claire

Claire

A de-priming adversary for Claude Code. Claire gives you an outside read on a decision, plan, draft, or argument — from a critic that was never told what answer you were hoping for.

She is built for decisions, not just code. Most adversarial plugins assume you are a developer reviewing a diff. Claire is for the judgement calls that sit underneath that: which research direction to take, whether an argument holds, how to frame a piece of writing, whether a plan survives contact with reality.

Why a separate critic, instead of asking Claude to check its own work

There is a documented blind spot here. When a language model is asked to review and revise its own reasoning without new external information, it tends to talk itself out of correct answers as often as wrong ones — the act of "reflecting" frequently reacts to the question's framing rather than to the substance. (See research on the self-correction blind spot in LLMs: self-revision tends to work when there is fresh ground truth to react to, not from pure introspection.)

A separately briefed critic does better — but only if its brief doesn't quietly hand it the conclusion. The well-documented failure here is anchoring: if the critic can see how you framed the problem, which option you already lean toward, or which side you wrote up as a worry versus a settled decision, it tends to evaluate your framing instead of the actual claim. Anchoring bias replaces adversarial force.

Claire's whole design is one response to those two facts:

1. She is context-starved by design. She gets the claim and the minimum needed to challenge it — not your reasoning chain, not your tool history, not the story of how you got there.
2. Her brief is leak-checked before she ever sees it. A separate read-only pass inspects the brief for tells that smuggle in your preferred answer (one option's costs painted vividly while the other's are glossed; your favoured side framed as a live worry and the rest as already-decided; a problem statement only one option happens to solve). A producer cannot reliably certify their own brief as neutral — that judgement is exactly the thing anchoring corrupts — so the check is done by a fresh reader, not asserted by you.
3. You see the neutral brief before dispatch. What Claire is actually being asked is shown to you first, so the de-priming is auditable rather than taken on trust.

That is the edge, stated plainly: not a smarter critic, but a genuinely blank-slate one whose brief has been checked so it can't agree with you by accident.

Commands

• /claire:challenge — Give me an adversary. Routes your request to the right kind of critic (a cold outside read, a plan-attacker, an actor role-play, a source-versus-claim check, a two-sides face-off, or a check that a test isn't rigged), strips the brief of the answer you expect, and shows you that neutral brief before dispatching. Use it for "attack this", "outside view on this", "what am I missing", "how would X react", "is my test rigged".

• /claire:blank — A pure cold read. Hand Claire a question or a piece of work with no surrounding context and get back the read of someone seeing it fresh, with no stake in the prior answer.

• /claire:doctor — Health- and conflict-check this install. Verifies dependencies and integrity, flags duplicate installs and agent-name clashes with your current workspace, and runs a live self-test confirming the de-priming enforcement is actually firing here (so you know whether strict mode is safe to turn on). Run it after installing or updating.

How she differs

A second opinion from the same model, asked neutrally, tends to drift back toward agreeing with you — it shares your priors and politely re-derives your view. Claire instead pairs a blank slate with a leak-checked, de-primed brief, so the disagreement you get is real rather than performative. The same de-priming discipline is what makes her useful well outside code review.

Enforcing the de-priming

The leak-check isn't a polite suggestion the assistant can talk itself out of. A built-in gate watches every critic dispatch and stays quiet only once the leak-auditor has actually cleared that exact brief — proven by a short-lived receipt written when the audit passes, not by the assistant claiming it did the step. If a brief never passes the auditor in two tries, the assistant is told to stop and show you the lean it keeps finding rather than push on regardless.

By default the gate warns (it never blocks your work on a public install). If you want it to hard-stop a critic that skipped the audit, set the environment variable CLAIRE_GATE_STRICT=1 on that machine — but first run a real /claire:challenge once and confirm the gate goes quiet on the genuine dispatch (i.e. receipts are being written on your setup), so strict mode blocks only true skips, not a payload-shape quirk.