive analysis—everything Claude might need.

You type your first message: "Help me draft our Q1 strategy."

Message: “This message exceeds the length for the conversation”

Here's what just happened: You gave Claude the equivalent of three novels to hold in its working memory before it even started thinking. It's like asking someone to juggle 47 balls while solving a math problem.

Here's why this matters right now:

AI usage is accelerating. You're probably using Claude more this month than you did all of last quarter. And just like budgeting money, if you don't budget your context window, you'll blow through it and wonder why Claude suddenly seems dumber or keeps cutting you off mid-conversation.

The difference is: With money, you see your bank balance dropping. With context windows, you just get frustrating error messages or weirdly generic responses and have no idea why.

Most working professionals and small business owners don't understand tokens. Honestly, they may never have even heard the terms “tokens” or “context window”, so they don't have a strategy for organizing information. They either:

• Dump everything into Projects and wonder why Claude seems "slower" or hits limits

• Use nothing and retype the same context every conversation

• Randomly pick tools without understanding the tradeoffs

This isn't just inconvenient. It's expensive. Whether you're using the web interface or the API, you're working with the same context limits. On the web, you hit "message limits" or get told "this conversation is too long." That's your context window running out. With the API, you pay per token. Either way, the problem is the same: too much loaded into Claude's brain at once.

So…No One Told You About Context Windows

Think of Claude's context window like RAM on your computer. It's working memory—everything Claude can "see" and think about in a single conversation.

Here's some math that should clarify what I mean:

• 150 pages of documents ≈ 75,000-150,000 tokens

• Claude's context window = 200,000 tokens

• Your actual working space = What's left after loading all your information

Load those 150 pages into a Project? You've used 50-75% of Claude's brain before the conversation even starts. Now you've got 50,000-125,000 tokens left for:

• Your actual questions and requests

• Claude's responses

• Back-and-forth conversation

• The work you're trying to produce

The part that is counterintuitive when you don’t understand context windows: Giving Claude more information upfront often makes it perform worse. It's cognitively overloaded before you even ask your first question.

Three Approaches to Sharing Information With Claude (And When Each One Wins)

• Claude Projects = Front-loaded, always-there context

• MCPs (Model Context Protocol) = On-demand retrieval via tools

• Claude Skills = Specialized workflows + knowledge combos

Think of it this way:

• Projects: The sticky notes on your desk you glance at constantly

• MCPs: The filing cabinet you search when you need something specific

• Skills: The Standard Operating Procedure manual that walks you through a process

Three Questions To Determine What To Use When

Before you put anything into Projects, MCPs, or Skills, ask:

Question 1: "Does Claude need this RIGHT NOW, or just SOMETIMES?"

RIGHT NOW = Projects

• Your brand voice guidelines (referenced in every piece of content)

• Core company information (mentioned constantly)

• Personal preferences/communication style

• Templates you use daily

SOMETIMES = MCPs

• Archive of past work (search when relevant)

• Large research repositories

• Historical data/records

• Extensive documentation

Real example:

A marketing consultant had 80 past client case studies in a Project. Claude was sluggish and kept hitting limits.

• The fix: Moved case studies to Notion (accessible via MCP). Kept a 5-page "greatest hits" summary in the Project.

• Result: 90% faster responses, no more context errors, and Claude could actually search for relevant cases instead of drowning in all of them.

Question 2: "How big is this, really?"

Do the napkin math:

• 1 page ≈ 500-1,000 tokens (rough estimate)

• Under 25 pages → Projects probably fine

• 25-100 pages → Consider splitting or using MCPs

• 100+ pages → Definitely use MCPs

Why this matters:

If you keep hitting message limits on the Pro plan, that frustrating "you've used your allocation" message. Restructuring your approach (Projects for core guides, MCPs for archives), those limits will stop being an issue. Same work, way less friction.

Question 3: "Is this stable or constantly changing?"

Stable = Projects

• Company mission/values

• Product specifications (until next release)

• Writing style guides

• Core methodologies

Constantly changing = MCPs

• News and updates

• Customer data

• Project status reports

• Market research

Why? Projects load once per conversation. If your data changes daily, you're working with stale information. MCPs fetch fresh data every time.

