Claude Code and Codex Change the Rules: Context Engineering Replaces Prompting

A large analysis of Context Engineering has been published on Habr AI, and its central thesis is highly practical: in agentic tools such as Claude Code, Codex and Cursor, the quality of the result depends less on how a request is phrased than on the entire environment around the model. The user's own text takes up a tiny share of the context. The system prompt, project instructions, memory across sessions, tools, dialogue history and tool call results have the decisive influence. In other words, the era of the "magic prompt" is ending, and context design is taking its place.

The key idea revolves around the nature of the context window. The author suggests thinking of it as the model's desk: the more stray papers pile up on it, the worse attention is spread across the signals that really matter. A long context does not merely cost more.

It also lowers quality through two effects. The first is context rot: attention gets smeared out, the model starts forgetting early constraints, fixating on old points and answering more vaguely. The second is reasoning shift: as context grows, part of the compute goes to processing the input rather than reasoning, so an answer may sound confident but be less well grounded.

The article cites an estimate that a long context can cut reasoning depth by up to 50 percent without any warning.

The author explains why this happens through the Transformer architecture. In its basic form, attention has quadratic complexity O(n^2): every token must be related to every other token. So adding documents, files and long conversations scales cost not linearly but quadratically. In practice this hits four parameters at once: response speed, cost, limits and accuracy. Hence the main thesis: good context engineering is not about "stuffing everything into the model" but about selecting the minimal set of high-signal tokens that maximizes the chance of the desired result.

The piece then breaks down what context in agentic systems is actually made of. Besides the model weights themselves, there are at least several controllable layers: the system prompt, project files such as CLAUDE.md or AGENTS.

md, memory about the user and the project, skills with ready-made workflows, MCP integrations, files loaded on demand and the entire tool_result history. The article stresses that the model does not "remember" the conversation on its own: the harness reassembles the history and sends it again every time. Because of this, every superfluous tool description, every unneeded MCP server and every long system file eats tokens over and over.

Saving them takes prompt cache, deliberate compact, fresh sessions via clear, and isolation of heavy research in subagents, so the main agent receives a short summary instead of mountains of intermediate data.

The most practical part concerns the cost of agentic mode. One request to an agent may turn out to be not a single call to the model but a whole chain of several calls with tool use. In the example of analyzing a failed deployment, one user message turns into four model calls and three tool executions. Without caching, such a loop quickly becomes economically pointless; with prompt cache the price drops sharply, but even then it is worth remembering that output tokens and hidden thinking cost more than ordinary input, and long tool definitions and system instructions are included in every round.

Hence the author's conclusion: a master is distinguished not by writing cleverer prompts but by building a reproducible environment once, with project rules, memory, relevant tools and workflows, and then making the agent work under that contract.

This is an important shift for the AI development market. Competition increasingly turns less on a nicely phrased request and more on the quality of context assembly, token discipline and engineering of the agent environment. Teams that learn to manage memory, tools and history as carefully as they once managed prompts will get not only more accurate answers but also predictable costs. That means the next stage in the evolution of AI tools will be defined not by the magic of the request but by the architecture of context.