Death to Vibe Coding: Long live Context Engineering
From prompt engineering to context systems
I’ve silently been hating the negative connotation that ‘vibe coding’ (coined by Andrej Karpathy here) has gotten in the industry. For those of us accelerating their coding with AI as a practice and study, it’s just an easy way for others to dismiss experiences they had months ago with ChatGPT spitting out code for them in a chat window that they copy and pasted in something that didn’t work.
This shift toward context engineering aligns perfectly with the evolution I've been tracking in AI-assisted development. While the industry spent two years perfecting prompt engineering techniques, we're witnessing the emergence of a fundamentally more powerful discipline: context engineering. This represents a maturation from crafting clever prompts to architecting intelligent systems that dynamically provide agents with precisely the information they need.
"the discipline of designing and building dynamic systems that provide the right information and tools, in the right format, at the right time, to give an LLM everything it needs to accomplish a task." — Google Deepmind researcher Philipp Schmid
Having spent considerable time researching AI coding agent contexts in my early work—projects that are now defunct but provided valuable insights into context management challenges—I've watched this evolution with particular interest. The principles we explored in those early implementations around dynamic context assembly and intelligent information filtering have now become foundational to the industry's approach.
This isn't just theoretical advancement—it's delivering measurable results. Google reports that 50% of their code is now AI-assisted with a 37% acceptance rate, while companies like Cursor have built $500M+ businesses entirely around superior context management. The message is clear: context engineering is the difference between "cheap demos" and production AI systems that deliver magical experiences.
Core Principles of Context Engineering
Dynamic Information Assembly
Context engineering systems don't just provide static information—they intelligently gather and synthesize relevant data based on the current task. This might include pulling relevant code snippets, documentation, recent changes, and related issues into a coherent context package.
What does this look like in practice? A repository that’s specifically for the context that’s easy for the agent to traverse, and perhaps even a llms.txt publishing artifact that’s auto-generated. I may write another article detailing this.
One of the most popular ways for AI coding agents is context7, a crowdsourced set of markdown with a set of MCP tools that allow for efficient grounding.
Temporal Context Awareness
Understanding when information is relevant is as important as what information to provide. Modern context engineering systems track project timelines, development phases, and user workflows to deliver timely insights. This temporal awareness manifests in several practical ways that dramatically improve AI assistant effectiveness.
Sequence diagrams prove invaluable for disambiguating software workflows that would otherwise require extensive back-and-forth to piece together in every coding conversation. By providing these upfront, you eliminate the cognitive overhead of repeatedly explaining system interactions. Most agentic coding tools now automatically inject date and time into their context flow, but if yours doesn't, this simple addition provides crucial temporal anchoring for the AI.
Git tools offer particularly rich temporal data that agents can leverage when considering any given task. The commit history, branch patterns, and change frequency all provide context about project evolution and current development focus. Complementing this with MCP access to your ticketing system creates a complete temporal picture—linking code changes to business requirements and project milestones. Teams that maintain good change log management practices find their AI agents become significantly more effective at understanding project context and making appropriate suggestions.
Multi-Modal Context Integration
The best context engineering implementations combine code, documentation, visual designs, user feedback, and system metrics into unified context representations that LLMs can process effectively.
Yes, most ai agents can review screenshots, use browsers, and apply visual decision making. Your milage may vary (or is it model may vary ) on image comprehension, especially when it comes to complicated software architecture diagrams. It can be done, and it’s better than no grounding on temporal data. However, you should convert it to text if you can.
If you are going to use diagrams, create or convert them into mermaid diagrams. It covers almost all major software engineering artifacts for communicating design, and since it is in text form, it is easy for LLMs to understand, and visually appealing for humans as well.
Industry Implementation Patterns
Leading AI development tools are implementing context engineering through several proven patterns:
Repository-Wide Understanding: Claude Code's approach of using CLAUDE.md files exemplifies superior context engineering over traditional RAG-based codebase indexing. Rather than generating embeddings of code snippets, this method provides AI assistants with direct file and codebase traversal tools alongside structured context files that document architectural patterns, naming conventions, and project-specific guidance. While different tools use various conventions (OpenAI's AGENTS.md, Cursor's .cursorrules), the CLAUDE.md approach following Anthropic's best practices has proven most effective for maintaining comprehensive project context (the codebasecontext.org specification is getting a major update soon to reflect these evolved practices).
Real-Time Context Updates: As developers work, context engineering systems continuously update their understanding of current goals, recent changes, and emerging patterns.
I keep the context repository nearby, and ask my coding agent to take notes on it’s findings as I find important nuggets that are particular to that context repo (I may use more than one for different concerns).
After a long coding session where we discover that there’s crucial grounding that could have prevented a lot of the wrong turns we took, I have the agent summarize that learning and add it to a FEEDBACK.md file for the agent that is running in the other repository, or I have the agent directly modify that repository.
Contextual Tool Integration: Rather than providing generic tool access, context-aware systems present the right tools at the right moment with pre-populated relevant parameters.
My article on MCP below:
This was before MCP blew up, it is the defacto-standard way of providing tooling to agents today.
Technical implementation strategies
Modern context engineering employs sophisticated techniques for managing information flow. Sliding window approaches process text in overlapping segments, maintaining continuity while respecting token limits. I’m also reviewing new research relating to sliding window approaches, perhaps I’ll write on that soon.
Conclusion
The transition from prompt engineering / vibe coding to context engineering represents more than just a new buzzword—it's a fundamental shift in how we architect AI-assisted development workflows. As the industry moves beyond simple chat interfaces toward sophisticated AI development partners, context engineering becomes the critical discipline that separates transformative tools from clever demos.
The future belongs to development environments that understand not just what you're trying to build, but why you're building it, how it fits into your broader system, and what information you need to succeed. Context engineering is the foundation that makes this vision possible.
For developers and organizations looking to harness AI effectively, investing in context engineering capabilities isn't optional—it's essential for staying competitive in an AI-enhanced development landscape.