Claude Code 将系统提示词缩短 80%，小模型也适用吗？
英文标题：🔶 Claude Code Cut Their System Prompt by 80%. Does That Work for Small Models Too?
Anthropic 最近为最新模型移除了 Claude Code 系统提示词中约 80% 的内容。Ante 团队在 terminal-bench 2.1 的 89 项任务上，以小型低成本模型 deepseek-v4-flash 测试了将完整请求提示词从约 3.4 万字符缩减至约 1.8 万字符的效果。结果显示，短提示词的性能与长提示词基本持平；在结果一致的 69 项任务中，输入 Token 中位数降低了 32%。该实验提示开发者，在新一代模型上线时，应优先审查并删除过时的通用流程指令，而非持续堆叠提示词。
988 B
148.8 KB
155.7 KB
257 B
938 B
148.8 KB
中文
Anthropic 最近为其最新模型移除了 Claude Code 系统提示词中约 80% 的内容。
直觉上，模型越智能，所需的引导、约束和示例就越少——当能力达到一定水平后，示例不再提供帮助，反而会限制模型发挥。
有意思的是，我们认为这些改变有数据支撑——但仅限于 Anthropic 最新的模型。显而易见的后续问题是：这一结论能否向_下游_市场延伸？小型、快速、低成本的模型，恰恰是人们预期会依赖细致手把手指导的模型。若将它们的提示词砍半，按传统观点，它们会直接失效。
我们替你做了这项实验。
实验设置
我们的 Agent Ante 提供 --short-prompt 模式：将系统提示词整合压缩（约从 5.0k 降至约 2.3k 个字符），同时缩短工具描述，使每次请求的完整提示词大约减半——从约 34k 降至约 18k 个字符。
我们使用真正小型、快速且低成本的模型 deepseek-v4-flash，在 terminal-bench 2.1 的完整 89 项任务套件上进行了 A/B 测试。就实际部署的 Agent 而言，它与前沿模型的距离已经相当遥远。
其他条件全部保持不变：相同的 Agent 构建版本、相同的固定数据集摘要、相同的沙箱池、相同的推理强度、相同的标志参数。两次运行唯一的差异，就是一个 CLI 标志。
结果
Token 对比排除了 20 项通过/失败结果发生变化的任务，然后对两组独立中位数进行比较。
性能：持平。 短提示词的得分实际上高出 2.3 分，但在每项任务只尝试一次的情况下，这处于噪声范围内——诚实的结论是“没有可测量的差异”。错误数量也完全相同。
在结果相同的任务子集中，输入 Token 降低 32%。 两次运行之间有 20 项任务的通过/失败结果发生变化，这也改变了这些 Agent 的工作时长，使它们的 Token 数量不适合进行同等条件比较。我们排除这 20 项任务，保留结果不变的 69 项任务，再比较两组独立中位数：长提示词为 509,498 个输入 Token，短提示词为 346,409 个。这降低了 32%。这是一个刻意筛选的子集，并非对整个套件成本的估算。
在全部 89 项任务中，输入 Token 总量的汇总几乎没有变化，这一点值得坦诚说明：对话历史占据了输入用量的大头，而少数采用短提示词的运行走了更长的解题路径，其额外消耗淹没了单次请求节省下来的 Token。运行成本同样只略有变化，从 4.25to4.25 变为 4.25to4.18。
**我们重点观察但没有发现的情况：**能力断崖。两次运行中有 20 项任务的结果发生翻转——11 项新通过、9 项新失败——但这种波动符合该基准测试中单次尝试的方差特征，而非某种模式。没有任何任务类别出现崩溃。在两次运行都成功解决的任务中，短提示词 Agent 使用的步骤甚至略少（886 次，对比 911 次）。
结论
因此，至少对于这个模型而言：答案是肯定的。我们将提示词减半，没有测得更差的结果；并且在结果相同的 69 项任务子集中，短提示词运行的输入 Token 中位数大约低了三分之一。
这不是我们预期的结果。长提示词的编写依据是：小模型需要比前沿模型更多的指令。显然并非如此——至少这个模型不是这样。像我们这样编写的提示词文本，会不断累积成针对两三代以前模型失败模式的补丁，但从来没有人回头删除其中任何内容。它从不会显现为一个问题，只会在每一次请求中悄悄花钱，直到有人进行了 A/B 测试。
这只是一次实验：一个模型、一个基准测试、每项任务一次尝试。32% 的 Token 结果来自一个经过筛选、结果相同的任务子集，而两次运行之间的任务结果波动也说明，多次尝试验证为何十分重要。±5 分的噪声范围意味着，其中仍可能隐藏着小幅性能回归，因此在切换任何默认设置前，我们会进行多次尝试测试。而且，能够安全删除的是通用流程指令——任务关键上下文则完全是另一回事。
但方向已经很清晰，并且与 Anthropic 在更高一层模型上发现的结果一致：当新一代模型到来时，你的第一步不应该是往提示词里增加内容，而应该是删除内容。
English / 原文
Anthropic recently removed ~80% of Claude Code's system prompt for their newest models.
