The GTM Guide to AI Context Engineering

How to turn Claude Code into a compounding knowledge asset (even if you don't code)

This post is supported by Schematic — monetization infrastructure for product & engineering teams.

Schematic decouples billing from code, giving product teams direct control over pricing, packaging, limits, add-ons, credits, trials, migrations, and upgrades — without burdening engineering.

The result: faster pricing experiments, better conversion, more upgrades, and less revenue leakage. Works natively with Stripe Billing. Used by high-growth teams like Plotly, Pagos, TightKnit, and Zep.

---

Dear GTM Strategist!

Earlier this year, Kyle Poyar and I published our 2026 State of AI for GTM report. The results were brutal. 53% of GTM leaders reported little to no impact from AI. Only 24% are seeing real returns. Nearly half don’t have a single AI agent in production.

Why is the gap so big? It isn’t the models. We all have access to the same LLMs. It isn’t prompt engineering. We all have the same cheat sheets.

The difference is structural. The 53% are treating AI like a chatbot. One-off conversations. Zero memory. Starting from scratch every time. The 24% are building Context Systems.

This shift has a name: Context Engineering.

It is the practice of designing the entire information environment around your AI, not just the prompts you type into it. It is the difference between asking AI clever questions and building a system that actually understands your business.

We are seeing a massive rise in GTM Engineering teams adopting this approach. And right now, the primary environment for building these systems is Claude Code.

In this post, I’m breaking down:

1. What context engineering is and why it changes the AI equation for growth teams

2. How Claude Code makes it practical, even if you’ve never opened a terminal

3. The four building blocks that turn AI from a chat window into a compounding knowledge system

4. Where to start (simpler than you’d think)

---

The Problem: AI Without Context Is Expensive Autocomplete

Here’s what actually happens in most companies right now.

A marketing manager opens ChatGPT. Types “write me a cold email for our product.” Gets back something that sounds polished but could have been written for any B2B SaaS company on earth. No understanding of the ICP. No awareness of the positioning. No knowledge of what worked before and what flopped.

So they try to fix it. “Make it more specific to fintech CFOs.” Better, but still generic. “Add a reference to regulatory compliance challenges.” Getting warmer, but the tone is off. “Make it sound less like AI wrote it.” Now it’s bland. “Actually, go back to version two but keep the compliance angle.” Five, eight, twelve rounds of back-and-forth, each prompt trying to inject a piece of context that the AI should have known from the start.

After 30 minutes of this prompt ping-pong, you end up with something that’s... acceptable. Not great. Acceptable. The next time you need an email? The exact same process. From zero. All that refinement you just did? Gone. The AI learned nothing from the exchange.

Most growth and marketing teams live in this loop, and it’s the reason behind those survey numbers. Every interaction starts from scratch. No memory, no accumulated knowledge, no understanding of your business. You’re re-onboarding a team member every single conversation. Except this team member has amnesia and forgets everything the moment you close the tab.

Imagine you hired a senior marketer, a really good one. Instead of briefing them once and building on that knowledge over time, you wiped their memory clean every morning and started over. You’d fire that person in a week.

Most teams are doing exactly that with AI. Then wondering why 53% see no impact.

With the prompt loop, you start from scratch every day. With the context system, the knowledge compounds.

Enter Context Engineering

In mid-2025, Shopify CEO Tobi Lütke posted something that caught the attention of the entire AI world:

“I really like the term ‘context engineering’ over prompt engineering. It describes the core skill better: the art of providing all the context for the task to be plausibly solvable by the LLM.”

Andrej Karpathy, one of the most respected AI researchers alive (former Tesla AI lead and OpenAI founding member), immediately agreed:

“Context engineering is the delicate art and science of filling the context window with just the right information for the next step.”

These aren’t fringe voices. The CEO of a $100B+ company and one of the most cited AI researchers in the world are saying the same thing: the way we work with AI is about to look very different.

Here’s how the shift breaks down:

Prompt engineering was about crafting the perfect question. Getting your words just right. Finding the magic sentence that made the AI produce good output. Focused on the input, one message at a time.

Context engineering is about designing the entire information environment around the AI. Not just what you ask, but what the AI already knows when you ask it. Your strategy, your ICP, your brand voice, your past learnings, your tool stack. Structured, persistent, and available at the moment it’s needed.

