How ChatGPT decides to render dashboards, charts, and reports

This is a detailed reference of the internal tool and skill instructions that govern how ChatGPT (and the Codex tooling behind it) renders data artifacts — dashboards, reports, charts, and tables. The core workflow is a strict pipeline: validate a manifest and a bounded data snapshot with validate_artifact before ever calling render_artifact, because failed render attempts can leave visible broken placeholder cards in the UI. Snapshots are capped hard (50 datasets, 2,000 rows each, 3MB, 200k inline characters) and must follow a canonical shape, with table-style {columns, rows} objects explicitly rejected. Much of the document is a decision tree for which rendering surface wins — native MCP artifacts first, then hosted Sites deployment via export_artifact_package, with static HTML, images, or plain tables only as progressively degraded fallbacks.

A recurring theme is ‘Work Mode,’ a trusted rendering context in which the model is forbidden from calling render_artifact, render_chart, or render_table because that path can silently drop standalone widgets. The spec is unusually prescriptive about presentation semantics: chart titles must stay neutral and descriptive rather than editorializing a takeaway, a ‘by ’ phrase in a title is treated as a binding contract that the dimension actually appears on an axis or legend, and reports must include specific block structures (a heading matching the title, at least one chart, one markdown block per editable section). It even dictates SQL hygiene — runnable SQL in source.query.sql, human summaries kept separate, and real table names in metadata.

The significance is less about any single rule and more about the visibility: this reads like extracted system-level guidance showing how much hidden logic sits between a user’s request and a rendered chart. It exposes the guardrails LLM vendors bake in to avoid broken UI states, misleading visualizations, and leaked internals — the closing lines explicitly bar emitting hidden reasoning, credentials, or secrets. For anyone building agentic or analytics tooling on top of these models, it’s a useful map of the constraints and failure modes the platform is quietly managing.