What they did
toshipon, an SRE and indie developer in Japan, does security reviews for clients with Claude Code. An earlier skill told the agent where to look. After using it on real engagements they wrote a second skill that controls how the agent reaches conclusions. Each suspected issue is written as a claim about what an attacker could do, along with what evidence would disprove it, before any evidence is gathered. An orchestrating session hands backend, frontend, secrets and compliance checks to subagents, which only return candidates. The orchestrator re-reads the real files before promoting anything. Checks that static reading cannot settle go on an UNKNOWN list. A small CLI keeps the state on disk and refuses to publish a report that breaks the evidence rules.
How it went
Reports got shorter and more useful, which the author sums up as six real findings in place of thirty mostly noisy ones. Looking for the disproving evidence first caught false alarms: an endpoint that looked like it rewrote production data only ran outside production, and a webhook already checked its signatures. A subagent flagged personal data in source comments as a high concern; the orchestrator opened the file and found keyboard-mash test strings. The author gives no time or cost figures.
Worth knowing
It is slower than a pattern scan by design and does not replace a security engineer; critical and high findings are flagged for human review. By default it only reads code and config and sends no traffic to the target. The author sells this kind of review, so the write-up doubles as promotion.