The simplest way to see the difference:

Prompt engineering is writing a clever brief for every single task. Context engineering is building the team that doesn’t need a brief, because the institutional knowledge, the playbooks, the quality standards, and the tool access are already baked into how the operation runs.

One scales with effort. The other scales with knowledge.

Why Growth and Marketing Teams Should Pay Attention

If you work in growth, marketing, or GTM operations, context engineering isn’t an abstract technical concept. It’s the answer to a specific, expensive problem you’re probably already experiencing.

The problem: Your team uses AI for content creation, email copy, competitive research, lead scoring, or campaign analysis. The output is fine. Usable. But generic. It “sounds like AI.” You end up rewriting half of it to match your voice and add the specifics only you would know.

The root cause: The AI doesn’t know your business. It doesn’t know your ICP is mid-market B2B SaaS companies with 50-200 employees who just raised Series A. It doesn’t know your brand voice is direct and practitioner-first, not corporate and safe. It doesn’t know that your outbound campaigns perform better when you lead with the industry-specific pain point instead of the feature list.

All of that context exists somewhere: in your head, in scattered docs, in your CRM, in the lessons your team has learned through trial and error. But the AI can’t access any of it. So it defaults to generic. Every single time.

Context engineering fixes this by building persistent structures that give AI access to the right knowledge at the right moment. Not dumping everything into a prompt. Not copying and pasting your brand guide at the start of every session. A system that loads the right context automatically and accumulates knowledge over time.

The 24% of teams seeing real AI impact and the 53% still getting nothing are using the same underlying models. The difference is the infrastructure around them.

Claude Code: Where Context Engineering Gets Practical

Let’s get tangible.

Claude Code is Anthropic’s command-line tool for working with Claude. If you’ve only used AI through chat interfaces (ChatGPT, the Claude website, Gemini), Claude Code works differently. That difference is the whole point.

It runs in your terminal (the text-based interface that developers use). If you’ve never opened a terminal, that might sound intimidating. What matters for growth and marketing teams: Claude Code is the first AI environment actually designed for context engineering.

In a normal chat interface, you type, the AI responds, and eventually the conversation gets long enough that the AI starts “forgetting” earlier details. No persistent knowledge. No structure. No way to say “always know these things about my business before we start.”

Claude Code changes that with four building blocks that turn disposable conversations into a persistent knowledge system.

I’ll walk through each one.

The Four Building Blocks

1. CLAUDE.md — Your Strategic Brain

Your strategic onboarding document for the AI, read automatically at the start of every session.

CLAUDE.md is a text file where you write down everything the AI should know about your business. Your positioning. Your ICP definition. Your brand voice guidelines. Your pricing model. The frameworks you use. Navigation rules so it knows where to find what. How your tools connect and when to use which one.

If you’ve ever onboarded a new marketing hire, you know the drill. The first two weeks are pure context transfer: brand guides, competitive landscape, audience research, past campaign performance, what worked and what flopped. CLAUDE.md is that entire onboarding compressed into a document the AI absorbs in seconds.

What makes it powerful is persistence. Unlike a chat conversation that evaporates after you close the tab, Claude Code reads your CLAUDE.md at the start of every session. The AI never needs to be re-briefed. It shows up already understanding your business, your standards, and your preferences.

It also evolves. When you learn something new (a sharper ICP insight, a campaign angle that outperformed, a positioning shift), you update the file. Over time, the CLAUDE.md becomes a living strategic document: your institutional knowledge, accessible to every AI interaction going forward.

As your system matures, the CLAUDE.md naturally evolves from holding everything directly into a navigation layer, pointing the AI to separate, dedicated files for your ICP research, your voice guide, your competitive analysis. You start with one well-organized document and gradually build it into a structured knowledge base. The CLAUDE.md stays at the center, but the depth behind it grows.

The companies getting real ROI from AI aren’t writing better prompts. They’re building better knowledge files.

You define your ICP, Brand Voice, and Playbooks here once, and the AI reads it before every session.

2. Skills — Your Specialist Playbooks

If CLAUDE.md is the strategic brain, Skills are the specialist methodologies it follows.

A Skill is a set of instructions that tells the AI how to perform a specific type of task. Not “write an email” but “write a cold outbound email using our 3-line framework, apply signal-based personalization from the prospect’s recent activity, use the messaging angles that worked best in last quarter’s campaigns, and follow our brand voice rules.”

