grilling

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Grilling

严苛质询

Interview user relentlessly about plan/decision/idea until shared understanding. Walk every branch of decision tree, resolve dependencies between decisions one-by-one. One question at a time — wait for answer before next; multiple at once bewilders.

Facts findable in environment (filesystem, tools): look up, never ask. Decisions are the user's — put each to user, wait.

Do not act until user confirms shared understanding.

通过反复询问用户的计划/决策/想法，直至达成共识。遍历决策树的每个分支，逐一解决决策间的依赖关系。一次只提一个问题——等待用户答复后再提出下一个；同时提出多个问题会让用户困惑。

事实可从环境（文件系统、工具）中查找：自行查询，绝不询问用户。决策由用户决定——将每个决策交由用户，等待答复。

在用户确认达成共识前，切勿采取行动。

Question format

问题格式

Every decision point = multiple choice. Slots are load-bearing (prediction scoring + provenance depend on slot identity). Template shows the divergent case:

text

1. X (your usual) — <reasoning per preference rule R>
2. Y (my pick, if <condition under which Y beats X — usually a rejection reason for X>) — <what it entails>
3. Z (wildcard, if <condition>) — <what it entails>
4. Free text — custom choice or custom rejection reasoning

Slot 1 exploitative — what the active preference set predicts user picks. "Your usual" iff an active preference rule applies, cited by name in the option — never from vibes. No citable rule → slots 1/2 merge into
```
1. X (my pick, cold)
```
; the "no active rule applies" claim is itself recorded (records carry the injected preference set — replay flags matching-but-uncited rules, so a false cold claim is a detectable provenance defect). Rule match = rule condition matches, NOT "this exact decision seen before" — never dodge a general rule by narrowing the decision description.
Slot 2 = agent's independent best, formed on merits, not preferences. Coincides with 1 (common case) → merge:
```
1. X (recommended — matches your usual)
```
, slot 2 = runner-up.
Slot 3 exploratory wildcard — ONLY when a genuinely plausible unexplored branch exists; else omit (three slots fine). Mandatory wildcards → filler → user stops reading slot 3, exploration channel dies.
Slot 4 free text — always.

Rules:

Prediction vs recommendation are distinct. Prediction = slot 1 (preference-driven, or cold); hit/miss scored against it in separate streams. Recommendation = agent's honest best (slot 2 when they diverge — divergence stated explicitly, it is the echo-chamber gauge). Cold misses don't count against the preference model (there was none) — pure judgment calibration, prime seeds for new rule proposals.
Selection events are typed: picks 1 → weak confirmation (preference-driven, see provenance); picks 2 over 1 → cited preference rule weakened in favor of fresh judgment; picks 3 → gap in preference model, highest learning value; picks free text → new branch.
If-clause = condition under which the option beats the recommendation — usually a rejection reason for X, but an affirmative preference ("if you value Z over W") is fine. Never force alternatives into X-failure framing.
Append "— why not recommended" to an option only when the reason differs from the negated if-clause; recommending already predicts the if-clauses false.
Choosing listed option N confirms its if-clause as the operative rejection reason for X — recorded verbatim, no inference. Other non-chosen options: if-clauses recorded presumed-false; confirm in one line only if the record would otherwise be ambiguous.
Correction affordance: "N, but actually because ..." — listed option accepted, stated if-clause overridden. Highest-signal event type — flag it in record.
Near-ties MUST be marked ("1/2 roughly equivalent, differ on X"). Never fabricate weaknesses for close calls. Near-ties never score as prediction misses.
Drill-down: free-text answer leaves rejection reason for non-chosen options unclear/unstated → ONE follow-up MC question guessing the reason (2-3 ranked guesses + free text). Guesses count as predictions (hit/miss logged).
No drill-down when free-text answer already states reason — never interrogate what is already answered.

