This analysis organizes the evidence behind context Engineering in the Million-Token Era, then explains the practical implications, trade-offs and current limits.
Read the evidence below as a decision trail: what changed, why it matters, which trade-offs shaped the result, and where the conclusion still depends on context.
The first step in learning AI is usually learning how to write prompt. But when you actually use the Agent tool to run a project, you will find that asking the right questions is just the starting point - the decision the AI makes in each subsequent step depends on what is in its current context, not how well you wrote the initial instructions. This is why Context Engineering is replacing Prompt Engineering.
- Prompt Engineering solves the quality of a single interaction, but the Agent tool is a multi-step autonomous execution - prompt is just the starting point, and the quality of subsequent decisions depends on what is in the context.
- Context Engineering is "designing a system so that AI has the correct context every time it makes a decision." The core operations are resident streamlining, on-demand loading, cross-session continuation, and context isolation.
- The 1M context window increases the upper limit of capabilities, but does not reduce the management complexity - the limitations of the attention mechanism (Attention) have not disappeared, and the number of instructions that the frontier model can stably follow is still about 150-200.
No matter how well written the Prompt is, it will fail if placed in a confusing Context.
Prompt Engineering solves the problem of "single interaction" quality: you ask a good question, and AI gives a good answer. This works really well in ChatGPT's question-and-answer scenario.
But the Agent tool is not one question and one answer. Tools such as Claude Code, Cursor, and Copilot are "multi-step autonomous execution" - after receiving the task, you decide which files to read, what commands to run, and how to verify the results. Anthropic in《Building Effective Agents》Agent is defined as "a system in which LLM dynamically determines its own processes and tools to use." Your prompt is just the starting point, and the quality of each subsequent step's decision depends on what's in its current context.
Real case: My blog-writer project has an 8-step transcription process, and the prompts are clearly written. But because CLAUDE.md is crammed with 300 lines (character settings, source lists of 14 companies, and three platform format specifications all mixed together), Claude has "attention drifted" when generating the format in step 6 and skips the fact-check in step 7. The problem is not that the instructions are poorly written, but that the context is polluted. Later I put CLAUDE.mdCompressed from 300 lines to 40 lines, splitting the format specifications and verification rules into Skills that are loaded on demand. With the same prompt, the execution stability is immediately improved.
No matter how good Prompt is, it will fail if placed in a confusing context. This is not a prompt problem, but a system design problem.
The four operational aspects of Context Engineering
Bojie Li's analysis of Claude's architectureThere is a sentence in the book that is very accurate: "Claude is already smart enough — intelligence is not the bottleneck, context is." Claude is already smart enough, the bottleneck is not the bottleneck, but the context.
The definition of Context Engineering: not "how to ask good questions", but"How to design a system so that AI has the correct context every time it makes a decision". The former is a one-time technique, while the latter is an ongoing system design.
From my practical experience, Context Engineering has four specific operational aspects:
Permanent information only contains things you need every time
According toAnalysis of HumanLayer, the number of instructions that the frontier model can stably follow is about 150-200, while the system instructions of Claude Code itself consume about 50. Your CLAUDE.md only has a budget of 100-150 items. Every extra "just in case" rule you add is diluting every rule that really matters. What I do isCompressed to 40 lines of core skeleton。
Load on demand: Provide it only when needed, use progressive disclosure to control token costs
AnthropicSkills mechanismUse Progressive Disclosure - only metadata (about 100 tokens) is loaded at startup, and the complete content is read only when triggered. Writing style guide 3000 words, no need to occupy context in every session. This is the same principle as "Just-in-time data retrieval", one of the four pillars of Context Engineering proposed by Bojie Li: using lightweight indexes to replace full preloading.
Cross-session continuation: use 200 tokens to exchange files for zero duplication of communication
HANDOFF.md records project progress, article-log.md records what has been written. Instead of letting the AI remember everything, we use the most streamlined structure to save the information that needs to be continued across sessions, and put it in a place where Claude can read it every time. 200-400 token handover files in exchange for session connections with zero duplication of communication.
Contextual Isolation: Don’t let research tasks pollute the main conversation
Subagent operates in an independent context window and only returns summary results. When you ask Claude to research a problem, each file it reads eats the context of the main conversation. Just send a sub-agent to investigate and come back with conclusions.
Core insight: Good Context Engineering is not about giving AI more information, but about letting it get the right information at the right time. "Put a little extra just in case" is the most common anti-pattern.
What this means
This is the third and final article in the "AI Agent Collaboration System Design" series. The underlying logic of the three articles is the same: you are not "using tools", you are "designing a system that allows the tools to operate stably".Chapter 1Let’s talk about layering.Part 2Talking about how to remember across sessions, this article raises it to the methodology level.