Your team already has SOPs for this. The content writer follows a different process than the paid ads specialist, who follows a different process than the conversion optimizer. Skills are those SOPs, encoded so the AI can follow them with the same consistency as your best team member on their best day.

What this looks like in practice:

→ Content creation skill — knows your editorial voice, follows your formatting standards, applies your quality checklist, runs a readability review

→ Outbound optimization skill — applies signal-based prospecting methodology, personalizes based on buying triggers, structures sequences using proven frameworks

→ Conversion audit skill — loads your product positioning first, benchmarks pages against conversion standards, recommends specific copy changes with rationale

→ Competitive analysis skill — knows where to find competitor intelligence, applies your analysis framework, outputs in the format your team actually uses

Skills compound with CLAUDE.md. A content creation skill doesn’t just follow generic writing best practices. It follows YOUR best practices, informed by YOUR brand voice and YOUR ICP definition from the CLAUDE.md. The layers feed each other. The more complete your system, the more specific and useful every output becomes.

3. MCP (Model Context Protocol) — Your Tool Connections

MCP is how Claude Code connects to the tools you already use. It’s the integration layer that lets AI pull data from and interact with your existing tech stack in real time.

Your marketing stack probably includes some combination of HubSpot, Clay, LinkedIn Sales Navigator, Google Analytics, your CMS, and more. You’ve already spent time and money connecting them through native integrations or automation platforms like Zapier or n8n. MCP does the same thing, but for AI.

Without MCP, the AI can only work with what you manually paste into the conversation. With MCP, it can pull live data from your CRM, reference documentation from your knowledge base, check analytics dashboards, or enrich prospect data through your prospecting tools. All as part of the natural workflow.

A concrete example: you ask the AI to figure out why your latest outbound campaign underperformed. Without MCP, you’d need to export data from three different tools, copy it into the chat, and explain what each metric means. With MCP, the AI pulls the campaign data, audience segments, and engagement metrics directly from the source. It analyzes them together, with full context of what your previous campaigns looked like.

MCP doesn’t mean the AI takes control of your tools. It means the AI can see what your tools contain when that information is relevant. Read access to your existing infrastructure, not a rip-and-replace of your stack.

Claude Code acts as the central hub, pulling live data directly from your existing stack (CRM, Analytics, Docs) without you copying and pasting.

4. Hooks — Your Automated Quality Gates

Most guides skip this building block entirely. For teams that care about quality and consistency, it might be the most important one.

Hooks are automated checks that trigger before or after the AI takes an action. Quality gates that fire automatically to enforce standards without anyone having to remember to manually review.

In a well-run marketing operation, there’s always a review process. Content gets edited before it publishes. Emails go through brand compliance. Campaign spend needs approval above a certain threshold.

Technically, a hook is a simple rule you set: “when the AI does X, automatically run Y.” Claude Code lets you configure these through a quick setup command. You define a trigger (before the AI writes a file, after it finishes a task, when a session starts) and an action (run a formatting check, send a notification, block the operation if it violates a rule). No coding required beyond describing what you want - Claude helps you build them.

What this looks like:

→ Auto-format everything the AI writes: Every time Claude creates or edits a file, a hook automatically applies your formatting standards. No more inconsistent document styles or manually fixing headers and spacing.

→ Protect files you don’t want touched: Set a hook that blocks the AI from modifying specific files: your approved brand guidelines, your finalized campaign briefs, your production configurations. It can read them for context but can’t change them.

→ Get notified when the AI finishes: Running a long analysis or batch task? A hook sends a desktop notification the moment Claude completes, so you can work on something else instead of watching the screen.

→ Enforce rules automatically: A hook that prevents mixing data between different client accounts. A hook that flags content missing source citations before it’s presented as final. A hook that checks every new document follows your naming conventions. Whatever your team’s quality rules are, hooks make them automatic.

Hooks encode quality expectations into the system itself, so consistency doesn’t depend on someone remembering to check every output manually. The same way a pre-publish checklist catches errors before they go live, hooks catch them before the AI even presents the work.

Claude Code walks you through the setup with examples. You can also type /hooks inside Claude Code and it will guide you through creating your first one interactively.

How It All Works Together

A growth marketer sits down Monday morning and asks the AI to draft cold outbound emails for a new campaign targeting fintech CFOs.

