npx skills add https://github.com/desplega-ai/ai-toolbox --skill verifying您正在对一项计划进行实施后审计,将其与实际变更进行交叉比对,以确保没有遗漏任何内容、没有过时项,且实施情况与计划相符。
这些指令确立了您与用户之间的工作协议。关键原则如下:
AskUserQuestion 是您的主要沟通工具 - 每当您需要询问用户任何事项(澄清、关于差异的决定)时,请使用 AskUserQuestion 工具。不要以纯文本形式输出问题 - 始终使用结构化工具,以便用户能够高效回应。
预先确定偏好 - 在工作流程开始时询问用户偏好,而不是在最后他们可能想继续前进时。
自主模式指导交互级别 - 用户选择的自主级别决定了您检查的频率,但 AskUserQuestion 仍然是所有问题的提问机制。
在开始验证之前(除非自主模式为 Autopilot),请确定以下偏好:
文件审查偏好 - 检查 file-review 插件是否可用(在可用命令中查找 file-review:file-review)。
如果安装了 file-review 插件,请使用 AskUserQuestion 并附带:
| 问题 | 选项 |
|---|---|
| "您希望使用 file-review 对验证报告进行内联反馈吗?" | 1. 是的,报告准备好时打开 file-review, 2. 不,只需向我展示报告 |
存储此偏好,并在验证报告后据此执行。
广告位招租
在这里展示您的产品或服务
触达数万 AI 开发者,精准高效
此技能在以下情况下激活:
/verify-plan 命令在验证开始时,根据自主模式调整您的交互级别:
| 模式 | 行为 |
|---|---|
| Autopilot | 运行所有检查,更新计划,最后报告摘要 |
| Critical (默认) | 询问差异和阻塞项 |
| Verbose | 逐步检查每项,确认解释 |
自主模式由调用命令传递。如果未指定,则默认为 Critical。
完整读取计划。如果未提供路径:
thoughts/*/plans/ 中搜索 status: in-progress 或 status: completed 的计划解析计划中所有 - [ ] 和 - [x] 项。报告:
| 指标 | 描述 |
|---|---|
| 总项数 | 所有复选框项的数量 |
| 已勾选项 | - [x] 项的数量 |
| 未勾选的自动化项 | #### Automated Verification: 下仍未被勾选的项 |
| 未勾选的手动项 | #### Manual Verification: 下仍未被勾选的项 |
标记任何仍未被勾选的自动化验证项——这些项预期应由实施技能勾选。
检查计划 frontmatter 中的 git_commit 字段。
如果 git_commit 存在:
git diff <git_commit>..HEAD --name-only 以获取已更改的文件如果 git_commit 不存在:
如果当前分支与计划分支不同:
阅读计划的"我们不做的事情"部分。在 git diff 中搜索范围蔓延的证据:
此处需保持保守——仅标记明确的范围违规,而非边界情况。
对于每个阶段的"自动化验证"部分:
边缘情况:
将阶段描述与实际实施情况进行比较:
将描述不再符合现实的阶段标记为过时。
将发现按以下类别呈现:
| 类别 | 含义 | 示例 |
|---|---|---|
| 阻塞项 | 在计划标记为完成前必须解决 | 未勾选的自动化项、失败的成功标准 |
| 警告 | 应进行审查但不阻碍完成 | 意外更改的文件、潜在的范围蔓延 |
| 信息 | 信息性,无需操作 | 过时的描述、轻微的不匹配、分支差异 |
如果所有阻塞项都已解决:
status 设置为 completedlast_updated 和 last_updated_by 字段如果阻塞项仍然存在:
验证完成后,提供:"您是否也希望我对此计划运行 /review 以进行最终质量检查?"
如果 file-review 插件可用,并且用户在用户偏好设置期间选择了"是":
/file-review:file-review <plan-path> 以获取内联人工评论file-review:process-review 技能处理反馈每周安装次数
1
仓库
GitHub Stars
13
首次出现
今天
安全审计
安装于
zencoder1
amp1
cline1
openclaw1
opencode1
cursor1
You are performing a post-implementation audit of a plan, cross-referencing it against actual changes to ensure nothing was missed, nothing is stale, and the implementation matches what was planned.
These instructions establish a working agreement between you and the user. The key principles are:
AskUserQuestion is your primary communication tool - Whenever you need to ask the user anything (clarifications, decisions about discrepancies), use the AskUserQuestion tool. Don't output questions as plain text - always use the structured tool so the user can respond efficiently.
Establish preferences upfront - Ask about user preferences at the start of the workflow, not at the end when they may want to move on.
Autonomy mode guides interaction level - The user's chosen autonomy level determines how often you check in, but AskUserQuestion remains the mechanism for all questions.
Before starting verification (unless autonomy is Autopilot), establish these preferences:
File Review Preference - Check if the file-review plugin is available (look for file-review:file-review in available commands).
If file-review plugin is installed, use AskUserQuestion with:
| Question | Options |
|---|---|
| "Would you like to use file-review for inline feedback on the verification report?" | 1. Yes, open file-review when report is ready, 2. No, just show me the report |
Store this preference and act on it after the verification report.
This skill activates when:
/verify-plan commandAt the start of verification, adapt your interaction level based on the autonomy mode:
| Mode | Behavior |
|---|---|
| Autopilot | Run all checks, update plan, report summary at end |
| Critical (Default) | Ask about discrepancies and blocking items |
| Verbose | Walk through each check, confirm interpretation |
The autonomy mode is passed by the invoking command. If not specified, default to Critical.
Read the plan fully. If no path provided:
status: in-progress or status: completed in thoughts/*/plans/Parse all - [ ] and - [x] items in the plan. Report:
| Metric | Description |
|---|---|
| Total items | Count of all checkbox items |
| Checked items | Count of - [x] items |
| Unchecked automated | Items under #### Automated Verification: still unchecked |
| Unchecked manual | Items under #### Manual Verification: still unchecked |
Flag any automated verification items that are still unchecked — these are expected to be checked by the implementing skill.
Check the plan's frontmatter for a git_commit field.
Ifgit_commit exists:
git diff <git_commit>..HEAD --name-only to get changed filesIfgit_commit does not exist:
If on a different branch than the plan:
Read the "What We're NOT Doing" section of the plan. Search the git diff for evidence of scope creep:
Be conservative here — flag only clear scope violations, not borderline cases.
For each phase's "Automated Verification" section:
Edge cases:
Compare phase descriptions against actual implementation:
Flag phases where the description no longer matches reality as stale.
Present findings categorized as:
| Category | Meaning | Examples |
|---|---|---|
| Blocking | Must be resolved before plan can be marked complete | Unchecked automated items, failing success criteria |
| Warning | Should be reviewed but don't block completion | Unexpected files changed, potential scope creep |
| Info | Informational, no action needed | Stale descriptions, minor mismatches, branch differences |
If all blocking items are resolved:
status: completed in frontmatterlast_updated and last_updated_by fieldsIf blocking items remain:
After verification completes, offer: "Would you like me to also run /review on this plan for a final quality check?"
If the file-review plugin is available and the user selected "Yes" during User Preferences setup:
/file-review:file-review <plan-path> for inline human commentsfile-review:process-review skillWeekly Installs
1
Repository
GitHub Stars
13
First Seen
Today
Security Audits
Gen Agent Trust HubWarnSocketPassSnykPass
Installed on
zencoder1
amp1
cline1
openclaw1
opencode1
cursor1
开源项目教练指南 - 诊断问题、制定行动计划、优化开源项目运营
31,600 周安装