/white-bear

name: white-bear
description: "Use when the user asks to \"check white bear\", \"audit prohibitions\", \"find negative framing\", or invokes /white-bear. Read-only audit: unnecessary competing-target mentions in LLM-facing prose."
user_invocable: true
allowed-tools: Read, Grep, Glob

White Bear Audit

A semantic audit of LLM-facing prose for the White Bear authoring principle: keep attention on the necessary path — a mention of a competing non-target (a forbidden act, a superseded path, a rejected alternative) earns its place only when it is load-bearing. Read-only — it emits structured findings and writes no fixes. The human author decides which to rewrite, mark as load-bearing, or dismiss.

Purpose

Surface prose that holds the model's attention on an unnecessary competing target, before it ships. **Prohibition framing** names a forbidden act, **superseded-path mention** names a retired path, **negated anchoring** names a rejected alternative — each keeps the non-target available as a competing action candidate. This drift survives deterministic structural checks — it is a meaning-level pattern, so a semantic reviewer catches what literal pattern matching cannot.

Inputs

**Manual invocation only** (interactive `/white-bear`):

• The caller passes target file paths or a glob; with no argument, the skill enumerates the in-scope set under the working tree HEAD.
• Files are read at their working-tree state — the post-edit, pre-commit content the author is about to ship.

Scope

**In scope** (LLM-facing prose where this principle applies):

• Skill instruction files (`*/skills/*/SKILL.md`), considered outside formal/definition blocks
• Agent system-prompt files (`*/agents/*.md`)
• Output-style files

**Out of scope** (positively framed by purpose, or the principle does not apply):

• Formal-definition blocks within instruction files — regions delimited by `── <NAME> ──` headers (FLOW, MORPHISM, TYPES, PHASE TRANSITIONS, and peers). Notation patterns are the content there.
• Fenced code blocks (` ``` ... ``` `) — code is content, and example code attached to a definition is part of that definition.
• Human-facing documentation (README files, design notes, reference material) — examples serve human comprehension there.
• Rule-tier and principle-tier prose authored for contributors — such prose admits intentional negative formulations as discriminant signals; a one-pass rewrite would erase calibration signals at intentionally preserved decision points.
• Session and context substrates outside this audit's surface.

What to evaluate

**The principle (stands alone).** LLM-facing instructions are followed more reliably when they state only what carries the intended behavior — a positive rationale ("X is Y because Z"), the current path, the affirmed characterization. Naming an unnecessary competing target holds attention on it — the white bear effect: "don't think of a white bear" surfaces the white bear. The human ironic-process effect does not transfer mechanistically to language models; each form carries its own operative ground:

• **Prohibition framing** ("do not use W") — weak negation processing: negative injunctions bind behavior less reliably than positive directives. The strongest-evidenced form.
• **Superseded-path mention** (a positive mention of a path the same instruction retires in favor of a replacement) — option availability: naming the retired path can keep it available as a competing action candidate. A conservative authoring heuristic rather than an established causal law.
• **Negated anchoring** ("X is not A but B" where the rejected alternative A carries no load) — the contrast anchors attention on the rejected alternative; the same discipline is recorded from practice as an editing convention (prefer positive predicates over negated anchoring).

For each in-scope file, consider every prose sentence outside formal blocks and code fences:

A **White Bear signal** is a sentence in LLM-facing prose that names a competing non-target — a forbidden act, a superseded path, or a rejected alternative — where stating only the intended path preserves the directive's force. The three named forms are working hypotheses, an open list rather than an exhaustive taxonomy; the constitutive test is necessity: a mention whose load-bearing meaning collapses without it stays compliant, and a mention the directive survives without is a finding.

**Section-level placement** — the principle's force varies by the section's role:

• **Runtime motivational prose** — Rules, Phase prose, agent system prompts: prose that directs what the LLM does at execution time. The named non-target becomes the foreground attractor during application, so apply White Bear avoidance at full strength; the rewrite tests below decide whether a mention stays.
• **Diagnostic substrate** — Anti-patterns sections, failure-case checklists, audit findings, review-vocabulary lists. The section's role is naming failure modes, superseded paths, or rejected alternatives for detection, so negative or failure-case wording is the content. Treat as compliant by purpose when the section role is visible. Surface a finding only when a rewrite would preserve both directive force *and* boundary meaning — usually `low` for human triage rather than `high`.

**Load-bearing boundary mentions** (compliant by purpose; surface only as `low` when the rewrite preference is judgment-dependent): a prohibition or competing-target mention stays when it encodes a genuine safety boundary the model observes, a contract it honors, a verification-by-design constraint, a legacy-input condition, a migration target, or a genuine fallback. A path may be retired for execution while its mention remains necessary to recognize or verify the boundary; removing such a mention would erase the boundary, so it stays.

**The rewrite test (constitutive core, all forms)**: a rewrite is valid when stating only the intended path — the positive restatement, the