Behind the scenes:

CLAUDE.md loads automatically — the AI already knows the company positioning, the detailed ICP profile for fintech CFOs (their specific pain points around regulatory reporting, their buying triggers tied to annual compliance deadlines), the brand voice guidelines, and the insight from previous campaigns showing that compliance-focused messaging outperforms feature-led pitches for this segment.

The outbound skill activates — it follows the team’s proven three-email sequence framework, applies signal-based personalization, and structures each email around documented best practices for this segment.

MCP connections pull live context — the AI checks the CRM for recent fintech engagement data and pulls relevant industry triggers from the prospecting tool to inform the personalization layer.

Hooks run quality checks — before presenting the drafts, automated gates verify brand voice compliance, confirm that every claim is substantiated, and check that the emails follow the approved format and length.

The marketer receives three emails that sound like their best copywriter wrote them after an hour of prospect research. Not because the AI is magic, but because the context system made the task, to borrow Lütke’s phrase, “plausibly solvable” before the first word was generated.

Multiply that across every task the team runs: content creation, lead research, campaign analysis, competitive intelligence, conversion audits. Each one drawing from the same persistent context, each one refined by accumulated learnings.

That’s when the compounding kicks in.

The Compound Effect

With traditional AI (chat interfaces, one-off prompts), knowledge doesn’t accumulate. Every session starts from zero. You might get marginally better at writing prompts over time, but the AI never gets better at understanding your business. Session one and session one hundred produce the same quality output.

With context engineering, the trajectory looks different:

→ Week 1 — You set up your CLAUDE.md with positioning, ICP, and voice guidelines. The difference is immediate. Output goes from generic to on-brand. That 30-minute editing ritual after every AI interaction? Gone. The AI sounds like your team because it understands who you are and who you’re talking to.

→ Month 1 — You’ve built your first skills, connected your CRM, and the system is learning from real work. Claude starts suggesting new skills based on the types of tasks you keep running: “I notice you do competitive analysis often, want me to create a dedicated skill for that?” You say yes, describe how you approach it, and in minutes you have a repeatable playbook the AI follows every time. Your team is producing in hours what used to take days.

→ Month 3 — Your system now has learnings from multiple campaigns baked in, ICP definitions refined against real conversion data, messaging angles tested and documented, competitive intelligence accumulating in the background. The AI knows which pain points drive replies for each segment, which positioning angles your enterprise buyers respond to versus mid-market, which content formats generate pipeline versus just engagement. Every new piece of work starts from a higher baseline than the last.

→ Month 6 — Your context system has become something that didn’t exist in your company before: a living, structured knowledge base that captures everything your team learns and makes it instantly accessible. New hires onboard through it. Campaign post-mortems feed directly into it. Competitive shifts get documented the day they happen. The AI surfaces patterns you’d miss manually, flags when messaging drifts from what works, connects dots across campaigns, segments, and quarters. Need to research five competitors at once? Claude can spin up parallel agents that work simultaneously and synthesize the findings into one report. You’re not saving time. You’re operating at a level that wasn’t possible before, regardless of team size.

The gap between teams that build this and teams that don’t widens every week. A compounding advantage that gets harder to close the longer you wait.

Chat AI stays flat because it forgets everything. A Context System learns from every campaign, creating a widening advantage over time.

You Don’t Need to Code Any of This

At this point, you might be thinking: this sounds powerful, but I’m not a developer. Claude Code runs in a terminal. There are files, configurations, tool connections. Is this really for me?

The hardest part of context engineering is knowing your business well, not knowing how to code.

If the terminal part still scares you, I have you covered. I spent the last week curating the best non-technical resources (free courses, setup guides, and “copy-paste” skills) specifically for GTM teams.

👉 Access the full Claude Code Resource List here

Writing a CLAUDE.md is writing a document. If you can write a strategy brief, you can write a CLAUDE.md. Skills are playbooks in text format. If you’ve ever documented a process for your team, you already know how to create one. MCP connections? Closer to setting up a Zapier integration than writing software.

What changes the equation: Claude Code builds itself with you.

Don’t know how to structure your CLAUDE.md? Tell Claude about your business and it will draft one. Not sure which skills you need? Claude will suggest them based on what you actually use it for: “You’ve asked me to write outbound emails five times this week. Want me to create a dedicated skill so I do it your way every time?” Need to connect your CRM? Describe what data you want access to and Claude walks you through the setup.

