dashi-ppt-skill: an agent skill that turns documents into browser-editable decks

The problem is the gap between generated slides and slides you can still change

Most document-to-deck tools optimise the first render and stop there. You get a file, you notice a wrong number on page seven, and the only way back is to regenerate the whole thing and hope the rest survives. The README of dashi-ppt-skill frames the project around that exact gap, stating that how you edit after generation matters more than generation itself. The skill is aimed at people who produce industry research, financing reviews, competitive analysis, trend reports, project updates, pitch materials and internal training decks, and who need the result to stay editable rather than frozen. The README is equally direct about who it is not for: scenarios that require pixel-by-pixel hand-crafted visuals. That boundary is honest, and it is the right place to start evaluating the tool.

Every generated page is an HTML editor, not a static slide

The architecture described in the README is a page made of a layout plus copy fields. The skill selects from 12 themes and, within each theme, a layout library the README counts at 1020 layout pages with 8576 adjustable controls across 20 page roles such as cover, table of contents, metrics, trends, comparison, process, risk and closing. Because the artefact is HTML, the agent can read it, modify it and validate it directly, which is the reason the project is packaged as an agent skill rather than a standalone application. The generated file is also the editor: the README describes a console attached to each page with more than twenty dimensions of adjustment, covering content, layout, module count, page emphasis, preset palettes and transition animation. Nine page transitions are listed. Text is edited in place by clicking it, and decorative elements resize as the character count changes. Media slots accept click or drag replacement, and the README notes that uploads are compressed automatically. Left-side thumbnails support drag reordering, and pages can be skipped, deleted or duplicated. A top bar switches to presentation mode, toggles light and dark, and resets all changes.

Install is one npx command, and the environment check is the real gate

Installation and update are the same command, and re-running it updates in place while keeping already-installed dependencies:

npx dashi-ppt-skill@latest

For networks in mainland China the README gives a mirror variant:

npx --registry=https://registry.npmmirror.com dashi-ppt-skill@latest

The README also provides a prompt you can hand to an agent instead of running the command yourself, asking it to install the skill from the same npx invocation. Requirements are stated as Node.js 20 or newer plus npm, and exporting PPTX or PDF requires Chrome, Chromium or Edge installed locally. That browser dependency is the constraint most likely to bite in a container or a locked-down CI machine, and it is worth confirming before you plan around the export features. Platform support is listed as tested rather than exclusive: Claude Code, Codex (which the README says can call image generation to supplement illustrations), Doubao with office mode enabled, and Marvis, Workbuddy, Dumate and Qclaw, which only need the skill file placed somewhere readable with SKILL.md present. Cursor and other local agents are described as usable if they can read and write files and execute shell commands. Plain web chatbots are explicitly not recommended because the generator needs a local Node.js environment.

The workflow assumes an agent that can run shell commands and write files

The README lays out five steps: describe the requirement (topic, audience, page count, the conclusion you want to land), pick one of 12 theme previews and confirm whether images or video are needed, let the skill assemble structured content and a deck plan, edit in the browser, then export. Step four is where the design differs from a generator that hands you a file. Sliders on the console change how many modules appear in a page and shift the logical emphasis, which the README ties to pacing during a talk. Layouts can be switched, chart types swapped, and palettes adjusted within a theme. The layout library includes radar, waterfall, treemap, funnel, heatmap, Sankey and Gantt charts, plus analysis-model layouts for SWOT, Porter's Five Forces, PEST, business model canvas and the double diamond. If your agent cannot execute shell commands, none of this is reachable, because the generator runs locally.

Export fidelity is claimed per node, and the README does not quantify the loss

Export produces an offline HTML bundle, a PDF, or an editable PPTX. The README describes the PPTX path as node-by-node reconstruction with text remaining editable, and ships a side-by-side GIF comparing the HTML version with the exported PPTX page by page. What it does not provide is a list of which HTML features survive the conversion. Entrance animations, interactive controls, light and dark mode switching and the nine transition animations are described as HTML-side capabilities, which implies the PPTX carries the static page rather than the interactivity. If your deliverable is a PPTX that a colleague will edit in PowerPoint, plan on verifying the conversion on a representative page yourself. The README's own framing supports that caution: the console and the in-browser editing are the product, and the export is the handoff format.

Where it is the wrong tool

The README's suitability list is unusually candid. Decks that need pixel-level manual customisation are marked unsuitable, and that is not modesty, it is a structural consequence of the model. You are choosing a layout from a fixed library and then tuning parameters, so a design that does not exist in the library cannot be produced by dragging a slider. The 1020 layouts and 20 page roles constrain the shape of what you can express. If your brand guidelines demand a specific grid, type scale and bespoke diagram, a design tool plus a human designer remains the correct path. There is also a maintenance question the material does not answer: the only release listed is readme-assets-v1, published 2026-07-07 and described in the repository as README demonstration assets for image hosting rather than a software version. That means no tagged software release is visible in the supplied material, so there is no changelog to read and no version pinning story to evaluate. The last push is dated 2026-09-08. Whether the project follows semantic versioning on npm is something you would need to check on the registry, not something this repository page establishes.

How it differs from Reveal.js and Marp

Reveal.js and Marp both treat the deck as source text that you author and then render. Reveal.js gives you an HTML framework with a JavaScript API, and Marp converts Markdown into slides through a fixed set of themes and directives. In both cases the authoring surface is the file, and the visual result is determined by the template you picked before writing. dashi-ppt-skill inverts the sequence. The agent drafts the content and the layout together, and the editing surface is the rendered page in a browser, with sliders and in-place text editing rather than a stylesheet or a directive. The trade-off is control for speed. With Reveal.js you can express any layout the browser can render, but you write it. With dashi-ppt-skill you get a structured deck in minutes and then adjust within the boundaries of the layout library. If your team already has a Marp theme that matches your brand and a Markdown pipeline, switching buys you the console and the PPTX export, and costs you the ability to define layouts from scratch. That is the actual comparison, and it depends on whether your bottleneck is authoring time or design freedom.

Licence and the cost of staying current

The repository is licensed AGPL-3.0, which is a strong copyleft licence with a network-use clause. The practical implication is that if you modify the skill and make it available to users over a network, the licence's source-disclosure obligations are generally understood to apply to your modified version. This is not legal advice, and the details depend on how you deploy it, so read the LICENSE file in the repository and get counsel if you plan to host it as a service or embed it in a product. On upgrade cost, the README states that install and update are the same npx command and that re-running it updates in place while keeping installed dependencies. That is a low-friction upgrade path, but it also means you track whatever is published at latest unless you pin a version yourself. Since no software release is tagged in the supplied material, there is no release channel to subscribe to, so pinning means pinning an npm version rather than a Git tag. Budget for reading the diff on the npm side before you let an unpinned npx command run in an environment that matters.

Editorial conclusion

Adopt dashi-ppt-skill if your workflow already runs through a file-reading agent such as Claude Code or Codex, you need structured decks fast, and you expect to keep editing after generation. Do not adopt it if you need pixel-level manual control, if your environment lacks local Node.js, or if AGPL-3.0 obligations conflict with how you plan to distribute the output. Before committing, verify one thing on your own machine: that Chrome, Chromium or Edge is installed and that npx dashi-ppt-skill@latest completes, because the PPTX and PDF paths depend on that browser being present.

Official sources

1. chuspeeism/dashi-ppt-skill on GitHub
2. Issues
3. License: AGPL-3.0
4. README
5. Releases