PPT Master: AI Agent Workflow for Editable PowerPoint Decks

What PPT Master is, and what it is not

The README sums up the design as "harness + model = agent". PPT Master owns the workflow, and the model you plug in sets the ceiling. In the words of the project FAQ, it supplies presentation-specific reasoning, project state, deterministic conversion and quality gates, while the selected model still decides how good the result can be. A weak or small-context model will produce a weak deck.

Hugo He, a finance professional who regularly reviews and edits decks, built the tool because he wanted AI-generated slides to stay editable in PowerPoint rather than flattened into images. The output is meant to be a draft you keep working on. The technical design document states that the user reviews the direction and owns the final-mile judgment in PowerPoint.

Beyond new decks, the README says PPT Master can also distill reusable brand, style, layout or deck templates from your references, fill an existing .pptx with new content while preserving its design, and add native transitions, animations and narration to a finished deck.

How the PPT Master workflow turns a document into a deck

Everything happens in the agent's chat panel: point it at a file or paste text and ask for a deck. The default run goes through these stages:

1. Source conversion. The agent converts your source material to Markdown before it plans any slide. If you only give a topic, the Generate route runs a topic-research stage to gather the factual baseline.
2. Project setup. The agent creates an active project under the working folder's projects/ directory and reports its exact path.
3. Stage 1 confirmation. You confirm the communication contract (purpose, audience, reading situation, canvas or format) and choose free design or a template.
4. Strategy. A "Strategist" role settles on the core message and picks a narrative mode: pyramid, narrative, instructional, showcase or briefing. The Why PPT Master page explains that a pyramid deck leads with the conclusion, while a briefing deck lays out the facts evenly.
5. Stage 2 confirmation. You confirm page count, the visual system, template application and production choices.
6. Page authoring and checks. An "Executor" role writes the slides as SVG pages, one after another, and a quality checker blocks export while errors remain.
7. Export. Scripts compile the SVG into DrawingML and save a timestamped .pptx in the project's exports/ folder. A backup of the authored SVG lets you re-export the same deck without re-running the model.

While pages are generated, the getting started guide describes a browser preview that prefers http://localhost:5050. You can select an element to change its text, colour, font or size, or attach an annotation and ask the AI to rewrite that region.

Default mode versus quick mode

If you explicitly ask for quick generation, the agent skips the confirmation round trips. Whatever you state is followed and whatever you leave out the agent decides on its own. Quick mode keeps the same toolbox (source conversion, images, native shapes, charts) but it is one-pass: it writes no design spec, and if the agent loses its context before delivery you start again.

PPT Master requirements and installation

Every installation path needs Python 3.10 to 3.14 and an installed, authenticated agent host that can read and write the working folder and run shell commands. The README lists three ways to get the code:

• Git clone (recommended). Clone the repository, run pip install -r requirements.txt, and update later with the bundled update_repo.py script.
• ZIP download. Download the repository ZIP from GitHub, or the skill-only ppt-master-skill-*.zip from the releases page (about 56 MB, without the bundled example decks). The FAQ notes the full repository is over 1 GB because of Git history and example assets.
• Skill marketplace. Run npx skills add hugohe3/ppt-master, or use /plugin marketplace add and /plugin install inside Claude Code. These fetch only the skill files, so you still run pip install -r requirements.txt from the installed location.

The Windows installation guide promises a working setup in under 10 minutes and flags one common mistake: tick "Add python.exe to PATH" in the Python installer. To check an installation, the FAQ suggests a Python import test for pptx and fitz, then asking the agent for a three-slide "Hello World" deck in quick mode. If a .pptx appears in the project's exports/ folder and opens, the setup works.

Images are optional: the agent can use its host's image tool, a configured provider, or a keyless web search whose quality the README calls uneven; free Pexels or Pixabay keys improve stock photos.

What "editable" and "native objects" mean in practice

According to the Why PPT Master page, the native output covers:

• Native shapes Preset geometry with working adjustment handles (block arrows, chevrons, callouts, flowchart nodes), connectors and freeform paths.
• Slide masters and layouts Real p:sldMaster / p:sldLayout inheritance, but only on template or structured routes.
• Text A multiline block exports as one editable text frame. Sizes are authored in px on a 1280×720 canvas and converted to pt on export (pt = px × 0.75), so a 24px body becomes 18pt.
• Equations Block and inline formulas compile to editable OMML for PowerPoint 2010+.
• Transitions, animations and speaker notes Written as real OOXML, not embedded video.

