Prompt engineering is chapter one. The real unlock is context engineering: a persistent, layered system that makes Claude Code smarter every time you use it. Here's what that actually looks like.
What is context engineering? Context engineering is the practice of building persistent, layered context systems — like CLAUDE.md files, memory, custom agents, and tool integrations — so that AI tools like Claude Code get better every time you use them, instead of starting from zero. Where prompt engineering optimizes a single message, context engineering optimizes the entire environment the AI operates in. It's the skill that separates people who get mediocre AI results from people who get extraordinary ones.
New chat, type a question, get an answer, close the tab. Every conversation starts from zero. That's like hiring a brilliant assistant every morning who has never met you.
A great prompt in a vacuum gets a generic answer. The same prompt inside a system that knows your goals, voice, project history, and constraints? That's a genuine collaborator. I call this a context architecture — and building it is the highest-leverage skill in the age of AI. If you've heard the debate about prompt engineering vs context engineering, this is what the "context" side actually means in practice.
Most people are on the left. The right is where things get wild.
Most people argue about which AI model is best. The people getting the most value aren't thinking about models at all. The model is the engine. Context is the fuel. A mediocre model with great context beats a frontier model with none.
And the engine keeps getting bigger. As of March 2026, Claude's Opus model has a 1 million token context window at standard pricing. Your entire codebase fits in one conversation. That doesn't replace good context engineering — it amplifies it. A million tokens of unstructured noise is still noise. A million tokens of well-layered context is a genuine co-founder.
Startup, MBA classes, personal CRM, content, research — all through Claude Code. Here are the 7 layers of context architecture, starting from the simplest.
A CLAUDE.md file in your home directory that Claude Code loads automatically every session. It tells Claude who you are, what you're working on, and how you like to work. Mine has a research index pointing to 80+ docs, an active project map, working style preferences, and explicit rules like "never change pricing without asking me."
The fastest way to start: run /init inside Claude Code. It scans your codebase and auto-generates a CLAUDE.md tailored to what you're building — file structure, tech stack, conventions it detects. You can edit from there. One practical note: keep it under 200 lines. Files that grow past that start seeing reduced adherence — Claude gets less precise about following every instruction. If yours is getting long, move detailed content into separate files and have CLAUDE.md point to them.
The power move: you can have multiple CLAUDE.md files. One at the root level (global context), one per project directory. Claude Code reads both. Your startup gets deployment safety rules. Your class project gets the assignment rubric. Same AI, completely different behavior.
# ~/CLAUDE.md (global — loads every session)
## Who Is Colton
- Founder & CEO of Rhythmicly (AI sleep tech)
- Dual MBA/MPH at UC Berkeley Haas
- Former Amazon PM
- Berkeley, CA — Pacific timezone
## Working Style
- Prefers direct communication
- Bias toward action over analysis
- When you notice perfectionism, call it out:
"You've thought about this enough. Decide and ship."
## Critical Rules
- Never change pricing without asking
- Never modify production configs without confirmation
Personal context documents: my values, 2026 goals, career vision. Not instructions for Claude — context about me as a person. When it helps me make a strategic decision or draft a pitch, it understands not just what I want to do, but why.
This is the "soul" layer. The difference between "help me write a LinkedIn post" and "help me write one that sounds like me, reflects my actual beliefs, and connects to what I'm building this year."
Here's what most people miss: Claude Code has auto-memory now. You don't have to build this layer manually. As you work, Claude writes notes to itself in ~/.claude/projects/<project>/memory/MEMORY.md — feedback you gave, decisions you made, patterns it noticed. It updates this file on its own across sessions. Tell it "don't mock the database in tests" once, and it remembers next week.
The AI gets better over time not because the model improved, but because the context did. I have memories for writing voice, project architecture decisions, technical gotchas, and pointers to external systems. Some of those I wrote. Most of them Claude wrote after I corrected it once and it made sure it wouldn't forget.
# Example memory file
---
name: Writing Voice
description: How Colton's content should sound
type: feedback
---
- Experience over authority. Show, don't claim.
- Sensory language over abstract concepts.
- Specific numbers = credibility.
- Sound like a smarter version of Colton, not a
corporate communications team.