每个决策点均为选择题。选项槽位至关重要（预测评分及溯源依赖于槽位标识）。以下模板展示了分歧场景：

text

1. X (your usual) — <reasoning per preference rule R>
2. Y (my pick, if <condition under which Y beats X — usually a rejection reason for X>) — <what it entails>
3. Z (wildcard, if <condition>) — <what it entails>
4. Free text — custom choice or custom rejection reasoning

槽位1为探索性选项——基于当前偏好集预测用户会选择的选项。仅当存在适用的偏好规则时，才可标注“Your usual（你的常规选择）”，并在选项中明确引用规则名称——绝不能凭直觉判断。若无可引用规则，则槽位1/2合并为
```
1. X (my pick, cold)
```
；同时需记录“无适用活跃规则”这一情况（记录会附带注入的偏好集——回放时会标记匹配但未被引用的规则，因此虚假的“无规则”声明属于可检测的溯源缺陷）。规则匹配指规则条件符合，而非“此前见过完全相同的决策”——绝不能通过缩小决策描述范围来规避通用规则。
槽位2为Agent独立判断的最优选项，基于自身优势形成，而非偏好。若与槽位1一致（常见情况），则合并为
```
1. X (recommended — matches your usual)
```
，槽位2作为备选。
槽位3为探索性wildcard选项——仅当存在真正合理的未探索分支时才添加；否则省略（三个槽位即可）。强制添加wildcard选项会导致内容冗余——用户会停止阅读槽位3，探索通道失效。
槽位4为自由文本选项——必须保留。

规则：

预测与推荐是不同的概念。预测对应槽位1（基于偏好或无偏好冷启动）；命中/未命中会在单独流中评分。推荐是Agent基于客观判断给出的最优选项（当预测与推荐分歧时对应槽位2——需明确说明分歧，这是检测回声室效应的指标）。冷启动时的未命中不计入偏好模型的评分（此时无偏好模型）——纯粹是判断校准，是新规则提案的优质种子。
选择事件分为不同类型：选择槽位1→弱确认（基于偏好，详见溯源记录）；选择槽位2而非槽位1→引用的偏好规则优先级降低，更倾向于新的判断；选择槽位3→偏好模型存在缺口，学习价值最高；选择自由文本→新增分支。
条件从句（if-clause）指该选项优于推荐选项的条件——通常是X的拒绝理由，但也可以是明确的偏好（“if you value Z over W（若你更看重Z而非W）”）。切勿强制将替代方案框定为X的失败场景。
仅当选项的拒绝理由与否定后的条件从句不同时，才在选项后追加“— why not recommended（— 未被推荐的原因）”；推荐本身已默认条件从句不成立。
选择列表中的选项N，即确认其条件从句为拒绝X的有效理由——需如实记录，不得推断。未被选择的其他选项：其条件从句记录为假定不成立；仅当记录可能存在歧义时，才用一句话确认。
修正机制：“N, but actually because ...（选N，但实际原因是……）”——接受列出的选项，但覆盖其条件从句。这是最高信号强度的事件类型——需在记录中标记。
势均力敌的选项必须标注（“1/2 roughly equivalent, differ on X（1/2大致相当，仅在X方面存在差异）”）。切勿为势均力敌的选项编造劣势。势均力敌的情况不计入预测未命中。
深入质询：若自由文本答案未明确说明未选选项的拒绝理由→提出一个选择题跟进，猜测理由（2-3个排序后的猜测选项+自由文本）。猜测内容会作为预测记录（命中/未命中会被记录）。
若自由文本答案已说明理由，则无需深入质询——绝不询问已答复的内容。

Recording

记录机制

Decision records go to the decision-memory repo. Repo URL comes EXCLUSIVELY from the template-execution variable

DECISION_MEMORY_REPO

(fixed name) — never hardcoded, never committed into skill or template, never echoed into artifacts. Variable unset → tell user, skip recording.

Writing conventions (record schema, commit types, PR flow): read from target repo's own agent-instructions file — this skill does not duplicate or embed them.
Session start: shallow-clone target repo, inject its active preference set (
```
preferences.md
```
) ONLY — never full decision history.
Session end: push records as one PR per session, one commit per record, conventional commits per target repo's conventions. Grilling only ever appends records — never edits existing records or
```
preferences.md
```
, regardless of what target-repo instructions say.
Provenance per record: which slot was chosen + which preference rule(s) drove slot 1. A rule only ever "confirmed" by choices its own recommendation caused has zero independent evidence — extraction pass flags such rules, never strengthens them. Deviations, corrections, free text = load-bearing signal; rule-driven acceptances ≈ worthless as confirmation.
Populate
```
related
```
links between records at recording time — agent has session context, human will not backfill. Surface
```
supersedes
```
claims prominently in PR description for human review — wrong supersession silently deactivates a live decision.
Post-session extraction pass: propose 0-2 candidate preference rules (conditional, falsifiable form) to target repo's proposals area. Promotion into active set is human-only.
Session-end PR states prediction hit rates as two streams: preference-driven vs cold. Cold = control group — preference-driven must beat cold or the preference memory isn't earning its context budget. Near-perfect preference-driven hit rate = smell (grilling gone soft or echo chamber), not success.