Charts and tables: two export options

By default, charts and tables become custom-designed SVG graphics converted to native PowerPoint shapes. You can move, recolour and restyle them, but there is no data workbook behind them. If you need Excel-style data editing, export with --native-charts-and-tables: supported charts and text-grid tables then ship as data-backed PowerPoint Chart and Table objects in a separate file. The FAQ notes that native objects may look slightly different across PowerPoint, Keynote, LibreOffice and WPS, which is why shapes remain the default.

What PPT Master does not produce

The project lists deliberate omissions: SmartArt, native WordArt and warped text, some decorative effects such as reflection and soft edges, and embedded or legacy objects such as OLE, video and macros. Formula editing is not supported in Keynote, WPS, LibreOffice or other non-PowerPoint clients. Microsoft PowerPoint is the primary validation target; other apps may remap animation effects.

PPT Master templates and reusing existing decks

By default PPT Master uses free design, so you do not need a template. When a deck must reuse a brand or a recurring format, the templates guide defines four workspace kinds: Brand owns identity, Style owns communication method and visual defaults, Layout owns reusable page structure, and Deck owns a recurring presentation together with its identity and structure.

Two routes handle an existing .pptx:

• Edit Native PPTX Give the agent your deck plus new material and ask it to fill the deck. Unchanged pages are restored byte-for-byte and only planned pages are edited.
• Create Template, then Generate Build a validated workspace from the reference, then author a fresh deck whose structure and page count can differ from the source.

Without a template you can still pick from 18 built-in visual styles, from swiss-minimal and editorial to dark-tech, brutalist and ink-wash. The FAQ is careful on one point: the repository contains no measured or guaranteed link between template reuse and lower token usage.

Outputs: formats, animations and narration

Besides 16:9 and 4:3 slides, the FAQ lists canvas formats for Xiaohongshu (3:4), WeChat Moments or Instagram (1:1), vertical stories (9:16) and A4 print. The output is still a .pptx file with native shapes.

Exported decks get a fade page transition by default and no element animation. You can opt in to automatic entrances or one of 203 native effect presets. For voice-over, the agent can turn speaker notes into per-slide narration, by default with edge-tts (about 90 locales), embed the audio in the PPTX and let PowerPoint export a narrated MP4. Cloned voices from providers such as ElevenLabs, MiniMax and OpenAI are also supported.

The examples site collects sample decks, and the README notes that every example is a single pass with no manual polish.

Models, cost and data handling

The README recommends Kimi K3 or Claude with a large context window (around 1M tokens) plus an AI image model. Kimi is also a project sponsor, so note that the FAQ separately names Claude (Opus or Sonnet) as the recommended and most tested model, and the author develops and tests mainly in Claude Code. Smaller context windows force split-mode runs, which the FAQ says introduces visible style drift between phases.

Cost is your model usage only: no PPT subscription, credits or per-seat fee, though the README admits high token use is the trade-off for page-by-page design.

On data, source documents are converted, SVGs generated and the PPTX exported locally. Outside traffic is the conversation with your AI model plus the optional services you enable: image generation receives prompts, web image search receives keywords and online narration receives speaker-notes text. Before converting confidential documents, check the data terms of the model provider you connect. One licensing detail for redistributors: the optional PDF converter depends on PyMuPDF, which is AGPL-3.0 rather than MIT.

Limits, and who PPT Master fits

The project's own list of limitations:

It fits consultants, analysts, IT and finance teams who send decks onward for editing, need masters, native shapes and real charts, and want material to stay on their machine. The docs point to finance, government and organisations with data residency requirements. It fits less well if you want instant slides with zero setup, where the project itself recommends a hosted SaaS tool. Other options are listed in the office and presentation tools category.

To evaluate it, open one of the example .pptx files in PowerPoint and click into the shapes and charts, then run the three-slide quick test with the model you would actually use. The PPT Master tool profile keeps a short summary for comparison.