It is quite intuitive that as models get smarter, they need less direction, fewer constraints, and fewer examples — past a certain capability level, the examples stop helping and start boxing the model in.
What's interesting is that we believe those changes are backed by data — but only for Anthropic's latest models. The obvious follow-up question: does the result survive going down -market? Small, fast, cheap models are exactly the ones you'd expect to depend on detailed hand-holding. Cut their prompt in half and conventional wisdom says they fall over.
We ran the experiment so you don't have to.
The setup
Our agent, Ante, ships a --short-prompt mode: a consolidated system prompt (~5.0k → ~2.3k chars) plus shortened tool descriptions, cutting the full per-request prompt roughly in half — from ~34k to ~18k characters.
We A/B tested it on terminal-bench 2.1 , the full 89-task suite, with deepseek-v4-flash — a genuinely small, fast, cheap model, about as far from frontier as agents get deployed in practice.
Everything else held constant: same agent build, same pinned dataset digest, same sandbox pool, same effort level, same flags. The two runs differ by exactly one CLI flag.
The results
The token comparison excludes the 20 tasks whose pass/fail outcome changed, then compares the two independent medians.
Performance: parity. The short prompt actually scored +2.3 points higher, but with one attempt per task that's inside the noise floor — the honest read is "no measurable difference." Error counts were identical.
Input tokens: 32% lower in the same-outcome subset. Twenty tasks changed pass/fail outcome between runs, which also changed how long those agents worked and made their token counts poor like-for-like comparisons. We exclude those 20 tasks, leaving the 69 whose outcome stayed the same, then compare the two independent medians: 509,498 input tokens with the long prompt versus 346,409 with the short prompt. That's 32% lower. This is a deliberately selected subset, not a whole-suite cost estimate.
The aggregate input-token totals across all 89 tasks are nearly flat, and that's worth being honest about: conversation history dominates input usage, and a few tasks where the short-prompt run took a longer solution path swamp the per-request savings in the sum. Run cost likewise changed only slightly, from 4.25to4.25 to 4.25to4.18.
What we watched for and didn't find: a capability cliff. Twenty tasks flipped outcomes between the runs — 11 newly passed, 9 newly failed — but that churn is characteristic single-attempt variance on this benchmark, not a pattern. No task category collapsed. On tasks both runs solved, the short-prompt agent even used slightly fewer steps (886 vs 911).
The takeaway
So, for this model at least: yes. We cut the prompt in half, measured nothing worse, and in the 69-task same-outcome subset the short-prompt run's median input-token count was about a third lower.
That's not the result we expected. The long prompt was written on the theory that a small model needs more instruction than a frontier one. Apparently not — or at least, not this one. Prompt text like ours accumulates as patches for the failure modes of models that are two or three generations gone, and nobody goes back to delete any of it. It never shows up as a problem. It just quietly costs money on every request until someone runs the A/B.
This is one experiment: one model, one benchmark, one attempt per task. The 32% token result comes from a selected same-outcome subset, and the run-to-run task churn shows why multi-attempt validation matters. A ±5pt noise floor means a small regression could still be hiding in there, so we'd run multi-attempt before flipping any default. And what you can safely delete is generic process instruction — task-critical context is a different thing entirely.
But the direction is clear, and it matches what Anthropic found one tier up: when a new model generation lands, your first move shouldn't be adding to your prompt. It should be deleting.