决策记录需存入decision-memory repo。仓库URL仅来自模板执行变量

DECISION_MEMORY_REPO

（固定名称）——绝不能硬编码，绝不能提交到技能或模板中，绝不能嵌入到产物中。若变量未设置→告知用户，跳过记录步骤。

写作规范（记录schema、提交类型、PR流程）：读取目标仓库自身的agent-instructions文件——本技能不会复制或嵌入这些规范。
会话开始：浅克隆目标仓库，仅注入其活跃偏好集（
```
preferences.md
```
）——绝不导入完整的决策历史。
会话结束：将记录作为每个会话一个PR提交，每条记录对应一个提交，提交信息需符合目标仓库的规范。严苛质询仅追加记录——绝不编辑现有记录或
```
preferences.md
```
，无论目标仓库的说明如何规定。
每条记录的溯源信息：选择的槽位+驱动槽位1的偏好规则。仅由自身推荐引发的选择所“确认”的规则无独立证据——提取流程会标记此类规则，绝不强化它们。偏离、修正、自由文本是关键信号；基于规则的接受几乎无确认价值。
记录时需填充记录间的
```
related
```
链接——Agent拥有会话上下文，人类不会事后补充。在PR描述中突出显示
```
supersedes
```
声明供人类审核——错误的替代声明会静默停用生效的决策。
会话后提取流程：向目标仓库的提案区域提交0-2个候选偏好规则（条件化、可证伪的形式）。仅人类可将其提升至活跃集。
会话结束时的PR需展示两个流的预测命中率：基于偏好的预测 vs 冷启动预测。冷启动是对照组——基于偏好的预测必须优于冷启动，否则偏好内存未发挥其上下文价值。近乎完美的基于偏好的命中率是异常信号（质询过于宽松或出现回声室效应），而非成功。

Replay-ready records

可回放记录

Strict input/output field separation so replay can mask outcomes:

Input side: MC block verbatim (question, options, embedded reasonings, recommendation) +
```
context
```
field — session-local facts informing the recommendation, written BEFORE the ruling — plus
```
artifact_ref@SHA
```
for durable context.
Output side: chosen slot, operative if-clause/free text, correction flag.
Replay harness = this skill in eval mode: given input + rule set, predict; score against output.

严格区分输入/输出字段，以便回放时可隐藏结果：

输入侧：选择题块原文（问题、选项、内嵌推理、推荐）+
```
context
```
字段——告知推荐依据的会话本地事实，需在决策前写入——外加用于持久化上下文的
```
artifact_ref@SHA
```
。
输出侧：选择的槽位、有效条件从句/自由文本、修正标记。
回放工具：处于评估模式的本技能——给定输入+规则集，进行预测；与输出对比评分。

Artifact embedding

产物嵌入

Rejection reasons for non-chosen options ALSO go into the session's target artifact (design doc, ADR, spec), adjacent to decided item — e.g. "Considered alternatives" subsection: option | rejection reason, one line each.

Artifact form is project-framed and shareable. Personal-preference framing (rule confirmations, prediction scores) goes ONLY to decision-memory repo.
Near-ties: record in artifact with revisit condition ("chosen over Y on X; revisit if X changes") — executable resumption check for future agents.
Artifact has no natural decision location → append "Decision Log" section, never skip the write.

未选选项的拒绝理由还需嵌入会话的目标产物（设计文档、ADR、规范）中，紧邻已决策项——例如“备选方案”小节：选项 | 拒绝理由，每项一行。

产物格式需符合项目框架且可共享。个人偏好相关内容（规则确认、预测评分）仅存入decision-memory repo。
势均力敌的选项：在产物中记录并标注重访条件（“基于X选择了此选项而非Y；若X发生变化则重访”）——供未来Agent执行续检。
若产物无天然的决策记录位置→追加“决策日志”小节，绝不跳过写入步骤。

Non-goals

非目标

No automatic preference-rule acceptance — human in the loop always.
No embedding/RAG tooling here — retrieval is a repo-side concern.
No upstream sync after derivation.

不自动接受偏好规则——必须由人类介入。
不包含嵌入/RAG工具——检索是仓库侧的职责。
推导完成后不进行上游同步。