Develop your presentation with an AI Agent. Explore designs beyond fixed templates using image generation, then build the chosen designs as editable PowerPoint slides.
An editorial essay with editable text and photographic collage.
Agentic slide generation needs both freedom of layout and control over the final deck.
Fixed templates keep layouts consistent, but restrict how content can be presented. Image generation allows freer design, yet leaves the work of maintaining consistency and turning those designs into accurate, editable slides.
SlidePoise gives that creative process a structured workflow, supported by tools for visual measurement and PowerPoint construction. A shared content plan, style and asset library guide the Agent as it develops designs, builds editable slides and reviews the resulting deck.
How SlidePoise works
The Agent plans what the presentation should communicate and how the slides fit together. An image model explores layouts using that content, your style references and selected assets. To turn a chosen design into PowerPoint, the Agent identifies its objects and their relationships while local tools measure their geometry. The Agent reviews the rendered slides against the plan and corrects problems in content and appearance.
System architecture
Loading the design example…
Plan the content and explore the design
The Agent plans each slide’s message and supporting content, then sends that plan to the image model with style guidance and reference images. The model can explore how to arrange the content without following a fixed layout. Open the sample plan to see the content and instructions used to generate the design.
Interpret the objects and measure their geometry
Once a design is chosen, its contents need to become PowerPoint objects. The Agent identifies what each element represents and which elements belong together. In this example, the bars, labels and values form one chart. OpenCV, a computer-vision library, measures its geometry so the renderer can rebuild it as an editable chart linked to its data.
During reconstruction, the Agent also corrects problems in the generated image, such as arrows that connect the wrong objects or overlapping text. It checks the rebuilt slides against the brief and reviews alignment, typography and colour across the deck.
Agent interpretation and OpenCV measurements are combined to build the editable PowerPoint chart.
Define the visual language
Asset libraries give the Agent a consistent set of icons, logos and PowerPoint components to draw on. It selects relevant files for each slide and includes them as references for image generation. During reconstruction, it places the original files in PowerPoint. Choose an example to see the selected icons, then inspect an icon to find it on the slide.
Remix IconOpen-source icons
The Agent searches for relevant icons and chooses outline or filled versions to suit the slide.
Small illustrations can lose detail when generated as part of a full slide. The Agent can regenerate an illustration at higher resolution with a transparent background, then check it against the original design before placing it in PowerPoint. Drag the divider to compare the initial illustration with the regenerated version.
Initial generation412 × 383 pixels
Regenerated as a separate layer1301 × 1209 pixels · Transparent PNG
Initial generationTransparent artwork
Working with SlidePoise
Tell your Agent who the presentation is for, what it should communicate and which material to use. You can review the plan and designs together or ask it to work independently. Expand a stage below to inspect the plans, designs and review records from the selected sample.
Sample presentations
Browse the sample decks and open their plans to see what each presentation was intended to communicate. Compare the AI-generated designs with the rendered PowerPoint slides, or select an object in the PowerPoint view to inspect its text, chart data or shape.
Loading the presentations…
Read-only · Select an object
This slide image could not be loaded. Try another slide or reload the page.
Editable PowerPointDrag the divider or use arrow keysAI-generated design
AI-generated content. Headers and footers are added in PowerPoint.
Make your style reusable
Develop a style with your Agent and save it as a Profile for future presentations. Use the local Console to adjust fonts, colours and visual references, and manage your asset libraries. Try it here with sample settings. Your own Profiles and libraries are saved on your computer.
Style
Set fonts, colours and references, and decide how closely the Agent should follow them.
Resources
Manage the icon libraries, logos and components available to the Agent.
System
Choose how images are generated and check local tools.
Open the local Console
After installation, ask your Agent to open the SlidePoise Console, or run this command in a terminal.
npx github:henryhyw/slidepoise console
Saved changes apply to future presentations. To change a presentation already in progress, ask your Agent.
Install SlidePoise
SlidePoise is built and tested with Codex and its native image tools. Claude Code and Qoder can use the same workflow with a connected image generator. You can also generate images in another app. The Agent prepares the prompt and reference files, then continues when you return the image.
Your environment
Agent platform
Can run local commands and inspect images
Python
3.10 or later
Node.js
18 or later, with npm 9+
Your Agent can check these and help install missing tools.
Ask your Agent
Install and set up SlidePoise for me from https://github.com/henryhyw/slidepoise.
Or use your terminal
npx github:henryhyw/slidepoise setup
What gets installed
Setup installs measurement and PowerPoint tools, registers the skill for detected Agent platforms, and prepares local slide previews. Your system may request administrator access for the preview tools.
An isolated Python environment in ~/.slidepoise/python with Pillow, python-pptx, NumPy and OpenCV.
PptxGenJS and JSZip in ~/.slidepoise/node.
Profiles and Library Sets in ~/.slidepoise, with the skill registered in the detected platform’s skills directory. Add --agent claude, --agent qoder or --agent codex to choose explicitly.
LibreOffice renders the PowerPoint and Poppler converts it into images for the Agent to review. Setup installs missing copies through Homebrew on macOS, winget on Windows, or apt on Debian/Ubuntu. Use an image generator connected to your Agent, or generate images in another app and return them to the Agent. Set this preference with your Agent or in the Console’s System settings.