上下文是预算不是仓库Context is a budget, not a warehouse

上下文工程管的是：每次推理时，窗口里该放哪些 token。 Anthropic 官方定义：提示词工程写指令，上下文工程在推理过程中 策划并维护最优的 token 集合——系统指令、工具、外部数据、消息历史都算。 窗口是预算，不是仓库：塞得越多，不等于用得越好。 效果未实测；以下机制均出自官方文档与论文。

Context engineering decides which tokens belong in the window at each inference. Anthropic's official definition: prompt engineering writes instructions; context engineering curates and maintains the optimal set of tokens during inference — system instructions, tools, external data, message history all count. The window is a budget, not a warehouse: stuffing more in does not mean getting more out. We have not run hands-on tests; the mechanisms below come from official docs and papers.

Anthropic 把这种现象叫 context rot： 上下文 token 越多，模型从中准确召回信息的能力越差。 根源在 transformer 架构：n 个 token 有 n² 对注意关系， 序列越长，捕捉这些关系的能力被摊得越薄。 训练数据里短序列更多，模型对长程依赖的经验也更少。

Lost in the Middle 论文在多文档问答与键值检索两个任务上发现： 相关信息在开头或结尾时表现最好，落在长上下文中段时显著变差。 显式支持长上下文的模型也一样。 结论：关键信息别埋在中段，窗口大不等于处处好使。

Anthropic calls the phenomenon context rot: as context tokens grow, the model's ability to accurately recall information from that context declines. The root is the transformer architecture: n tokens form n² pairwise attention relations, and the longer the sequence, the thinner the capacity to capture them gets stretched. Short sequences dominate training data, so models also have less experience with long-range dependencies.

The Lost in the Middle paper tested multi-document QA and key-value retrieval: performance is highest when relevant information sits at the beginning or end, and degrades significantly when it lands in the middle of a long context. Explicitly long-context models behave the same way. Takeaway: do not bury key information mid-context; a big window is not uniformly good.

OpenAI 的提示词缓存按前缀匹配： 请求先查提示词开头部分是否已缓存，命中就复用，未命中则全量处理并缓存前缀。 超过 1024 token 自动启用，不用改请求。 由此推出摆法：指令、示例、工具定义放前面，用户相关的变化内容放最后。

三个细节决定命中率： 工具定义与顺序必须逐次完全一致； 多轮历史只在 messages 数组末尾追加，不重排； 响应里的 cached_tokens 字段可查命中量，长期不用的前缀会被逐出。 缓存按组织隔离；官方说明缓存过程不存储数据，符合零数据保留条件。

OpenAI's prompt caching matches on prefixes: a request first checks whether the beginning of the prompt is already cached; a hit reuses it, a miss processes everything and caches the prefix. It activates automatically above 1024 tokens — no request changes needed. The placement rule follows: instructions, examples, and tool definitions up front; user-specific variable content at the end.

Three details decide the hit rate: tool definitions and their order must be exactly identical across requests; multi-turn history is only appended to the end of the messages array, never reordered; the cached_tokens field in the response reports hits, and prefixes unused for a while are evicted. Caching is organization-scoped; the official docs state no data is stored during the process, making it eligible for zero data retention.

另一种省预算的方式：不预载数据，只保留轻量标识符。 文件路径、存储的查询、链接，运行时用工具按需加载。 Anthropic 以 Claude Code 为例： 写定向查询、用 head/tail 看片段，大数据对象从不整体进窗口。 标识符的元数据自带信号：目录、命名、时间戳暗示用途与相关性。 agent 由此逐层探索，工作记忆只留当前需要的部分。

取舍官方也写明了：运行时探索比预检索慢； 缺引导时 agent 会滥用工具、追死胡同、漏掉关键信息。 混合策略常见：一部分数据预载求快，其余留给自主探索。

Another way to save budget: do not pre-load data — keep lightweight identifiers. File paths, stored queries, links; load on demand at runtime with tools. Anthropic cites Claude Code as its example: write targeted queries, inspect slices with head/tail, and never load whole data objects into the window. Identifier metadata carries signal: directories, naming, and timestamps hint at purpose and relevance. The agent explores layer by layer, keeping only what is currently needed in working memory.

The official trade-off is explicit: runtime exploration is slower than pre-computed retrieval; without guidance the agent misuses tools, chases dead ends, and misses key information. Hybrid strategies are common: pre-load some data for speed, leave the rest to autonomous exploration.

任务超过窗口时，Anthropic 给了三条路：

1. 压缩（compaction）：接近上限时摘要历史、重开窗口。先调召回再调精度；清理旧工具结果是最轻的一档。
2. 结构化笔记：agent 定期把笔记写到窗口外，之后再读回。进度与依赖不因窗口重置而丢。
3. 子代理：各子代理用干净窗口做深度探索，只回传蒸馏后的摘要，主代理管综合。

选择按任务特征：压缩保对话连贯，笔记适合有里程碑的迭代开发， 子代理适合并行探索有收益的研究分析。

When a task outgrows the window, Anthropic offers three routes:

1. Compaction: summarize history near the limit and reinitialize the window. Tune recall first, then precision; clearing old tool results is the lightest touch.
2. Structured note-taking: the agent periodically writes notes outside the window and reads them back later. Progress and dependencies survive window resets.
3. Sub-agents: each explores deeply in a clean window and returns only a distilled summary; the lead agent synthesizes.

Choose by task shape: compaction preserves conversational flow, notes suit milestone-driven iterative development, and sub-agents suit research and analysis where parallel exploration pays off.

1. 边缘情况清单：把每条规则都塞进提示词。官方建议改用多样、典型的示例。
2. 工具集膨胀：功能重叠、选择模糊。人都分不清该用哪个工具，agent 更分不清。
3. 压缩过狠：丢掉当时不起眼、后来才关键的细节。压缩提示要在复杂轨迹上调。
4. 等更大窗口：官方判断是窗口再大也有相关性问题，策展省不掉。

系统提示本身也有两个极端：硬编码 if-else 太脆，空泛指引太虚。 官方建议从最小提示起步，按失败模式逐步加指令和示例。

1. Edge-case laundry lists: stuffing every rule into the prompt. The official advice is diverse, canonical examples instead.
2. Bloated tool sets: overlapping functionality, ambiguous choices. If a human cannot say which tool to use, the agent cannot do better.
3. Over-aggressive compaction: dropping details that look trivial now but matter later. Tune the compaction prompt on complex traces.
4. Waiting for bigger windows: the official judgment is that relevance problems persist at any size — curation cannot be skipped.

System prompts have two extremes of their own: hardcoded if-else logic is brittle, vague guidance is empty. The official advice: start minimal, then add instructions and examples based on observed failure modes.

• Anthropic Engineering · Effective context engineering for AI agents
• OpenAI Cookbook · Prompt Caching 101
• Lost in the Middle（arXiv 2307.03172）

核验日 2026-11-07，三条均直抓成功。 OpenAI 文档主站本次网络不可达，官方内容取自其官方仓库 raw 文件。

• Anthropic Engineering · Effective context engineering for AI agents
• OpenAI Cookbook · Prompt Caching 101
• Lost in the Middle (arXiv 2307.03172)

Verified 2026-11-07; all three fetched directly with HTTP 200. The OpenAI doc site was unreachable from this network; official content was taken from a raw file in its official repository.