Context Engineering: The AI Skill Marketers Actually Need

Table of contents

What is context engineering?
Why prompt engineering hits a ceiling
The four building blocks of context engineering
The compounding effect
Fewer rules, more judgement
Getting started: what to do this week
The bottom line
About Growth Method
Frequently asked questions
Further reading

Most marketing teams are stuck in the AI hamster wheel.

You open ChatGPT. You spend 10 minutes briefing it on your company, your audience, your brand voice. You get a decent first draft. Then you close the tab and the AI forgets everything.

Tomorrow, you start from scratch.

According to the 2026 State of AI for GTM report, 53% of GTM leaders report little to no impact from AI. Nearly half don’t have a single AI agent in production. The problem is the workflow, not the models.

What is context engineering?

Tobi Lütke, CEO of Shopify, put it simply:

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.

Tobi Lütke, CEO, Shopify

Andrej Karpathy, former Tesla AI lead and OpenAI founding member, expanded on this:

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

Andrej Karpathy, former Tesla AI lead and OpenAI founding member

Think about the difference this way. Prompt engineering asks: “How do I phrase this question?” Context engineering asks: “What does the AI need to know before I even ask?”

Why prompt engineering hits a ceiling

Prompt engineering was a great starting point. It taught marketers that how you ask matters. But it has a fundamental limitation: it treats every AI interaction as a one-off conversation.

A prompt can tell the AI to “write a LinkedIn post in a conversational tone for B2B SaaS marketers.” That’s useful. But it can’t tell the AI about your last three campaigns, what messaging landed with your ICP, which channels are underperforming, or what your CEO said about positioning in last week’s all-hands.

As Maja Voje describes, a marketing manager using ChatGPT for cold emails might spend 30 minutes and a dozen rounds of back-and-forth to get something usable, then lose all of that refinement when the session closes.

The teams seeing returns from AI are building better context, not writing better prompts. It’s the same shift underway in agentic marketing more broadly: the advantage moves from the person who can write a clever one-off prompt to the person who builds the system the AI works from every time.

The four building blocks of context engineering

Voje outlines a practical framework for building a context engineering system. While the specifics reference Claude Code (Anthropic’s AI coding tool), the principles apply to any AI workflow.

1. A persistent knowledge base

The foundation of context engineering is a document that the AI reads at the start of every session. It is a living document, not a one-off brief, and it includes:

Your company positioning and value proposition
Ideal customer profiles with real pain points
Brand voice guidelines (3-5 concrete descriptions of what good sounds like, not vague adjectives and not a list of prohibitions)
Current campaign priorities and lessons
Links to supporting resources

This means every interaction starts with the AI already understanding who you are, who you serve, and how you communicate. No more re-briefing.

One refinement since we first wrote this. Newer AI tools increasingly maintain their own memory. Thariq Shihipar, a member of technical staff at Anthropic, describes the shift in Claude Code: “Instead, Claude now automatically saves memories that are relevant to the work and to you.” So keep the knowledge base for the things that rarely change (positioning, ICPs, voice) and let the tool accumulate campaign lessons itself as you work, rather than hand-editing them into the document after every project.

2. Reusable playbooks

Instead of writing custom prompts for every task, context engineering uses encoded playbooks: standardised processes for repeatable marketing tasks. These capture your best practices, quality standards, and proven frameworks so every output follows the same standard.

For example, a “campaign brief” playbook might include your brief template, required fields, past examples of high-performing briefs, and specific criteria for approval. The AI follows the playbook every time, producing consistent results without the marketer reinventing the process.

Two things make playbooks work better with current models. First, don’t load every playbook at the start of every session. Anthropic calls the alternative progressive disclosure, which Shihipar describes as “loading the right context at the right times.” The brief playbook loads when you’re writing a brief, the reporting playbook when you’re reporting. Everything else stays out of the way.

Second, prefer real references over prose descriptions. Anthropic’s advice is to replace simple specs with rich ones: instead of a paragraph explaining what a good brief looks like, point at an actual high-performing brief, a live landing page, or a scoring rubric. The model does far better with a concrete example to match than with adjectives to interpret.

3. Live tool connections

The most powerful context is what the AI can access directly, not what you type. Model Context Protocol (MCP) and similar integrations let AI pull live data from your CRM, analytics platforms, and marketing tools without manual copy-pasting.

Instead of telling the AI “our open rate was 24% last month,” the AI reads it directly from your email platform. This eliminates the biggest source of AI errors: stale or incomplete information provided by the human.

