the Basics
Taking your AI skills to the next level — practical tools, real strategies.
Brendon Frazer — May 28, 2026
Claude's New Tools
Cowork & Design — built for professionals
MCPs & Context
Extend AI's reach and understand its limits
The Database Problem
When AI should write the tool, not be the tool
Agents & Safety
Automated workflows — and the guardrails they need
A real-time AI collaborator that works alongside you in a shared workspace. Think of it as a teammate who can see what you're working on and contribute in real time.
Shared Canvas
Work together on documents, spreadsheets, and code in real time — Claude sees your edits as you make them
Proactive Suggestions
Claude doesn't just wait for prompts — it notices patterns, spots issues, and offers improvements while you work
Context-Aware
Understands the full project context. Ask follow-up questions without re-explaining your setup
Think about collaborative tasks: drafting proposals, reviewing contracts, refining marketing copy. Cowork turns those from back-and-forth into side-by-side.
Create visual content directly inside Claude. Landing pages, presentations, dashboards, and prototypes — described in words, delivered as designs.
Describe → Design
Tell Claude what you need in plain English. It generates polished visual layouts, complete with styling and structure.
Iterate Instantly
Don't like the color scheme? Want more whitespace? Just say so. Refine designs through conversation, not design tools.
This presentation was built using AI. Real estate flyers, event pages, internal dashboards — anything visual that used to require a designer for a first draft.
Model Context Protocol — a standard that lets AI connect to your tools. Think of MCPs as plugins that give AI access to your calendar, email, files, databases, and more.
You
Ask a question or give a task
AI + MCP
Connects to external tools via protocol
Your Tools
Calendar, email, files, databases, CRM
MCP vs API — What's the Difference?
API: A way for developers to write code that talks to a service (e.g., pulling weather data into an app). Requires programming.
MCP: A plug-and-play connection that lets AI use a service directly — no coding needed. Think of APIs as the engine, MCPs as the steering wheel.
Both platforms are adding ways to connect AI to external tools. Here's how they compare.
| Feature | ChatGPT | Claude |
|---|---|---|
| Approach | Actions & Connectors | Model Context Protocol (MCP) |
| Setup | Toggle on from a marketplace — like an app store | Configure via settings or desktop app — more hands-on |
| Who Makes Them | Vetted by OpenAI, official partners | Open-source community, anyone can build one |
| Categories | Productivity (Slack, Notion), data (Sheets, Zapier), search | Dev tools (GitHub, DBs), files, search, custom/internal tools |
| Customization | Use what's available — limited configuration | Build your own, connect to internal systems, full control |
| Best For | Non-technical users wanting quick integrations | Teams wanting deep, custom workflows |
Every AI has a 'context window' — the amount of information it can hold in its working memory at once. Think of it as the size of the AI's desk.
What does this mean in practice?
128K tokens ≈ a 300-page book. Sounds like a lot, but complex tasks eat context fast. Long conversations, large documents, and detailed instructions all compete for the same space.
Bigger doesn't mean better
Even models with 1M+ token windows (like Claude Opus 4.7) start hallucinating more reliably past the ~200K mark. More context = more room for the AI to get confused, fill in gaps with made-up details, or lose track of instructions. Bigger windows are useful, but reliability drops before you hit the limit.
Context windows are getting bigger, but some tasks will always exceed them. Here's how to work around the limits.
Chunk Your Input
Break large documents into sections. Process them one at a time. Combine the results.
Summarize First
Ask AI to summarize each section, then work with the summaries for analysis and comparison.
Be Specific
Don't paste everything. Give AI only what it needs for the current question. Less noise = better answers.
Change the Approach
Sometimes the answer isn't fitting more in. It's asking AI to build a tool that processes the data for you.
You have a massive Excel file — thousands of contacts. Your database is full of issues that need cleaning.
.con
Instead of .com
email @gmail
Random spaces in emails
John Smith ×3
Duplicate entries
5551234567
Inconsistent formatting
Why you can't just paste it into AI
It's not just about row count — it's rows × columns. A "small" CSV with 3,000 rows and 25 columns is 75,000 cells. Each cell is a piece of context. That modest spreadsheet can easily consume 500K+ tokens — exceeding most context windows and well past the reliability threshold. Scale that to 50K rows and you're looking at millions of data points. No AI can process that reliably in-context.
Instead of making AI process your data, ask it to write a script that does the processing. AI is better at building tools than being the tool.
Paste the database into AI
Hits context limits. Misses patterns. Inconsistent results. Can't handle 50K rows.
Ask AI to write a cleaning script
Handles any size. Consistent rules. Catches every instance. Reusable on future data.
Before you run anything
Always run scripts on a copy — never your master file. If something goes wrong, your original data is safe. If you're not comfortable reading code, have someone who is review it first. AI-generated scripts are usually solid, but they can make assumptions about your data structure that aren't right. Trust but verify.
An agent doesn't just answer questions — it takes action. It can use tools, make decisions, and complete multi-step tasks on its own.
You ask, it answers
One exchange at a time. No tools. No follow-through.
You delegate, it executes
Multiple steps. Uses tools. Makes decisions. Completes the task.
What's a repetitive task in your business you wish someone else would just handle?
You don't need to code. These three platforms make it easy to create your first agent.
ChatGPT Custom GPTs
Build a specialized chatbot with custom instructions, files, and tools. Share it with your team or keep it private.
Claude Projects
Create a project with persistent context, custom instructions, and uploaded files. Your AI workspace for any recurring task.
Gumloop
Visual drag-and-drop agent builder. Describe what you want in plain English with Gummie and it builds the automation.
AI tools are getting more powerful — and more autonomous. Here's what that looks like, and why caution matters.
What is OpenClaw?
An open-source AI assistant that runs 24/7 on your computer. You talk to it through WhatsApp, Telegram, or text. It manages your email, calendar, files, and tasks — and builds new capabilities on its own.
Why It Matters
310K+ GitHub stars. Fastest-growing open-source AI project. The creator was hired by OpenAI within 4 months. This is the direction AI assistants are heading — always-on, deeply integrated into your life.
The Security Reality Check
Cisco's security team called OpenClaw "a security nightmare." China restricted government use. The tool can access everything on your computer — files, emails, passwords, browsing history. This applies to all autonomous AI tools: always contain what they can access, run them in sandboxed environments, never give full system access, and review what permissions you're granting. The more powerful the tool, the more important the guardrails.
Power without containment is a liability. Always ask: what can this tool access, and what's the worst it could do?
- Claude Cowork and Design put AI inside your workflow, not beside it.
- MCPs turn AI from a chatbot into a connected tool that works with your systems.
- Know your context window limits — bigger doesn't mean more reliable.
- When data is too big for AI, ask it to build the tool instead. Always run on copies.
- Agents let you go from asking questions to delegating tasks. Start small.
- As AI tools get more powerful, security and containment matter more, not less.
Your Cheat Sheet
Everything from today in one place.
Quick reference, prompts, and links.
Brendon Frazer