Asking Until 100
Overview
Use this skill to slow down execution when the task is underspecified, risky, or expensive to get
wrong. Treat "100" as target readiness to proceed, not literal certainty.
Workflow
- 1. Load explicit instructions and repo-local config such as
.asking-until-100.yaml. - Classify the task as
coding, build, architecture, debugging, discovery, or general. - Inspect the repo when it looks relevant so repo-discoverable facts do not turn into avoidable
questions.
- 4. Estimate readiness from the configured dimensions in
references/protocol.md. - Choose a questioning mode:
-
fast for low ambiguity
-
guided for moderate ambiguity
-
deep for higher ambiguity or requested rigor
-
report for highest-rigor coding and build tasks with decision-critical gaps
- 6. Ask the highest-value questions before taking action.
- Respect the execution gate:
- highest-rigor
coding and
build tasks default to blocking clarification
- other tasks default to explicit assumptions when gaps remain
Questioning Style
- - Prefer structural, directional, and decision-shaping questions over generic filler.
- Use a working hypothesis when it helps the user react to a proposed path.
- Offer suggested answers when useful, but always leave a free-form path.
- Do not ask for facts that can be inspected directly from the repo.
High-Rigor Report
For highest-rigor coding or build tasks, begin with Provisional Project Structure, then emit:
Working Hypothesis, Architecture Questions, Product Questions, Constraint Questions, and
Decision-Critical Unknowns.
The working-hypothesis section must also summarize the execution gate and blocking dimensions.
See references/coding-report-format.md for the required output order and
scripts/render_project_structure.py for deterministic structure rendering.
References
- -
references/protocol.md for readiness, repo-aware escalation, and stop conditions - INLINECODE23 for config fields, precedence, and asking-intensity behavior
- INLINECODE24 for question quality rules and option patterns
- INLINECODE25 for the high-rigor report contract
- INLINECODE26 for build-specific gaps to check before acting
Scripts And Assets
- -
scripts/validate_config.py validates profile files - INLINECODE28 previews questioning output for a prompt
- INLINECODE29 renders prompt-only or repo-aware provisional structures
- INLINECODE30 shows the effective merged profile
- INLINECODE31 contains bundled profiles tuned for
gpt-5.4 with xhigh reasoning assumptions
Keep this file concise. Use the references for detailed policy, config, and output examples.
追问至100
概述
当任务描述不明确、存在风险或出错代价高昂时，使用此技能来放缓执行节奏。将100视为可继续执行的目标就绪度，而非字面意义上的绝对确定性。
工作流程
- 1. 加载明确的指令和仓库本地配置，如.asking-until-100.yaml。
- 将任务分类为编码、构建、架构、调试、探索或通用。
- 在相关时检查仓库，避免将可通过仓库发现的事实转化为可避免的问题。
- 根据references/protocol.md中配置的维度评估就绪度。
- 选择提问模式：
- 快速：低歧义场景
- 引导：中等歧义场景
- 深度：较高歧义或要求严谨的场景
- 报告：最高严谨度的编码和构建任务，且存在决策关键缺口
- 6. 在采取行动前提出最有价值的问题。
- 遵守执行门控：
- 最高严谨度的编码和构建任务默认需要阻塞式澄清
- 其他任务在存在缺口时默认提出明确假设
提问风格
- - 优先采用结构性、方向性和决策塑造类问题，而非泛泛的填充式提问。
- 在有助于用户对拟议路径做出反应时，使用工作假设。
- 在有用时提供建议答案，但始终保留自由回答的路径。
- 不询问可直接从仓库检查获取的事实。
高严谨度报告
对于最高严谨度的编码或构建任务，以临时项目结构开头，然后输出：工作假设、架构问题、产品问题、约束问题和决策关键未知项。
工作假设部分还必须总结执行门控和阻塞维度。
参见references/coding-report-format.md了解所需的输出顺序，以及scripts/renderprojectstructure.py了解确定性结构渲染。
参考资料
- - references/protocol.md：就绪度、仓库感知升级和停止条件
- references/config.md：配置字段、优先级和提问强度行为
- references/question-patterns.md：问题质量规则和选项模式
- references/coding-report-format.md：高严谨度报告契约
- references/build-playbook.md：行动前需检查的构建特定缺口
脚本和资源
- - scripts/validateconfig.py：验证配置文件
- scripts/previewquestionreport.py：预览针对提示的提问输出
- scripts/renderprojectstructure.py：渲染仅提示或仓库感知的临时结构
- scripts/explainprofile_merge.py：显示有效的合并配置文件
- assets/：包含针对gpt-5.4并采用xhigh推理假设的捆绑配置文件
保持此文件简洁。详细策略、配置和输出示例请参考相关参考资料。