I run 2-3 Claude Code terminals simultaneously. Without coordination, they'd duplicate work or miss context from each other. The Command Center solves this: a dashboard (2-minute overview), weekly sprint files, a review queue, and daily logs. Every session reads the dashboard first. Every session logs what it did. Cross-session awareness, zero manual coordination.
An execution tier system decides what Claude does autonomously (research, deploying), what it drafts for my review (emails, content), and what only I can do (meetings, filming). The system knows which is which.
Saved workflows invoked with slash commands. /meeting-prep pulls CRM history, checks calendar, and generates a brief. /content-plan reads analytics and drafts the next week. /investor-prep builds a full dossier. Claude Code now has a plugin ecosystem with over 9,000 community-built skills you can install — but the most powerful ones are the ones you build yourself.
I have ~20 custom skills. Each encodes domain knowledge that would take 10 minutes to explain. Instead: one command, it just runs. To create one, add a SKILL.md file in .claude/skills/ — it's plain markdown where you teach Claude a workflow in natural language, not code. Describe the steps, the tools to use, the output format. Claude reads it when you invoke the slash command and executes the whole thing.
# Some of my skills:
/meeting-prep → CRM lookup + calendar + brief
/meeting-debrief → Save notes, update CRM, create tasks
/content-plan → Analytics + series framework + drafts
/investor-prep → Investor dossier + warm paths + angles
/new-activation → Build a personalized activation page
/handoff → End-of-session state transfer
/brand-voice → Load voice guidelines for content
Specialized sub-versions of Claude for specific domains:
Each agent has its own instructions, model selection (analyst = most powerful, CRM = fastest), and tool access. They run as sub-processes returning results to the main session.
Everything above lives inside Claude Code. This layer connects to the outside world. Through MCP (Model Context Protocol) servers, Claude directly accesses Google Calendar, Gmail, Drive, and Sheets. It queries my Supabase database. It manages n8n automation workflows. It taps buttons and takes screenshots in the iOS Simulator. The most useful MCPs for builders: GitHub (PRs, issues, code review), database servers (Supabase, Postgres — read and write your data directly), and Google Workspace (calendar, email, docs without leaving the terminal).
As of March 2026, Claude Code shipped MCP Elicitation — MCP servers can now ask you questions mid-task. Your database MCP can prompt "which table?" instead of guessing. Your deployment server can confirm "deploy to staging or prod?" before pushing. It makes the whole system conversational instead of fire-and-forget.
On top of that, 14 automated workflows run on n8n Cloud: content generation Saturdays, ops digests Mondays, daily anomaly alerts, hourly CRM syncs, queue aging warnings.
This is where AI stops being a tool you open and becomes infrastructure running in the background. Always on.
My context engineering system took months to build. Yours will look completely different. What matters is the principle: context compounds. Every layer makes every other layer more valuable.
Start with Layer 1. Five minutes. Add layers as you feel the edges of what's possible. Whether you're vibe coding a side project or running a startup through Claude Code, the same principle applies: better context in, better results out. And if typing feels slow, try /voice — Claude Code has native push-to-talk voice input now. Hold spacebar, talk, release to send. It's surprisingly good for thinking out loud about what you want to build.
This is what I teach in my AI activation and expansion sessions — how to build a context architecture that makes Claude Code feel less like a tool and more like a collaborator who actually knows your work. If you want to go deeper, I offer 1:1 activation sessions (build your first context system together) and expansion sessions (level up what you've already got). Reach out here if that's interesting.
Your existing AI already knows a lot about you. The fastest way to start context engineering: have it write a summary, then paste it into Claude Code as your first CLAUDE.md file.
Open whichever AI you use most (ChatGPT, Claude.ai, Gemini) and paste this:
You know a lot about me from our conversations. I'm going to start using a new AI coding tool and I need to give it context about who I am.
Write a comprehensive summary of everything you know about me. Cover:
1. Who I am — name, background, what I do
2. What I'm currently working on and building
3. My goals for the next 6-12 months
4. How I like to work and make decisions
5. My communication style and preferences
6. Things I care deeply about
7. Anything else that would help a new AI work with me effectively
Make it specific, personal, and honest — not a generic template. This is going to be read by an AI that's never met me, so include the things that matter most.
Review what it generates, edit anything that's off, paste it into Claude Code, and tell it: "Remember all this stuff about me." It saves to memory and loads in every future conversation. Months of accumulated context, ported in 5 minutes.