菜单
Generate a merge-gate quiz from the ledger, cross-checked against the diff — questions target what was actually uncertain.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/leoncuhk/defog --skill fog-quiz命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
During implementation, log deviations and newly discovered fog to the ledger instead of stopping or silently improvising. Use whenever implementing against a plan.
Maintain a FOG.md unknowns ledger for a task. Use when a task has unresolved decisions or unverified constraints, when the user mentions unknowns or the fog ledger, or when another fog-* skill needs the ledger rules.
Resolve open decisions in the ledger by interviewing the user, one evidence-priced question at a time. Use when FOG.md has open KU decision items, or the user asks to be grilled about a plan.
基于 SOC 职业分类
正在显示 SKILL.md
| name | fog-quiz |
| description | Generate a merge-gate quiz from the ledger, cross-checked against the diff — questions target what was actually uncertain. |
| disable-model-invocation | true |
The ledger can only grade what it logged — so first reconcile ledger against diff: scan the actual changes for anything no plan step, resolved item, or deviation accounts for. Unaccounted changes are unlogged deviations — the highest-risk items in the room. Each must be dispositioned before the gate can pass: the user owns it (moves it to Resolved with a why) or it is reverted or split into its own change.
Then a report, grouped by intent: what changed, how it interacts with existing code paths, and the 2–3 mental-model updates the user should walk away with.
Then the quiz, scaled to the fog it guards: if reconciliation found no unlogged deviations and the ledger holds fewer than ~4 entries, one round of 2–3 questions suffices. Otherwise a round is 5–8 questions drawn from the ledger — resolved decisions, deviations (especially just-discovered ones), standing assumptions — asked and graded one question at a time, never in bulk. Mix recall ("what happens to in-flight jobs during deploy now?") with prediction ("someone calls X with a stale token — what do they see?"). Weight toward what would bite an unaware maintainer.
Grade honestly. A miss is either a gap in the user's model or a change too clever — say which. Never pass out of politeness. A round passes when every question is answered correctly. After a failed round: walk the missed ledger entries with the user, then run a second round of variant questions on the missed ground — never reuse questions verbatim. Two failed rounds → recommend splitting or simplifying the change, not a third quiz.
Pass = merge-ready. Offer the ledger's death rite (fog skill): graduate the whys worth keeping, delete the file.