4. Automated quality gates

The final layer is automatic checks that run before or after AI output. These enforce standards without manual review:

Formatting and style checks
Compliance flags for regulated industries
Approval workflows for sensitive content
Notifications when tasks complete

Quality gates turn AI from a tool that requires constant supervision into a system you can trust.

The compounding effect

The power of context engineering is that it compounds. Voje maps out what this looks like over time:

Week 1: Your knowledge base eliminates the daily re-briefing. Output relevance improves immediately.
Month 1: First playbooks created, tools connected. The system starts learning from real work patterns.
Month 3: Campaign lessons, refined ICPs, and tested messaging angles accumulate. Every new task starts from a higher baseline.
Month 6: A living knowledge base emerges that captures institutional lessons and informs all future work.

This is the gap between teams getting value from AI and teams that gave up after a few underwhelming ChatGPT sessions. The difference is infrastructure, not model quality or prompt sophistication.

Fewer rules, more judgement

There’s a temptation, once you start building context, to keep adding rules. Every time the AI does something you don’t like, you add another line: don’t do this, always do that. Anthropic’s experience suggests this backfires with newer models.

In July 2026, Thariq Shihipar wrote that Anthropic “removed over 80% of Claude Code’s system prompt for models like Claude Opus 5 and Claude Fable 5 with no measurable loss on our coding evaluations.” The reason:

Overall, we found that we were overconstraining Claude Code, both through our system prompt and in our CLAUDE.md files and skills.

Thariq Shihipar, member of technical staff, Anthropic

The specific failure was instructions fighting each other. Shihipar gives the example of a single request containing both “leave documentation as appropriate” and “DO NOT add comments” as the system prompt, skills, and user request clashed. Anthropic’s fix was to replace the pile of rules with a description of the outcome: write code that reads like the surrounding code.

The marketing translation is direct. A brand voice document that reads as twenty prohibitions will produce cautious, flat copy. One that describes what good sounds like, with a couple of real examples, gives the model something to aim at. And when you spot the AI doing something odd, check your context for two instructions that contradict each other before you add a third.

Getting started: what to do this week

You don’t need to build the whole system at once. Start with the highest-leverage piece: your persistent knowledge base.

Create a document (a simple text file will do) that captures:

Who you serve and what makes you different
Buyer personas with real pain points (not demographics)
Brand voice guidelines: 3-5 concrete descriptions of what good sounds like, with an example of each
Recent campaign lessons: what worked, what didn’t, and why
Your current marketing stack and how tools connect

This is the one document worth loading at the start of every session. You’ll notice the difference immediately.

From there, identify your most repetitive marketing task (campaign briefs, weekly reports, content outlines) and build your first playbook. Document the process, the quality criteria, and a few examples of good output. Keep it separate from the knowledge base so it loads only when that task comes up.

The bottom line

Context engineering is the practical difference between AI that wastes your time and AI that speeds up your marketing.

The shift is simple in concept: stop treating AI like a chatbot and start treating it like a new team member who needs proper onboarding, clear processes, and access to the right tools and information.

Marketing teams that build these systems now will compound their advantage every quarter. Those who keep starting every AI conversation from scratch will keep wondering why the robots aren’t delivering.

About Growth Method

Growth Method is where this context lives for a marketing team. It is the agentic marketing platform for B2B teams: agents plan, launch, and analyse campaigns from your live marketing data, with your positioning, playbooks and connected tools already in place. That is context engineering applied to the marketing team’s own operations: instead of re-briefing an AI on your ICP, voice and current campaigns every session, that context persists in the platform, so every agent starts from a higher baseline than a blank chat window ever could.

We are on-track to deliver a 43% increase in inbound leads this year. There is no doubt the adoption of Growth Method is the primary driver behind these results.

Laura Perrott, Colt Technology Services

Get started and stop briefing the AI from scratch every morning.

Frequently asked questions

What is context engineering?

Context engineering is the practice of designing the entire information environment around an AI model: not just the prompt you type, but the knowledge, data, tools, and memory that shape every response. It is built from four blocks: a persistent knowledge base, reusable playbooks, live tool connections, and automated quality gates that compound in value the longer you use them.

How is context engineering different from prompt engineering?

Prompt engineering asks “How do I phrase this question?” while context engineering asks “What does the AI need to know before I even ask?” Prompt engineering treats every AI interaction as a one-off conversation, so refinements are lost when the session closes. Context engineering builds a persistent system, the knowledge base, playbooks, tool connections and quality gates, that carries over between sessions and compounds over time.