prune-prompt-pollution

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Prune Prompt Pollution

清理提示词污染

Agent-facing documents steer by what they activate. A prohibition names the forbidden behavior and makes it more available, not less; an absence declaration spends tokens on what is not there; a stale meta-narrative teaches the agent to ignore the document itself; a strawman warning names the temptation and primes it. This skill rewrites those four patterns without weakening the constraints the document carries. It is guidance, not a script.

面向Agent的文档通过激活的内容来引导其行为。禁令会明确禁止的行为，反而会让该行为更容易被联想到，而不是减少；缺失声明会浪费token在不存在的事物上；过时的元叙述会让Agent学会忽略文档本身；稻草人警告会点明诱惑，反而诱导Agent去关注它。本Skill会重构这四种表述模式，同时不削弱文档承载的约束条件。这是指导性方法，而非固定脚本。

Scope

适用范围

The skill reads prose whose job is to steer the agent's current behavior; records and human-facing text stay with their own standards.

In scope, automatic — rules files (
```
AGENTS.md
```
/
```
CLAUDE.md
```
), skill bodies and descriptions, prompts and visible strings, and current-state design/plan/process documents.
In scope, on request — task and session files, which mix live guidance with process record; prune them only when asked.
Out of scope — code, commit messages, decision records, postmortems, and player- or user-facing copy. These record history or rationale, or are read by people rather than by the model.

The keep list below is the safety valve: a load-bearing negative is restated as its consequence, never deleted.

本Skill处理用于引导Agent当前行为的散文式文本；记录类文本和面向人类的文本遵循各自的标准。

自动处理范围 — 规则文件（
```
AGENTS.md
```
/
```
CLAUDE.md
```
）、Skill主体与描述、提示词及可见字符串，以及当前状态的设计/计划/流程文档。
按需处理范围 — 任务和会话文件，这类文件混合了实时指导与流程记录；仅在收到请求时才对其进行精简。
不适用范围 — 代码、提交信息、决策记录、事后分析报告，以及面向玩家或用户的文案。这些内容用于记录历史或原理，或供人类而非模型阅读。

下方的保留列表是安全阀：承载关键约束的负面表述会被重构为对应的结果，绝不会被删除。

The four patterns

四种问题模式

Negation priming — "don't think of an elephant", "never use X", "avoid Y". The negation is a weak modifier the strongly-activated concept overruns. Restate the positive target: "write one-line comments", "use the official seam". A prohibition stays only as a hard guardrail you cannot phrase positively; even then, pair it with the positive target so attention lands on what to do.
Absence declarations — "there is no X", "this file does not contain Y". If the document's job is to state what is present, delete the absence line; if the absence is load-bearing (the agent must not assume X exists), state the consequence: "when X is absent, the lookup answers 503".
Stale meta-narrative — "this doc used to say", "the old version covered", "previously we recommended". The reader needs the current rule, not its history. State the current rule; the history belongs in git log or a decision record.
Strawman warnings — "some people think X, but we don't", "you might be tempted to Y, but don't". The warning names the temptation and primes it. State the correct pattern directly; the temptation is not the reader's problem.

否定诱导 — "不要想大象"、"永远不要使用X"、"避免Y"。否定是一个弱修饰词，会被强激活的概念盖过。重构为正向目标："编写单行注释"、"使用官方接口"。只有当无法用正向表述来表达时，才保留禁令；即便如此，也要搭配正向目标，让注意力集中在该做的事情上。
缺失声明 — "没有X"、"本文件不包含Y"。如果文档的作用是说明存在的内容，删除缺失声明的行；如果缺失是关键约束（Agent不能假设X存在），则说明对应的结果："当X不存在时，查询返回503"。
过时元叙述 — "本文档过去曾说明"、"旧版本涵盖了"、"我们之前推荐"。读者需要的是当前规则，而非历史。直接说明当前规则；历史内容应放在git日志或决策记录中。
稻草人警告 — "有些人认为X，但我们不这么认为"、"你可能会忍不住做Y，但不要做"。警告点明了诱惑，反而诱导了Agent。直接说明正确的模式；诱惑不是读者（Agent）需要考虑的问题。

Keep these

需保留的内容

Hard guardrails with no positive phrasing — "never store literal secrets in credentials" cannot be restated positively without losing the constraint; keep it and pair it with the positive target ("store only
```
apiKeyEnv
```
references").
Counterfactual regression pins — "without X, Y happens" states the invariant that keeps the behavior safe; keep it.
Load-bearing absences — "this environment has no database" is a constraint the agent must honor; keep it or restate it as its consequence, never delete it.
Suppression justifications — empty-catch explanations and ignore-reasons are required prose; fix a false reason, never delete it.
Measured bounds — "(measured: 512 nests ≈ 0.15s)" calibrates a constant; the word "measured" is load-bearing.

无正向表述的硬性约束 — "永远不要在凭证中存储明文密钥"无法通过正向表述保留约束，因此保留该禁令并搭配正向目标（"仅存储
```
apiKeyEnv
```
引用"）。
反事实回归锚点 — "没有X的话，会发生Y"说明了保障行为安全的不变量；保留此类表述。
关键缺失声明 — "此环境无数据库"是Agent必须遵守的约束；保留或重构为对应的结果，绝不删除。
抑制理由说明 — 空catch块的解释和忽略原因是必要的文本；修正错误的理由，但绝不删除。
度量边界 — "（实测：512个nests≈0.15秒）"用于校准常量；"实测"一词是关键内容。

Workflow

工作流程

Read the document as the agent will: what does each sentence activate?
Classify every suspect passage: negation priming, absence declaration, stale meta-narrative, or strawman warning.
Restate the surviving constraint as a positive target or a consequence, and delete the pollution around it.
Verify the edit preserves every obligation, invariant, precondition, and postcondition the original carried.

以Agent的视角阅读文档：每句话会激活什么内容？
对所有可疑段落进行分类：否定诱导、缺失声明、过时元叙述或稻草人警告。
将保留的约束重构为正向目标或结果，并删除周围的污染性表述。
验证编辑后的内容保留了原文所有的义务、不变量、前置条件和后置条件。

Examples

示例

Pollution	Restated
"Don't use `var` "	"Use `const` for bindings that never reassign; use `let` when they do."
"This file has no tests"	"Coverage lives in `test/` ; add one runnable check per non-trivial branch."
"The old README recommended X"	"Use X."
"You might want to cache this, but don't"	"The lookup is O(1); no cache."

污染性表述	重构后表述
"不要使用 `var` "	"对于永不重新赋值的绑定使用 `const` ；需要重新赋值时使用 `let` 。"
"本文件没有测试"	"测试覆盖率文件位于 `test/` ；每个非平凡分支添加一个可运行的检查项。"
"旧版README推荐X"	"使用X。"
"你可能想要缓存这个，但不要这么做"	"查询操作是O(1)复杂度；无需缓存。"