AI Agent Output → Editable Work Artifact: a reproducible Inplain test
One public Markdown brief and CSV become PPTX, DOCX, and XLSX. Download every file, inspect the Open XML verification report, and see the limits we do not claim to solve.
The claim being tested
AI agents are increasingly good at producing research, analysis, tables, and recommendations. The unresolved step is the handoff: turning that output into something another person can read, edit, review, and share without returning to the original chat. We call that result an editable work artifact.
The useful question is not whether a product can generate an attractive screenshot. It is whether the same verified source can survive as a presentation, document, spreadsheet, or live link; whether the recipient can inspect the result; and whether the product states its format boundaries honestly.
| Public input | Generated output | What the verifier checks | |
|---|---|---|---|
| Deck | Weekly Growth Markdown + CSV metrics | PPTX with five slides | Open XML package, slide count, text runs, shape nodes, tables, charts |
| Doc | The same decision summary and source labels | DOCX report | Open XML package, paragraphs, headings, text runs, tables |
| Sheet | The same CSV plus the Markdown narrative | XLSX workbook | Open XML package, worksheets, cells, numeric cells, formulas, charts |
Download the source before reading the conclusion
- Weekly Growth W23 · Markdown source
- Weekly Growth W23 · CSV source
- Generated PPTX sample
- Generated DOCX sample
- Generated XLSX sample
- Machine-readable verification report · hashes, checks, and boundaries
The source names the reporting date, owner, decision summary, three verified data sources, and three next actions. The generated files should preserve those facts. If they do not, that is a product defect, not a difference of opinion.
What the current structural test verifies
The verifier opens each Office file as a ZIP-based Office Open XML package. It checks required package entries, computes a SHA-256 hash, and counts native structures. For PPTX it counts slides, text runs, shapes, tables, and charts. For DOCX it counts paragraphs, heading paragraphs, text runs, and tables. For XLSX it counts worksheets, cells, numeric cells, formulas, and charts.
This is deliberately narrower than a visual benchmark. A valid shape node proves that the presentation contains editable structure rather than one flattened screenshot. It does not prove that PowerPoint, Keynote, and WPS will render every font and line break identically. Visual comparison across Office clients remains a separate test.
What we do not claim
- No formula-preservation claim. The current XLSX export keeps native cells and worksheets, but it does not preserve formulas from the live Sheet model.
- No native Excel chart claim. The chart source data can be exported; interactive charts remain part of the live web artifact.
- No pixel-identical web-to-Office claim. Motion, responsive layout, embedded media, and browser interactions cannot all survive a file-format fallback.
- No independent endorsement claim. Inplain publishes this test. Independent reviewers and customers must make their own assessment.
How this differs from a Gamma alternative page
Gamma is a strong presentation-first product. Microsoft 365 and Google Workspace are strong native-office environments. Quarto, Marp, and Observable are strong code-first publishers. The Inplain category is narrower: it is a delivery layer after an AI agent finishes useful work. Structured source enters through MCP, a CLI, or the web workspace; a human receives an editable, link-first artifact; Office files remain compatibility outputs when a downstream workflow requires them.
That positioning should be evaluated through the whole chain, not a single generated slide. Can the agent create the artifact? Can a human edit it? Can the agent revise it without discarding the human work? Can the same facts appear in a presentation, report, and workbook? Can the recipient open one stable link? The public samples cover the export portion of that chain. Public share links, version history, and repeated agent edits require separate product tests.
Security and maturity evidence
Export structure is only one adoption criterion. Review the current security posture, release history, pricing and credit model, and MCP capability matrix before using Inplain for sensitive or business-critical work. Those pages distinguish available features from roadmap items and name current limitations rather than presenting them as completed certifications.
How to reproduce the test
The sample generator and verifier live in the public Inplain repository. From a checked-out copy, install dependencies and run:
pnpm evidence:officeThe command regenerates all three files from the checked-in Markdown and CSV, then writes the JSON verification report. A reviewer can compare hashes, inspect the XML package, and open each file in their own Office client.
What evidence is still missing
This benchmark does not replace independent reviews or customer proof. The next meaningful evidence is an authorized case showing an external user completing the full loop: an agent produces the source, a human edits the artifact, the agent makes a later revision, the recipient opens the same link, and the final Office fallback is accepted in a real workflow. Until those cases exist, Inplain should be described as an emerging agent-to-artifact platform with reproducible first-party tests, not as a universally proven Office replacement.