refreshing-anthropic-guidancelisted

Sweep the local Claude Code harness against Anthropic's current guidance and diff it against what was true the last time this ran, then enter plan mode with a remediation plan. **No code, config, or agent/skill/hook file is edited by this skill** — the only write is the tracker (Step 5); every actual fix goes through the user-approved plan, the same as `auditing`. **This skill must only run in the main (hub) agent, never inside a subagent.** Step 6 calls `EnterPlanMode`, and Step 3 dispatches subagents via the Agent tool; spokes carry `disallowedTools: Agent` and cannot do either. If you find yourself executing this skill as a subagent inside a larger task, stop and surface the refresh request back to the hub instead. ## Why this exists, not just `researching-anthropic-guidance` That skill answers "what does Anthropic recommend for X" once, for whatever X the invoking task names. Nothing in the repo ever asks the inverse question — is what's _already built_ still what Anthropic recommends? The harness hardcodes a lot of Anthropic-owned surface (model IDs in every agent frontmatter, hook event names, `settings.json` schema keys, `maxTurns`/ `disallowedTools`/`permissionMode` semantics) and none of the repo's `check:*` gates compare any of it to upstream — they're all closed loops that check the repo's own prior decisions against itself. This skill is the missing outward check, run periodically rather than once. ## Steps ### 1 — Read the tracker and establish anchors Read