Skills sit in a different category: Use them when you have a repeatable process that needs specific tools. For example:

• "Competitive analysis workflow" (searches web + your database + creates report)

• "Customer research synthesis" (pulls from multiple sources + applies framework)

A Smart Hybrid Approach

Here's the setup that works for a boutique consulting firm with 5 employees:

In Projects (15 pages total):

• 3-page company overview

• 4-page service descriptions

• 5-page client communication guidelines

• 3-page proposal template

Via Notion MCP:

• 200+ past project reports

• Client research database

• Industry analysis archive

• Meeting notes and recordings

As a Skill:

• "Client Discovery" workflow (searches past similar projects + applies proprietary framework)

The result: Claude has constant access to core identity but can search vast archives as needed. Context window stays clean. Responses stay fast. Limits stay manageable.

Here’s Exactly How This Works

I've worked through this exact challenge with my own setup, so let me show you what works. Here's how a marketing consultant should structure their information:”

In Your Project (12 pages):

1. "About Me" (2 pages)
   - Your approach to marketing
   - Your specific methodology
   - Services you offer
   - Types of clients you work with

2. "How I Communicate" (2 pages)
   - Your writing voice
   - Email templates you use
   - Proposal structure preferences
   - Client communication guidelines

3. "Core Frameworks" (8 pages)
   - Your customer research process
   - Your content strategy framework
   - Your campaign planning checklist
   - Key questions you always ask clients

In Notion (accessible via MCP):

- Database of 50+ past client projects
- Industry research you've collected
- Competitor analysis reports
- Case study archive
- Meeting notes from client calls

Why this works:

When you ask Claude to "draft a proposal for a B2B SaaS client," it already knows:

• How YOU write proposals (from the Project)

• Your methodology (from the Project)

• What to include (from your templates in the Project)

When you ask "what approaches worked for similar clients?", Claude searches your Notion database via MCP and pulls relevant examples.

You get personalized responses every time WITHOUT overloading Claude's working memory.

Four Mistakes That Will Cost You Money

Mistake #1: The "Everything Drawer"

Putting every document "just in case Claude needs it." This is like bringing your entire closet on a weekend trip.

Fix: Audit what Claude actually references. You'll find 80% of your Project content never gets used.

Mistake #2: The "Bare Minimum"

Using no Projects or MCPs, retyping context every conversation. You're paying for repetitive token usage.

Fix: Create a small Project with your true essentials (under 10 pages). Use it as your baseline.

Mistake #3: The "Wrong Tool for the Job"

Putting searchable databases in Projects or frequently-needed guides in MCPs.

Fix: Use the 3-question framework above.

Mistake #4: Not Testing

Never checking how much of your context window you're actually using.

Fix: In a new conversation, ask Claude: "How much of your context window am I currently using?" (Note: Claude can't see exact percentages, but you can estimate based on your token counts)

Where Skills Fit (And Why You Should Care Even If You Don't Have Them Yet)

Skills are the newest addition, and here's why they matter: They're like teaching Claude a specific workflow that combines multiple tools and knowledge sources.

The Matrix Analogy that Makes This Click

Remember when Neo needed to fly a helicopter? Trinity: "Tank, I need a pilot program for a B-212 helicopter." Upload. "Let's go."

That's what Skills do. They're pre-loaded programs that give Claude instant, specialized capabilities when you need them.

Think of it this way:

• Projects = What Claude knows about YOU

• MCPs = Where Claude looks up information (like Tank uploading knowledge when needed)

• Skills = Pre-built programs for HOW Claude executes specific types of work

Even if you don't have custom Skills yet, understanding this third category helps you think about your information architecture differently.

When Skills become available to you (or as you start building them), think about repeatable processes that need:

1. Multiple information sources

2. Specific tools or APIs

3. A consistent workflow

Examples:

• Content creation that pulls from your brand guidelines (Project) + past performance data (MCP) + applies SEO tools

• Customer research that searches multiple databases + applies your analysis framework

• Competitive analysis that monitors web sources + compares to your internal data

Skills are powerful because they're like loading programs into Claude to give you instant access to specialized capabilities when you need them.