Context Engineering is not a concept exclusive to Claude Code. Any agent tool - Cursor, Windsurf, Copilot, tools in the future - will face the same architectural challenges. Context management, task verification, and agent division of labor can be directly transferred to this set of ideas. For those who want to enter the field of AI, rather than learning how to operate a certain tool, understanding "how to design a collaborative system between people and AI Agents" is a transferable core competency.
The threshold is not technology, but a change in the way of thinking: from "What do I want AI to do?" to "What system do I want to design to make AI work stably?" Anthropic said it well - the most successful Agent implementations do not use complex frameworks, butSimple, composable patterns. Unpack your CLAUDE.md, write the handover documents, and automate the verification process - these do not require writing a single line of code, but the effect is greater than any prompt technique.
Can Context Engineering and Prompt Engineering coexist?
Completely coexist and complement each other. Prompt Engineering optimizes the quality of "single input", and Context Engineering optimizes "the overall information environment when AI makes decisions." A good prompt works best in a good context. But if there is only a good prompt but no good context, the Agent will deviate sooner or later during the multi-step execution process.
Is 1M context window enough? Do I still need to manage context?
1M solves the problem of “can’t pretend”, and Context Engineering solves the problem of “attention allocation”. According to HumanLayer analysis, the number of instructions that the frontier model can stably follow is still about 150-200, which will not double just because the window changes from 200K to 1M. The space is large enough to allow fewer compression triggers, but the instruction budget remains unchanged.
Don’t even engineers need to understand Context Engineering?
Needed. The core of Context Engineering is "information architecture design", not programming. The PM decides which rules are permanent, which ones are loaded on demand, and how to perform cross-session handovers - these are all tasks of information management and process design. There is no need to write any code to use Claude Code's CLAUDE.md, Skills, and HANDOFF.md.
Series introduction: AI Agent collaboration system design
This article is the third article in the "AI Agent Collaboration System Design" series. The three articles can be read independently, or you can build a complete knowledge system in sequence.
- Why does Claude Code "become stupid" after being used for a long time? Use the system design thinking of six-layer architecture to solve the problem — Context management within a single session: How to split the bloated CLAUDE.md into streamlined instructions + skills loaded on demand.
- Does AI feel like amnesia every time? Use HANDOFF mode to solve cross-session memory gaps — Memory continuation across sessions: Use handover files to allow new sessions to take over the work without repeated communication.
- From Prompt Engineering to Context Engineering — why asking questions is no longer enough(This article) - Raise the practices of the first two articles to the methodological level and explain why contextual design is more important than prompt techniques.
This article is original content from gwarket, written by Aaron Huang.
1M Context Window increases the upper limit, but the attention bottleneck remains unchanged
Anthropic officially announced Claude in March 20261M token context window is fully open(GA). Both Opus 4.6 and Sonnet 4.6 are supported, with no premium added to the standard rate (Opus 4.6 maintains $5/$25 per million tokens). Max, Team, and Enterprise users of Claude Code automatically use 1M space. Opus 4.6 scored 78.3% in the MRCR v2 benchmark, which is the highest score of the frontier model in this context length.
The intuitive reaction may be: if the space becomes 5 times larger, wouldn’t it be nice to just stuff everything into it?
Not so.The finiteness of Attention does not disappear as the window becomes larger.According to HumanLayer's analysis, the number of instructions that the frontier model can stably follow is still about 150-200. Claude Code system instructions eat about 50. When 300 lines of CLAUDE.md are crammed into a 1M window, the attention allocation problem is exactly the same as when it was 200K - LLM gives priority to the head and tail ends of the prompt, and the content in the middle receives the least attention.
The real value of 1M lies elsewhere:
- A single session can do larger tasks — Read the entire codebase at once and directly analyze hundreds of pages of files without slicing.
- Compression trigger frequency reduced — 92% of 1M has a lot more space than 92% of 200K. Some users reported that compaction events were reduced by 15%, and early decisions are retained longer.
- Improved media processing capabilities — Increased from 100 to 600 images or PDF pages
But the essence of compression remains the same: it is summary, and summary means that details are lost. No matter how big the window is, your instruction budget is still the same.
To use an analogy: 1M is like a desk changing from 1 square meter to 5 square meters. If you are not good at tidying up, the table will still be messy no matter how big it is. Context Engineering is “a methodology for organizing your desktop.” 1M increases the upper limit of capabilities but does not reduce management complexity. Context Engineering addresses the latter.
Moreover, when everyone has a 1M window, the difference is no longer "whose desk is bigger", but "who has a better desk". Hardware restrictions are getting fewer and fewer, and soft design capabilities are becoming increasingly important.
What to take away
The article's value is in the evidence and trade-offs behind context Engineering in the Million-Token Era, not in treating the conclusion as universal.