I’m not a developer. I studied economics in college, spent two years as a product manager at Addepar, and I edit content for a living. I’ve also built and run production AI systems: agent workflows that research, draft, and quality-check content for venture-backed B2B companies, connected to live search data. The content those systems produce ranks on page one of Google for competitive keywords and gets cited in AI answers.
I’m starting with my background because the coding barrier that kept people like me from building real software is gone. What you need now is your judgment and taste.
The Coding Barrier Just Collapsed
When Andrej Karpathy coined “vibe coding” in early 2025, the joke worked because it was already half true: models had gotten good enough to “forget that the code even exists.” With Claude Code and the Model Context Protocol, a non-engineer can now build a working system that reads live data and carries out multi-step work on its own.
For example, take the content system I built. Two years ago it would’ve needed an engineering team: API integrations with SEO platforms, an orchestration layer, evaluation logic, a review pipeline. I built it with agent tooling and a folder of configuration files, and I never wrote a function in the process.
Context Is the New Bottleneck
Anyone can point an agent at a frontier model now. Every competitor calls the same APIs and gets the same baseline capability, so a model with no context produces the same generic output for everyone. What separates a production system from a demo is everything the model sees before it acts: the data, the standards, the domain rules, and what “great” looks like.
Shopify CEO Tobi Lütke calls this “the art of providing all the context for the task to be plausibly solvable by the LLM.” Karpathy endorsed the same shift, describing context engineering as “the delicate art and science of filling the context window with just the right information for the next step.” Anthropic now publishes engineering guidance that treats context as the scarce resource agents have to be designed around. It’s also a totally different job from prompt engineering: a prompt tunes the wording of a single request, while context engineering covers everything the model knows about the work before it starts.
What “Encoding Context” Actually Looks Like
In my content system for B2B companies, context lives in three layers. The first layer is judgment stored as configuration. Every client engagement gets a set of files the agent loads on each run: style rules with specific banned constructions, a document on what the company sells and who buys it, and a running list of the failure patterns I keep catching in drafts. One rule bans superlatives that can’t be verified. Another requires a linked primary source for every statistic. Each file is a decision I only had to make once.
The second layer is live data, connected over MCP. The agent reads keyword research, search performance data, and the actual results pages for the queries we’re targeting, the same sources I’d check myself. Without them, the model can only guess at what’s working.
The third layer is evaluation. The definition of good is written down as a checklist the agent audits its own draft against before any human sees it: each claim needs a source, comparisons need specifics, and anything that could sit unchanged on a competitor’s site gets rewritten into something only your company can claim.
When the output is weak, the fix is almost never in the code. During my testing, lots of early drafts came back clean but generic, and the cause was always the same: some piece of judgment was still in my head instead of in the context.
The Part You Still Can’t Encode (Yet)
Everything the system produces still gets a human review before it ships. The model will always miss some stuff: a claim that sounds right but isn’t, or a piece that would make the client look bad in front of a reader who knows the field. When we catch a mistake that’s likely to come up again, it becomes a new rule in the context files, and the system gets a little better each time. Some mistakes are one-offs, though, and catching them stays a human job.
Real engineering hasn’t disappeared either. Infrastructure and security are still engineering problems, and I wouldn’t ship customer-facing software this way. What’s changed is that a whole class of internal production systems no longer waits in an engineering backlog.
Why This Shifts Who Builds
The leverage question has moved from “can you build it” to “do you understand the work well enough to write it down.” The person who knows a domain’s failure patterns is better positioned to build the system than the person who knows the framework, because most of what the system does is enforce those failure patterns as checks. A senior editor knows which claims need verification before publishing. A compliance analyst knows which transactions deserve a second look. That judgment was always valuable, and for the first time the people who hold it can turn it into working systems without waiting on an engineering team.
For the past decade, the standard advice to domain experts was to learn to code. Agent tooling has made that advice mostly obsolete in about two years.
To be clear, none of this makes engineers less valuable! Writing down standards, testing against them, and treating quality as a process are engineering habits. The rest of us just need them too now.
What This Means Going Forward
If context is the bottleneck, the scarce skill is getting the judgment and standards out of subject matter experts’ heads and into artifacts a model can act on. The next wave of production AI systems will come from people whose titles say nothing about software. I built mine between editorial deadlines, and the hardest part was getting what I actually know about the work into something that AI could act on. That’s the job worth getting good at now.
Hassan Rashid is a managing editor at GrowthX, a service-as-software startup that raised a $12 million Series A, where he builds AI-powered content systems for B2B technology companies including Ramp and Vercel. He has also led SEO for the healthcare startup Alpaca Health, and writes about context engineering and putting AI agents to work on real production systems. He previously spent two years as an associate product manager at Addepar, the wealth management platform used to manage and advise on more than $9 trillion in assets.
