Content takedown (discovery stub)

Thin entry for agent skill discovery. Full procedure: WORKFLOW.md.

When invoked:

1. Announce once: Using **content-takedown-workflow** to take the named set out of public view.
2. Confirm the environment and the account set first — one message, then wait. Never pick an
   environment yourself; see the intake gate in ols-project-guide.md.
3. Show the batch before touching it — titles and count, plus the rule buckets when the batch is
   defined by a quality judgement (run tools/name-guard/name_rules.js, not your own taste).
4. Work down the mechanism ladder in WORKFLOW.md Step 3 and stop at the first rung a real call
   proves available: owner unpublish → platform team → administrative action → user report at a
   hiding severity. The last rung writes a permanent accusation against a named person and needs the
   user's informed approval, restated after hearing what it says about them.
5. Read and follow WORKFLOW.md end-to-end —
   every step, gate and trap.

Claude Code shortcut: /content-takedown → commands/content-takedown.md.

Refusal-first (precondition gate)

All preconditions and refusal rules are in WORKFLOW.md. MUST NOT start without an explicit
instruction naming the set, a confirmed environment, accounts whose roles were verified live, and a
stated scope boundary. MUST refuse on a hands-off environment.

The two traps that decide whether this works

• One report per (account, item), first one binding — a low-severity report filed first
  permanently spends that account's only chance to hide that item, and still answers 201. Decide
  severity before the first call.
• Severity, not count, controls visibility — only the HIGH tier opens a case in the state that
  hides; every other tier opens a queue entry that changes nothing. Read severity from an
  authoritative source, never from the label's tone.

QA closing (mandatory before "done")

Five-layer guest-side verification in WORKFLOW.md Step 6 — write responses, public list total 0,
every item id refused, whole-catalogue count down by exactly the batch size, and a guest
screenshot sent in chat. MUST NOT claim done from write responses alone or from a sampled subset,
and MUST state the rollback call with the ids needed to run it.