Why Brand Voice Holds or Slips Inside Claude

TRANSMISSION RECEIVED · PLANET COGNITION-1 · COORDS [0.12, 0.34]

Anthropic has located a small cluster of “glowing” neural patterns inside Claude, nicknamed J-space. It functions as the model’s central broadcast station: only a handful of concepts can enter at once, but once they do, they are amplified across the entire network. The finding reframes why some prompts cohere and others drift.

Inside the global workspace

The research, published as “A global workspace in language models,” borrows a decades-old idea from cognitive science called Global Workspace Theory. In the human brain, the theory proposes that many specialized modules process information in parallel, while a shared workspace integrates a few of those signals into a single coherent broadcast the whole system can access. Anthropic found an analogous structure inside Claude: a sparse, high-amplitude set of features that lights up when the model does integrative thinking and stays quiet during routine generation.

J-space is tiny. It accounts for less than a tenth of the model’s internal activity. The vast majority of Claude’s computation—fluent grammar, common facts, continuing a sentence—happens in specialized circuits that never touch the workspace. To find J-space, researchers used circuit analysis tools to trace which features fire together when Claude integrates information, then confirmed the result by lesioning those units: suppressing them degraded multi-step reasoning while leaving fluent generation intact.

What does reach J-space gets broadcast. A concept that enters can simultaneously drive many downstream tasks. In one experiment, swapping the word “France” for “China” in the workspace caused the capital, language, continent, and currency to all update together, as if a single switch rewired four answers at once. This is the signature of global access: one concept, many consumers.

The workspace switches on for cognitively demanding work—multi-step reasoning, summarization, rhymed verse, reconciling conflicting constraints. For a simple continuation, it stays dark. This explains why a model can be grammatically flawless yet miss a brand’s soul: the generation runs on autopilot circuits and never routes the brand context through the workspace. The workspace is also capacity-limited: only a few concepts can occupy it at once, which is why overloading a prompt with a dozen brand pillars tends to dilute them rather than reinforce any single one.

The most counterintuitive finding is the white-bear effect. When researchers told Claude not to think about a particular concept, the concept lit up in J-space anyway, and neighboring features flagged “damn” and “failure” lit up too, as if the model recognized it had fumbled. Negative instructions do not merely fail; they actively push the forbidden concept into the broadcast. Coherence, then, is not a property of the text but of whether the governing concepts reached the workspace.

Why marketers should care

For marketers, J-space is the missing explanation for a familiar frustration: the same brief sometimes yields a campaign that feels unified, and sometimes a string of competent-but-disconnected sentences. The difference is whether the brand’s core concepts entered the workspace. There is also the white-bear trap. Instructions like “don’t mention the competitor” or “avoid hype words” can backfire by pushing exactly those ideas into broadcast. Positive framing—“respond with our durability claim”—puts the right concept on stage instead. Once a concept is broadcast, it propagates, so the goal is to load the right concept once at the top rather than restating brand rules at every turn.

How to use it

• Anchor tone before splitting tasks. Pin the brand’s core in the first lines of the prompt (“premium, restrained cool, no sloganeering”), then ask for multiple variants—the concept enters the workspace once and is shared.
• Make multi-step tasks externalize their thinking. List priorities, pick an angle, then write—pushing intermediate results into the workspace so the next step inherits them.
• Replace “don’t” with “instead of”. Say “use our durability claim” rather than “don’t mention the competitor”—put the concept you want into the broadcast, not the one you are forbidding.

// END OF LOG