blank

/claire:blank — cold outside read (Claire)

A shortcut to the blank-slate primitive in /claire:challenge. Steps:

0. Version line. Run the status script adv-status.sh (at the plugin root — the parent of the skills/ directory this skill lives in) and print its one-line output at the very top of your response (e.g. claire v0.2.0 · up to date). This lets the user confirm, across machines, that this machine is on the latest version. If it reports commits behind origin, surface that the plugin can be updated with git pull before proceeding.
1. Take the user's decision / situation (in $ARGUMENTS or the surrounding message).
2. Apply the de-priming checklist from /claire:challenge Step 2 — translate the situation into neutral, two-sided, jargon-free plain language; strip your own lean and the options you have already ruled out.
3. Leak-check the brief, then show it. Pick a fresh, unique receipt id (any short token). Dispatch brief-leak-auditor with [CLAIRE-RECEIPT:<id>] on its own line, then the situation text itself — blank-slate carries no persona/attack-license preamble, so the audited text is just the situation (the auditor ignores the id line as ambient). If you paste in any document the advisor will read (a draft, notes), audit the whole situation + document as you will send it — the brief the auditor reads is exactly what the gate injects into the advisor, so never audit the framing alone with the document added after. On a clean verdict a receipt is written, keyed by that id, storing the audited situation; you then dispatch the advisor with the same [CLAIRE-RECEIPT:<id>] and the gate injects that stored situation, replacing whatever else is in the dispatch. If it leans, follow the bounded fix loop in /claire:challenge Step 3.1 — paste the auditor's neutral rewrite verbatim, re-audit, cap at 2 cycles, and escalate to the user if it still leans rather than proceeding. A clean verdict writes the receipt the gate needs; quoting an id without an actual clean audit writes no receipt and the gate fails closed. Do NOT self-certify the brief as neutral — that is the failure this step exists to stop.
4. Dispatch claire:blank-slate-advisor (the namespaced name — never bare blank-slate-advisor, which a workspace's own same-named local agent could shadow) carrying the same [CLAIRE-RECEIPT:<id>] — the gate injects the audited situation; no files, no rationale, no history reach the advisor.
   • If blank-slate-advisor is not loaded (its agent-file changes need a session restart, so after a git pull the named agent may be absent): do not fall back to a built-in Plan/Explore agent. Under injection the gate would overwrite its inlined persona with the audited situation, losing the cold-read guardrail. Tell the user the Claire agents aren't loaded on this machine and to restart so the plugin agents load, then re-run.
5. Return Claire's read verbatim, then offer to unpack it. If the user wants to track whether her calls hold up over time, log it for outcome-scoring: adv-score.sh add blank "<one-line topic>" (prints an id), and later adv-score.sh verdict <id> right|wrong|partial. See README → Outcome scoring.

Debug mode (CLAIRE_DEBUG) is governed by /claire:challenge Step 3: if a [CLAIRE TRACE] block appears in your context, surface the apparatus (the auditor verdict, the gate/receipt trace lines, the raw dispatch) as a clearly-marked TRACE section after Claire's read; otherwise keep it behind the scenes and show only the neutral brief and her read.

The de-priming standard lives in /claire:challenge Step 2 — this skill points to it rather than restating it, so there is one canonical copy.