本2026-07-08文章宣布新增51条记录，编号156–206，来源包括 Superpowers、Anthropic 插件生态和 Matt Pocock skills。当时多数条目是名称与问题摘要，不是已完成验证的方法。
重建版本读取冻结的来源文件，检查实际机制，并与现有目录去重。来源 confidence 标签不等于质量分数或模型实验。当前介绍、场景、正反教学构造和边界，采用与旧模式相同的编辑标准。
原摘要编号保留在迁移记录中。部分条目已合并或撤下，因此历史新增51条不代表当前仍有51个独立有效方法。
]]>This 2026-07-08 post announced51 additional records, numbered156–206, from Superpowers, Anthropic’s plugin ecosystem and Matt Pocock’s skills. Most new records were name/problem summaries rather than complete tested methods.
The rebuild reads the frozen source files, checks each observed mechanism and deduplicates it against the catalog. A source confidence label is not a quality score or a model experiment. Current descriptions, scenarios, good/bad constructions and limits follow the same editorial criteria as older methods.
The original summary numbering remains part of the migration record. Some entries now merge into earlier identities or have been withdrawn, so51 additions does not imply51 active methods in the current edition.
Current source revisions · Reviewed catalog · Admission rules
]]>本2026-05-26文章宣布收录155个编号方法。这是历史目录规模，不是效果测量。旧版聚合插件来源说明没有提供逐模式的固定版本证据。
当前版本逐项评估任务、机制、正反例、可检查结果和边界。重复身份合并，错误或薄弱内容撤下；有用但未定位原文的方法保留“来源未确认”标签。
当前实例链接到固定版本，不代表最早发明归属。最新数量来自有效元数据，不从这篇历史标题推断。
存续方法有独立的中英文页面与具体场景。旧编号和入口继续通过合并或撤下记录定位。
]]>This 2026-05-26 post announced a catalog of155 numbered methods. That is a historical collection size, not a measurement of their effectiveness. Its aggregate plugin-source claims did not provide per-pattern immutable evidence.
The current edition assesses each method’s task, mechanism, contrasting examples, observable result and boundaries. Duplicate identities merge, incorrect or weak content is withdrawn, and useful methods without located sources retain a Source unconfirmed label.
Current source instances link to fixed revisions; they do not establish earliest invention. The current catalog count comes from active metadata rather than this historical title.
Each surviving method has an independent English/Chinese page and a concrete use case. Old numbers and entry points remain resolvable through merge or withdrawal records.
]]>保留本文2026-04-19的实验语境。旧报告描述了重复运行的 prose/tree 对比，但当前仓库没有原始输出，无法重新计算当时的数字；这些结论不作为当前效果证据。
两组提示词的输出词汇也不同：树形版本给出固定 action 字符串，散文版本允许自由句子。这混合了分支表示与输出契约两个变量。决策树本身不保证模型决策稳定或正确。
任务：为失败部署选择一个已获授权的行动。
Bad：
选择最好的缓解措施并解释。
Good：
支持且授权回滚时，提出 action=rollback。
否则，如果授权的 feature flag 能禁用变更，提出 action=disable_flag。
否则，报告 action=investigate 及缺失的事实。
返回含 action 与 evidence 的 JSON，不执行缓解操作。
这个构造让分支、未知状态和输出枚举可检查。输出 schema 与语义判断准确性仍应分别验证。
旧样例保留在 eval/decision-tree-ab/，运行器需要 Claude Code CLI。Patternfoo 是当前可配置的评测工具。运行模型实验须显式选择，不属于网站静态检查。
新的对比应固定事实、模型配置、输出枚举与预算，只改变分支表示方式。报告每次结果和失败，并保存原始输出后再提出数值主张。
]]>This article retains its 2026-04-19 experiment context. The earlier report described repeated prose/tree runs, but the raw outputs are absent from this checkout. Its numerical reproducibility claims cannot be recomputed here and are not current evidence.
The two prompt forms also differed in their output vocabulary: the tree supplied fixed action strings while prose allowed free-form sentences. That confounds branch presentation with output-contract enforcement. A tree alone does not guarantee stable or correct model decisions.
Task: choose an authorized action for a failed deployment.
Bad:
Choose the best mitigation and explain it.
Good:
If rollback is supported and authorized, propose action=rollback.
Otherwise, if an authorized feature flag disables the change, propose action=disable_flag.
Otherwise report action=investigate and the missing fact.
Return JSON with action and evidence. Do not execute a mitigation.
This construction makes the branches, unknown state and allowed output values inspectable. Check schema compliance separately from semantic decision accuracy.
The retained fixtures are in eval/decision-tree-ab/; its runner requires Claude Code CLI. Patternfoo is the maintained configurable evaluation tool. Running either tool is an explicit model experiment, not part of the static site checks.
A new comparison should hold facts, model configuration, output enums and budget constant; change only the branch representation. Report individual outcomes and failures, and retain raw outputs before making numerical claims.
]]>原2026-04-19文章用条件熵和注意力解释提示词结构，但仓库没有将具体措辞与这些模型内部量关联的测量。修订版保留实用方法，移除未经支持的因果保证。
明确输入、所需结果及未知信息。例如诊断配置问题时，先定位加载器、配置路径和已观察到的症状，再讨论修改。
把“清楚报告”改成必需字段和允许值，并提供缺失输入或未知分支。格式校验能检查契约，不能证明答案为真。
固定任务与事实，只修改导致一个具体失败的指令，解释可检查的差异并说明边界。本站示例是教学构造，不是模型效果测量。
每个分支都需要的约束放在行动附近。任务专属参考材料附明确的读取条件。保留来源位置，不把记忆或摘要当作当前事实的证明。
]]>The original 2026-04-19 article explained prompt structure through conditional entropy and attention. This repository does not contain measurements linking its suggested wording to those internal model quantities. The revision keeps practical methods and removes unsupported causal guarantees.
Specify the input, required outcome and what is not known. For a configuration diagnosis, identify the loader, config path and observed symptom before asking for a change.
Replace “report clearly” with required fields and allowed values. Include a missing-input or unknown branch. A format validator checks the contract, not the truth of the answer.
Keep task and facts constant. Change the instruction responsible for one concrete failure, explain the expected difference, and state the boundary. Examples here are teaching constructions, not measured model outputs.
Put always-needed constraints near the action. Attach an explicit reading condition to task-specific references. Preserve a source location instead of treating a memory or summary as current proof.
Teaching-example guide · Reference-reading guide · Editorial criteria
]]>