vigiles: Skill for Claude Code

Turn a prose rule that no off-the-shelf linter rule matches into a custom lint rule — the opt-in synthesis step. This is the hand-off target of the custom rule (⚙) lane in a vigiles audit rule map: audit maps a rule there when it looks enforceable but nothing off-the-shelf fits. You invoke this skill explicitly; nothing here runs on its own, and installing vigiles never starts synthesizing anything.

The point of this skill is not "write a checker." A model can write a plausible-looking checker in seconds. The point is the discipline that makes the result trustworthy: synthesize the rule and an independent test that encodes the rule's real intent, run the checker against adversarial cases it didn't author, and abstain — hand it back as prose — if it leaks. A checker that matches a rule's surface but not its intent gives false confidence (the measured failure mode: most naively-synthesized checkers silently leak). Shipping a green check nobody should trust is worse than shipping nothing.

Arguments

$ARGUMENTS — a prose rule to enforce. Either a custom rule (⚙) line copied from a vigiles audit report, or free text. Examples:

• "we keep telling people not to import directly from antd — use our design-system barrel"
• "people forget to use our custom logger instead of console.log"
• "wrap outbound API calls in our withRetry helper"
• "route handlers must go through the withAuth wrapper"

The pipeline

prose rule
  → detect language + linter (ask if ambiguous)
  → REUSE first: does an off-the-shelf rule already fit? → use it, STOP (no synthesis)
  → clarify intent + read the codebase (what exactly is the violation? the fix?)
  → SYNTHESIZE: the rule  +  an INDEPENDENT intent-encoding test (adversarial cases)
  → TRUST GATE: run the checker on that test; precision AND recall must be 1.0
        pass → KEPT     (safe to enforce; wire it in)
        leak → ABSTAIN  (hand back as prose; never ship a checker it can't prove sound)