In a recent LinkedIn post, Dan Sherrard-Smith delves into the emergent capabilities of Anthropic’s Claude AI, highlighting a critical, yet often overlooked, aspect of its reasoning process. Sherrard-Smith draws attention to recent research from Anthropic that reveals an internal workspace, dubbed ‘J-space’, which appears to be a key driver of Claude’s more complex problem-solving abilities.
The core of Sherrard-Smith’s analysis centers on the distinction between Claude’s visible output and its internal cognitive processes. While many users might assume the AI’s reasoning is directly reflected in its step-by-step output, the research suggests otherwise. According to Sherrard-Smith, the ‘real thinking’ occurs in this hidden ‘J-space’.
“Most people think Claude reasons the way you see it reason. The visible chain-of-thought. The step-by-step output. But that’s not where the real thinking happens.”
The Emergent ‘J-space’ Workspace
As Dan Sherrard-Smith explains, this ‘J-space’ is not explicitly programmed but emerges organically during the AI’s training. He likens it to an internal notepad where Claude holds active concepts and uses them to steer its answers. Crucially, the contents of this workspace are not visible in the AI’s output.
To illustrate the significance of J-space, Sherrard-Smith points to experimental findings detailed in the Anthropic research. He notes:
“The researchers tested it directly. → They swapped an internal ‘spider’ pattern for ‘ant’ mid-question about legs. Claude’s answer flipped from 8 to 6. The visible text didn’t change. The hidden workspace did. The answer changed.”
Furthermore, Sherrard-Smith highlights that even when the J-space was entirely deleted, Claude could still perform basic chat functions and recall facts. However, its ability to complete multi-step problems significantly degraded. This observation underscores the ‘J-space’ as the locus of deeper, more complex reasoning within Claude.
Rethinking Prompting Strategies
Based on these insights, Dan Sherrard-Smith argues for a fundamental shift in how users prompt Claude, particularly for complex tasks. He contends that optimizing only for the AI’s written output misses the crucial pre-computation phase.
Stocking the ‘Notepad’ Upfront
In Sherrard-Smith’s view, effective prompting involves more than just refining the final output. Instead, it requires providing Claude with a richer, more structured context before it begins generating a response. He elaborates on this strategy:
“If you’re only optimising what Claude writes back to you, you’re working on the wrong layer. The reasoning that drives complex outputs happens before the first word appears. The fix: give Claude more structured problem context upfront. Not ‘write me a LinkedIn post about X.’ Give it the goal, the constraint, the audience, the format. All before it starts.”
He frames this approach not as writing a better prompt, but as effectively preparing the AI’s internal workspace for the task ahead. Sherrard-Smith suggests that this shift is particularly impactful for applications requiring multi-step completions, such as lead generation sequences, sales emails, and content systems.
By understanding and leveraging the concept of Claude’s emergent ‘J-space’, users can move beyond surface-level interactions to unlock more sophisticated and effective AI-driven outputs, as Dan Sherrard-Smith has clearly articulated in his recent post.
📝 About This Content
This article is based on insights shared by Dan Sherrard-Smith on LinkedIn.
📅 Originally posted on July 14, 2026 | View original post on LinkedIn →