The AI chat interface treats every task the same: type a message, get a response, paste the output somewhere else, repeat. Anton Krylov's Idea Cells proposal, published in his Sphera newsletter, argues that this is not a minor UX quibble but a fundamental category error — collapsing cognitively distinct forms of work into a single conversational thread obscures what each task actually requires. His alternative draws on the Jupyter notebook model but extends it well beyond code execution to cover the full spectrum of knowledge work.
The cell taxonomy Krylov proposes is architecturally specific. Terminal cells handle command execution and persist stdout and artifacts in place rather than letting them evaporate from a chat scroll. Writer cells expose structured editing operations — rewrite, expand, shorten, tone, continue — rather than inviting freeform prompting, producing more stable and comparable outputs. The idea cell is perhaps the most opinionated: it enforces a structured thesis format covering market framing, customer pain, business model, moat, a 90-day plan, risks, kill criteria, and a validation checklist. Downstream research, risk, and execution cells handle the operational sequencing that follows ideation. At the more exotic end, a class of formal reasoning cells — conjecture, example, counterexample, lemma, and proof-gap — decompose mathematical or logical work into auditable units closer to theorem-proving workflows than ordinary chat.
What gives the canvas its structural integrity is the linking system. Rather than feeding one cell's output as a raw text blob into the next prompt, links route specific outputs into named input fields of downstream cells — keeping multi-stage workflows typed and traceable rather than collapsing into copy-paste chains. The entire canvas is also versioned in a Git-like model, with every change to cells, links, and outputs preserved as inspectable history rather than overwritten in place, integrating version control directly into the workflow rather than treating it as an external concern.
Krylov presents Idea Cells as a design concept rather than a shipping product, which may be why its specificity is easier to appreciate than to dismiss. Most proposals for moving beyond chat AI remain abstract enough to evade scrutiny — 'agents,' 'workflows,' 'orchestration' — terms that gesture at a direction without committing to a shape. Idea Cells commits to a shape. Named cell types, typed slot routing, and versioned state are all claims that can be tested against actual use. The notebook displaced the REPL as the default environment for serious computational work not through argument but through contact with daily practice. Whether the same logic applies to knowledge work more broadly is the question Krylov is implicitly posing.