Start With A 5-Minute Audit

I know, that's a lot of new information. So here's exactly where to start: a 5-minute audit that'll show you what needs to change. Do this right now:

1. List everything currently in your Claude Projects

   Add up approximate pages (documents × average pages per doc)

2. Honestly answer: How often does Claude reference each item?

   • Every conversation = Keep in Project

   • Weekly = Keep in Project

   • Monthly or less = Move to MCP

   • Never = Delete it

3. Check for the "search candidates"

   • Large archives? → MCP

   • Historical data? → MCP

   • "Just in case" documentation? → MCP

4. Identify your true essentials

   • What does Claude ALWAYS need to know about you/your business?

   • Aim for under 20 pages total

   • This is your Project core

5. Create your hybrid setup

   • Core essentials → Projects

   • Searchable archives → Notion/Google Drive (accessible via MCP)

   • Repeatable workflows → Skills (if you have them)

The Real Impact

One of the frustrations I hear the most from business owners we work with at Hello Alice is "It keeps giving generic advice" or "It's not understanding my product."

The problem? They are at one of two extremes. They have nothing in projects or they have too much in project, because no one has explained the context window to them. So every conversation either started from zero, or with an overload.

When I walk them through a simple way to organize projects Claude's responses transformed overnight. It stopped making suggestions that violated their brand guidelines. It referenced their specific user segments. It understood their service and delivery philosophy.

The difference between having good information architecture and having none isn't just efficiency. It's whether Claude can actually be useful at all.

The Bottom Line

You don't need to become a token-counting expert. You just need to stop treating all information the same way.

The mental model that changes everything:

• Projects = What Claude always needs to remember about you

• MCPs = What Claude might need to look up

• Skills = How Claude should do specific types of work

Get this right, and Claude stops feeling like a tool you have to constantly brief. It starts feeling like a team member who actually knows your business.

Get it wrong, and you're either wasting resources on unused context or retyping the same background information 47 times.

Here's what I want you to do this week:

Open your Claude Projects right now. Scroll through everything you've loaded in there. Ask yourself: "When was the last time Claude actually referenced this?"

If the answer is "never" or "I don't know"—delete it or move it to an MCP.

You'll be shocked at how much faster Claude responds. How much more specific its answers become. How it suddenly "gets" what you're asking for.

That's not because Claude got smarter. It's because you stopped drowning it in information it didn't need.

The difference between a mediocre AI assistant and an indispensable one isn't the technology. It's how you organize the information you give it.

I

Start with your "core 10". Before loading anything into Projects, write down the 10 things Claude absolutely needs to know about your business to be useful. Turn those into a single 10-page document. That's your foundation. Everything else is optional.

II

Create a "reference" document instead of uploading everything. Instead of uploading 50 case studies, create a 3-page summary with the patterns and lessons. Include links to the full versions in Notion. Claude gets the insights without the cognitive load.

II

Test in a clean conversation. Once a month, start a fresh conversation outside your Project. Ask Claude to help with your typical work. Notice what information you have to provide. That's what belongs in your Project. Everything else is noise.

I

"If I only had 10 pages to give Claude, what would they be?" This question forces prioritization. You can't hedge. You can't say "well, this might be useful." You have to pick what actually matters. Most people discover their current Project could be 70% smaller without losing any value.

II

"Am I organizing this for Claude's brain or for my filing system?" Just because your documents are organized a certain way on your computer doesn't mean that's how Claude should access them. Your annual reports from 2018-2024? That's a filing system (MCP). Your current product positioning? That's Claude's brain (Project). Don't confuse the two.

Your context window is a budget, not a storage unit.

Every page you load into Projects is a withdrawal from Claude's working memory. Most people treat their Projects like cloud storage—dump everything in and hope Claude figures it out.

But Claude's context window isn't storage. It's more like RAM. It's active processing space. When you overload it, Claude doesn't just slow down—it starts missing things, giving generic responses, or hitting limits.

The teams getting the most value from Claude understand this: Less context, better organized, strategically accessed = smarter outputs.

Think about your own brain. You can't hold 50 documents in your working memory. You remember key principles and look up specific details when you need them. That's exactly how you should structure Claude's information access.

• Projects = Working memory (what you remember)

• MCPs = Reference library (what you look up)

• Skills = Procedural knowledge (how you do things)

Get the balance right, and Claude transforms from a tool that sometimes helps into a team member that consistently delivers.

P.S. - If you found this useful, the single best thing you can do is forward it to one person on your team who's also using Claude. The number of people still dumping their entire Google Drive into Projects is... concerning.