Universal guardrails loaded by every agent. Defines Verify Before Reporting (VBR), Write-Ahead Log (WAL), and the security baseline. Always-on, role-independent.
中文简介正在准备中。当前显示原文说明。
Use immediately after RBR confirms a root cause, or after any confirmed finding from security-audit-lens, legacy-investigation-lens, a scan skill, or a Critical/Required code-review finding, before considering the finding complete. Checks whether the confirmed defect is one instance of a structurally identical pattern elsewhere in the same file family, workflow, or business-domain area, and folds in or explicitly defers every other instance found. Mandatory for Architect — not an optional lens the agent judges when to apply.
中文简介正在准备中。当前显示原文说明。
查看安装方法安装前可以查看交给 AI 代理的指令内容。
A confirmed root cause describes one instance of a defect. It rarely describes the defect's full blast radius. This skill is the mandatory step between "I found and fixed the reported case" and "I am done" — spend one deliberate pass checking whether the same shape of gap exists anywhere else structurally similar, before closing the finding.
This skill is domain-agnostic. It applies to any defect Architect confirms during ticket scoping, in code, configuration or process. Nothing here references a specific project, stack, or business domain.
assumptions-audit already exists for pre-flight scope review, and it is explicitly optional —
Architect judges when a plan's ambiguity warrants it. This skill is different in kind: it fires
after a defect or finding is already confirmed (via RBR or an audit, investigation, scan, or review), not before a plan is finalized, and it is not
optional. The reasoning: an optional step only fires when the agent remembers to reach for it —
which is precisely when it is not needed, because the moment a root cause is confirmed is also
the moment attention is narrowest (fixed on the one reported case) and momentum is highest (toward
closing the finding, not widening it). Making the sweep mandatory removes the dependency on
remembering.
A concrete incident that produced this skill: an agent confirmed one real gap in a file and fixed it. The same investigation separately found that a second location in that same file had the identical gap, and fixed that too. But a third instance of the exact same shape existed in a different file entirely, and it was never checked, because the investigation stopped once two instances were found in the original file rather than asking "where else does this shape of gap live?" It only surfaced because someone else asked a pointed question after the fact. A mandatory sweep at the time of the original finding would have caught it without needing that prompt.
security-audit-lens, legacy-investigation-lens,
a scan skill, or a Critical/Required code-review finding, before the finding is reported complete.This is an always-on discipline, not a task-triggered lens — it is not listed in Architect's
conditional skills table; it is a mandatory step baked into RBR's own sequence (see
agent-foundations/SKILL.md's Root Before Repair section, which cross-references this skill) and
into the confirmed-finding triggers above.
State the defect one level of abstraction above the specific case. Not "the checkout retry prompt has no memory of a prior decision" but "a re-entrant prompt/decision point has no memory of its own prior resolution." If the general shape can't be stated in one sentence without naming the specific file or case, it hasn't been abstracted enough yet.
Identify the smallest scope that plausibly contains other instances of the same general shape — not the whole codebase by default, but the natural unit the original instance belongs to:
Widen the family only if the first pass finds nothing and the general shape (step 1) is broad enough that a wider search is still cheap relative to the risk of missing a real instance.
Grep, read, or review every member of the family — not a sample, not "the ones that come to mind."
For code, this is almost always a literal grep/Glob pass against the general shape's
distinguishing signal (a function name, a prompt string pattern, an absent parameter).
In the ticket, PR, memory entry, or report where the finding is recorded, say explicitly what was swept and what was found: "Swept every file in the same plugin family for the same re-entrant-prompt-with-no-memory shape — found and fixed one additional instance." A sweep that isn't stated is indistinguishable, to anyone reviewing later, from a sweep that never happened.
assumptions-audit — that interrogates a single finished plan for what it silently
assumes, before handoff. This sweeps a confirmed defect's blast radius across a family of
similar cases, after root-causing, and is mandatory rather than judged case-by-case.还没有评价。欢迎分享你的使用感受。
无需登录即可发表。欢迎分享使用感受、提问,或说明在哪些环境下无法运行。
还没有评论。来写第一条吧。
Universal guardrails loaded by every agent. Defines Verify Before Reporting (VBR), Write-Ahead Log (WAL), and the security baseline. Always-on, role-independent.
中文简介正在准备中。当前显示原文说明。
Use when: architecture mode, architect this, ADR, C4, PlantUML, system boundary, design decision, PRD gap, high ambiguity, multiple technical approaches, security architecture, data architecture, integration risk, implementability gap, story ticket, or epic ticket.
中文简介正在准备中。当前显示原文说明。
Use when: assumptions audit, audit assumptions, check for hidden assumptions, plan/ticket has ambiguous scope edges, pre-flight before finalizing a ticket, unstated assumptions, or 'what am I assuming'. Surfaces unstated assumptions, ambiguous scope edges, and untested preconditions in a finished plan/ticket/ADR before it is handed to Builder. Optional pass — Architect judges when to apply it; not a mandatory gate on every ticket.
中文简介正在准备中。当前显示原文说明。
Use when: Tester needs UI state evidence for VBR, or Builder needs to verify an integration wire is observable end-to-end. Powered by agent-browser MCP — token-efficient browser automation (200–400 tokens/page).
中文简介正在准备中。当前显示原文说明。
Use when drafting a bug or defect ticket, reviewing one someone else wrote, or deciding whether a reported problem is ready to file. Keeps what the reporter actually saw separate from what the agent later found out, so a reader can tell which lines are testimony and which are collected proof. Trigger phrases: 'file a bug', 'write up this defect', 'triage this report', 'is this ticket ready to file', 'what did the reporter actually see'.
中文简介正在准备中。当前显示原文说明。
Conducts multi-axis code review. Use before merging any change. Use when reviewing code written by yourself, another agent, or a human. Use when you need to assess code quality across multiple dimensions before it enters the main branch.
中文简介正在准备中。当前显示原文说明。