You don’t need to have it all figured out before you start. Open Claude Code and start writing the way you’d brief a new team member on their first day: everything about your company, your customers, your positioning, what campaigns worked and what flopped. Don’t worry about structure or format. Claude takes what you give it, asks follow-up questions, and helps you shape it into a proper context system that gets smarter every time you use it.

Over the past year, development has been turned into conversation. The technical barrier hasn’t disappeared. It’s been absorbed by the AI itself. What’s left is the part only you can do: clearly articulating what you know about your market, your customers, and what actually works.

Where to Start

You don’t need to build everything at once. Here’s a realistic first month:

Week 1: Your CLAUDE.md

Start here. Copy this structure, fill it in, and you’re already ahead of most teams using AI:

→ Company & Positioning - Who are we? What do we do? Who do we serve? What makes us different?

→ ICP Definition - Who’s our buyer? What pains them? What triggers a purchase decision?

→ Brand Voice - 3-5 concrete rules. Not “professional and friendly” but “direct, practitioner-first, no corporate fluff, every claim backed with data”

→ What Works - Learnings from recent campaigns. Which angles drove results? What messaging fell flat? What surprised you?

→ Tool Stack - What tools does your team run on and how do they connect?

Don’t overthink it. Tell Claude about your business, type it out like you’re explaining to a colleague, and let it help you turn that into a structured document.

Week 2-3: Your first skills + tool connections

By now you’ll notice patterns in how you use the system. Claude will too, and it’ll suggest skills based on your most common workflows. Let it create them. Review, adjust, and suddenly the tasks you repeat every week are running on documented playbooks instead of starting from scratch.

This is also when you connect your first tool: your CRM, your prospecting platform, your analytics, whatever your team touches most. Ask Claude to help set it up. Start with read-only access and build trust from there.

Week 4: Review and compound

Look at what the system produced this month. What was strong? What missed? Update your CLAUDE.md with those learnings. Every correction, every insight, every pattern you add now informs every future interaction. The compound effect kicks in here.

This is also a good time to set up your first hooks. After a month of using the system, you know what quality issues come up: formatting inconsistencies, files that shouldn’t be edited, notifications you wish you had. Tell Claude what you want to enforce and it will create the hooks for you.

Even a basic CLAUDE.md with your positioning, ICP, and voice guidelines will produce noticeably better output than starting every AI session from zero. Once you experience that difference, once you see the AI produce something that sounds like your team wrote it, you’ll understand why context engineering matters.

The Bigger Picture

Context engineering isn’t just an AI technique. It’s a forcing function for strategic clarity.

Building a context system requires you to precisely know (not assume, not vaguely sense) who your customer is, what your positioning is, how your brand sounds, and what works in your market. Most teams think they know these things. They try to write them down and realize the gaps are bigger than expected.

Which is exactly the point. The act of engineering your context forces you to get clear on your strategy.

In a world where AI handles more of the execution layer every quarter, strategic clarity becomes the ultimate competitive advantage. The teams that build context systems will compound their knowledge, their output quality, and their speed. Every week. Every month. Every quarter. The teams that don’t will keep starting from zero and wondering why AI hasn’t delivered on the promise.

This isn’t about a single tool or a trend. It’s a shift in how teams capture, structure, and apply what they know. The companies that figure this out first don’t just get better AI output. They get better strategy. Better onboarding. Better institutional memory. A compounding knowledge asset that grows more valuable with every interaction and never leaves when someone does.

Infrastructure, not hype. If you’re reading this thinking “we should probably start,” yes. You should.

Happy context engineering!

---

Fast-Track Your Context System

The hardest part of Context Engineering isn’t the tech. It’s writing the source material. You need to document your GTM strategy, your playbooks, and your frameworks before the AI can learn them.

You can spend weeks writing this from scratch. Or you can feed my brain directly to your AI.

I brought back my GTM in a Box bundle for a limited time. It includes my full 350-page GTM Strategist book (PDF), my 100-step launch checklist, my GTM Masterclass course, and all my core frameworks.

The Context Engineering Hack: Upload the PDF book, the and the Checklist and the transcript of the videos (they are included in the package) directly to Claude and say: “You are now an expert GTM Strategist trained on Maja Voje’s methodology. Use this context to draft my CLAUDE.md file.”

You get an instant, world-class GTM brain for your AI.