autonomous-investigation — Claude Code skill: a protocol

What autonomous-investigation does

This is a contract, not a report generator. The skill spells out how investigation skills should behave: the agent may ask at most a few clarifying questions, and once that budget runs out or nobody replies, it carries on and flags its guesses. An unattended run therefore slows down rather than stalling.

Work starts with a three-point plan: what gets searched, which kinds of sources, how facts will be told apart from interpretation. Then each key claim gets one label. Fact means a checkable URL is attached. Inference means the evidence is cited and you judge the jump. Assumption means a working guess kept for later validation. A separate gaps list holds whatever could not be found, and every domain names its own forbidden fabrications.

Section order stays fixed from run to run, so output N and output N+1 diff cleanly. Signal strength is graded by how many independent channels agree: one is a watch item, two make a working hypothesis, three or more justify briefing leadership. Collection stays within public sources, and the skill rules out pretexting or breaking terms you accepted.