Summary

PPT Agent addresses LLM-generated presentation flaws like hallucination, overlap, and poor layout. It's a code-driven framework using a state machine and multi-agent collaboration to automatically produce professional PPTX files from simple text prompts, offering a robust solution for high-quality, automated presentation generation.

How It Works

A state machine drives independent sub-agents for Research, Outline, Style, and Planning, preventing context contamination. The "Yin-Yang" philosophy enforces foundational layout constraints while granting AI creative freedom. A pixel-level Visual QA loop audits HTML screenshots via LLM, resolving conflicts with DOM/CSS rewrites. Data/rendering layers are decoupled via validated JSON contracts, supporting stateless breakpoint recovery by analyzing existing artifacts. A high-performance Puppeteer engine with event hooks ensures rapid, accurate rasterization and SVG parsing for dual PNG (fidelity) and SVG (editability) PPTX exports.

Quick Start & Requirements

Operates as an "Agent Skill" within a compatible agent environment (npx skills add sunbigfly/ppt-agent-skills). Trigger generation via natural language prompts (e.g., "Generate a 15-page roadshow deck on 2026 embodied intelligence trends, dark tech style."). Outputs (web preview, dual PPTX) are saved to ppt-output/runs/<RUN_ID>/. No specific OS/hardware prerequisites beyond the agent environment are detailed.

Highlighted Details

• Sub-agent Stage Isolation: Independent agents for Research, Outline, Style, Planning with strict context separation.
• Pixel-Level Visual QA: LLM audits HTML screenshots; resolves layout issues via DOM/CSS rewrites.
• Stateless Breakpoint Recovery: Resumes from interruptions by scanning existing artifact files.
• "Yin-Yang" Design: Enforces physical constraints while granting AI typography/layout privileges.
• Optimized Rasterization: Puppeteer engine with event hooks for fast, accurate screenshots and SVG parsing.
• Dual PPTX Export: PNG (visual fidelity) and SVG (editability) formats.

Maintenance & Community

Recently updated (v4.1, April 9, 2026), indicating active development. Mentions friendly links with the LINUX DO Community. No specific contributor, sponsorship, community channel, or roadmap details are provided.

Licensing & Compatibility

Released under the permissive MIT license, allowing commercial use, modification, and distribution, including within closed-source applications.

Limitations & Caveats

Requires a specific "supporting agent environment" capable of running "Agent Skills," a potential adoption blocker. The complex multi-stage workflow may present integration challenges. Visual QA relies on LLM interpretation, introducing potential edge cases.