Comparing many things across many attributes collapses into prose where no fact can be traced, counted, or rechecked on its own.
← Back to patterns
Spreadsheets for Images Living Documents as a UX Pattern in AI
Visual & Generative
EmergingSemantic Grid
One Prompt, N Independent Runs
Ask a chat to compare six vendors on three attributes and you get eighteen facts flattened into prose — one silently missing, none traceable, and rechecking a single one costs a full regeneration. Semantic Grid makes the prompt a column definition: authored once, fanned out across rows as independent runs, each cell carrying its own source and re-runnable on its own. Where Parallel Alternatives is N variants of one task, this is one task over N subjects.
Metaphor
Work Queue
Interaction
Column Definition, Cell Inspect, Per-Cell Re-run
Use case
Vendor Comparison, Literature Review, Lead Enrichment, Competitive Research
Control mechanisms
Editable Column Prompts · Per-Cell Provenance · Single-Cell Retry · Run Counter
Reference examples
Elicit · Clay · Airtable AI Fields
Interactive Example
The live side-by-side is the comparison.
This prototype already contains the contrast between chat and the pattern response. It stands as the primary comparison surface for this page.
The table is a work queue, not a rendering of an answer.
Chat has exactly one unit of work: the turn. That is fine when the question has one subject. It stops being fine the moment the question has a data axis — six vendors, fifty papers, two hundred leads — because the answer then holds dozens of independent findings that can each be right, wrong, missing or stale on their own, and the transcript gives you no handle on any of them individually.
Semantic Grid takes the axis the work already has and makes it the axis of the interface. Authoring collapses to one act per attribute; execution fans out to one run per subject. That asymmetry is the whole pattern — and it is why the grid is not a nicer way to display an answer but a different shape of work.
The deep root is fill-down itself: a formula authored once and evaluated independently per row, the move that made the spreadsheet the most successful end-user programming environment ever built. What is new is that the column definition is natural language and each cell carries its own evidence — so dragging a formula down a column now works over questions that have no formula.
Ask once per attribute, run once per subject, verify once per cell.
DOPE Lens
How this pattern improves AI interaction quality.
Observability
Every cell carries its own source chip and confidence dot, and the counter states the run exactly — 18 cells · 3 low-confidence · 1 re-run · 4,210 tokens
Directability
Edit a column prompt and only that column re-runs; add a row and only its cells fill — the unit of work is the cell, not the table
Predictability
The schema is declared before anything runs: N rows × M column prompts is N×M cells, each answering the same stated question
Explainability
Open any cell for the snippet that produced it, with its own tokens and latency — provenance is per answer, not per document
Fit
Where it earns its place.
Avoid
when
when
The question has one subject: a single deep answer beats a one-row table wearing a schema.
In the wild
Where this pattern already ships.
The pattern is shipping but still maturing across products. Entries below show how different teams have approached it; expect variation as the space settles.
Reference implementations
Elicit · Clay · Airtable AI Fields
Detailed write-ups for each implementation are in progress.
Research Frame
Why this pattern exists, and where it breaks.
Make the prompt a column definition, fan it out across the rows as independent runs, and give every cell its own source and its own retry.
A transcript holds one answer per turn, so eighteen facts arrive as a single blob and rechecking any one of them costs a full regeneration of the other seventeen.
A grid of confident-looking cells invites reading it as a dataset — without per-cell provenance and a visible low-confidence count it industrialises unverified claims.
Common questions
What people ask about this pattern.
When should I avoid the Semantic Grid pattern?
+
The question has one subject: a single deep answer beats a one-row table wearing a schema.
What problem does Semantic Grid solve?
+
Comparing many things across many attributes collapses into prose where no fact can be traced, counted, or rechecked on its own.
Why is chat the wrong fit for this?
+
A transcript holds one answer per turn, so eighteen facts arrive as a single blob and rechecking any one of them costs a full regeneration of the other seventeen.
Grounded in Research
Not just opinion.
The papers, guidelines, and usability studies this pattern rests on.
Marc Levoy
Stanford University · Proc. SIGGRAPH '94, pp. 139–146 · 1994
The grid recast as a substrate for expensive computation rather than arithmetic — cells hold images and volumes, so "larger formulas, longer computation times, and more complicated intercell dependencies" become the design problem. Answers it with the two extensions the fan-out pattern still runs on: cells can be selectively disabled, and multiple cells edited at once.
Adam Wiggins
Elicit (Ought) · engineering blog · 2024
First-party account of the fan-out: papers in the rows, extractions in the columns, and "fifty papers with twenty data extractions is already 1,000 individual cells to be generated". Names both halves the pattern needs — click a cell to flip through the quotes behind it, and treat per-cell errors and retries as ordinary rather than exceptional.
Related Patterns
Curated connections, not just neighbors.
Hand-picked relationships — what this pattern pairs with in practice, what it replaces, what it builds on, and what builds on it.
Often paired with
Compute Budget
A grid is rows × columns runs; Compute Budget prices it before it fans out and meters every cell you re-run.
Extends
Parallel Alternatives
Semantic Grid takes the parallel runs of Parallel Alternatives and aims them at N subjects, not N variants of one task.
Extends
Templates
A column prompt works like a template: written once, run for every row, with only the row's subject changing.