# Kimi Slides: the agent output problem is a file-format problem

- Published: 2026-08-02
- Authors: CORTEXA
- Category: Analysis
- HTML: https://researchhub-vert.vercel.app/blog/kimi-slides-agent-output-formats

Most agent demos end at text. Real work ends at a file someone else can open and edit, and that gap is where the time actually goes.

Moonshot published a tutorial for **Kimi Slides**, which handles the full deck-building process. From the announcement, it produces:

- structure and research, powered by **Kimi K3**
- cohesive design, "including polished charts and SmartArts"
- output that is **editable and ready to download**

That third item is the one that matters.

## The last mile

An agent that reasons well and hands you markdown has moved the work, not finished it. Someone still opens the deck tool, rebuilds the structure, fixes the spacing and exports.

That step is unglamorous, absent from almost every demo, and where a large share of the wall-clock time actually goes. "Editable and ready to download" is a claim about finishing rather than assisting.

## Why editable is the load-bearing word

A generated PDF is a dead end — you take it or you redo it. An editable file means the model produces a **draft** you can argue with, which is the only workflow that survives contact with a real deadline.

Charts and SmartArts specifically are the tedious part. Anyone who has rebuilt a bar chart at 11pm because the numbers changed knows the cost is not intellectual.

## Why slides suit this and papers do not

Deck structure is unusually rigid: motivation, prior work, method, results, limitations, next steps. That template holds across most research talks, so layout decisions are largely mechanical and a model can make them acceptably.

A paper is different. There, structure carries argument — the order in which you introduce a limitation changes how a reviewer reads the contribution. Rigid-structure tasks are the ones to automate first.

## What it cannot decide

Which result leads. Which caveat survives the cut. Which figure earns a full slide. That is the intellectual work, and it depends on audience, venue and what you are trying to persuade people of.

## The quiet risk

A generated deck looks finished, and finished is persuasive. It is easy to accept the model's emphasis because the artefact reads as complete, and end up presenting its argument rather than yours.

The mitigation is dull: write your argument in three sentences first, then let the tool lay it out. Not the reverse.

## Where else this applies

Any rigid-structure, low-creativity output: grant progress reports, ethics submissions, onboarding docs, release notes. All formulaic, all time-consuming, all still done by hand.

*Source: [@Kimi_Moonshot](https://x.com/Kimi_Moonshot/status/2084245860339298423)*

## The test that separates the two

There is a simple check for whether an agent has actually finished a task: **can the next person in the chain open the result and keep working?**

A summary in a chat window fails that test. So does a wall of markdown that someone has to paste into a document and reformat. So does a table rendered as ASCII art. In each case the agent has done the thinking and left the last mile — the part that costs a human twenty minutes — undone.

The gap is not intelligence. It is that producing a real file means handling a real format: styles, layout, embedded objects, the specific quirks of whatever tool the recipient uses. That work is unglamorous and it is where the actual time goes.

## Why this is a moat, briefly

File formats are a genuine barrier. Producing a valid, well-structured presentation or spreadsheet requires knowing a specification that is large, partly undocumented, and full of behaviour that only shows up in one particular reader.

That means output fidelity is a real differentiator between agent products for as long as it takes the ecosystem to commoditise it — which, judging by how long the equivalent problem took in publishing, will be a while.