P-Forge

P-Forge (formerly Course Gen) turns a topic or syllabus into a finished deck, at /p-forge. It produces either a narrated, animated course or a plain presentation, and saves both into your Obsidian vault.

What it does

Courses and presentations are siblings under one P-Forge root:

• ~/Obsidian/maven-os/P-Forge/Courses/<slug>/
• ~/Obsidian/maven-os/P-Forge/Presentation/<slug>/

Pick the mode with the Course Generator / Presentation Generator cards at the top of Generate (and in Settings):

• Course Generator - modules plus per-word-synced voiceover plus animation (the classic pipeline).
• Presentation Generator - a flat set of slides, a theme, an export format, and no voiceover (generation skips narration entirely; faster and key-free).

Flow

1. Generate tab - pick the mode, type a topic, or paste a full syllabus / day-by-day plan (unlimited modules and slides; “Auto-size” builds one module per listed item). Click Outline (free - Kimi drafts the module/slide structure for you to review), then Generate.
2. It plans and writes the course with your chosen model per step (Kimi by default), renders an infographic per slide, narrates every slide, embeds the audio so each slide auto-plays, adds native PowerPoint entrance animations synced to the narration, and writes speaker notes. Output: one self-contained .pptx plus outline.md, syllabus.md and course.json in the course folder.
3. Library / History - re-open or re-assemble past courses. Re-assembly is free - assets are cached.

In-app preview

Every deck in the Library has a ▸ PREVIEW button that opens a full-screen viewer: thumbnail rail, live slide rendering, Play narration (per-segment highlight), speaker notes, animation timeline and auto-advance. Content slides are drawn as native animated cards - real slides, not rendered pictures - so text stays sharp and the animation timeline is live.

From the viewer header you export to:

• .pptx
• .pdf
• .odp (via LibreOffice)
• SCORM 1.2 - a self-contained HTML5 package for any LMS.
• Standalone narrated HTML deck (narrated courses only) - the course packaged as the deck it already is, playable in any browser.

Google Slides is present but disabled until Google sign-in is wired.

Themes & templates

• Theme picker (Presentation mode) - built-in Navy&Gold / Midnight&Cyan / Graphite&Amber, plus any JSON theme you drop in ~/.maven/coursegen/themes/. One theme colours both the preview and the .pptx.
• Template (Settings) - takes a .pptx/.potx base (local path or URL) and an aspect ratio (16:9 / 16:10 / 4:3).

Deck-template library

Beyond the colour themes, P-Forge ships a library of fully-designed deck templates - Executive Dark, Neon Deck, Gold Noir, Nordic Sea, Data Report, Minimal Light, Paper Sketch and more. Each is a complete layout system (title, section, content, quote, stat and image slides) that any engine can fill. Pick one in Templates; images can be drawn onto slides by any image engine you have configured.

Settings

Shared live with the Generate tab:

• Per-step model pickers for the text pipeline - Step 1 Module planning, Step 2 Slide outline, Step 3 Slide content. Each offers the same concrete model lists as the AI Mastermind’s ⚙ Models panel (Claude Opus 4.8 / Sonnet 5 / Haiku 4.5 / Fable 5 · Kimi K2.x + Moonshot v1 · Antigravity Gemini/Claude/GPT-OSS · Codex GPT-5/o-series · every installed Ollama model). A single-model user leaves them on Kimi; a multi-model user mixes per step.
• Voice engine - ElevenLabs · Google · Local Kokoro · Local Supertonic - each with a Listen preview per voice.
• Infographic engine - matplotlib (default; fast, free, exact text) · Codex CLI (free, SVG) · Kimi · Claude · fal.ai · Google - to compare quality.
• Module / slide counts - 0 = auto.

Voice engines in detail

• Supertonic - the one to reach for outside English: on-device, free, 31 languages including Russian and Ukrainian (not Norwegian - the model has no no/nb). Its ten voices are styles (M1-M5 male / F1-F5 female), not languages: the same style speaks whatever the course language is, and the Listen preview is spoken in that language. About 0.6× realtime on a CPU; the weights (~1.9 GB) self-download on the first synthesis.
• Kokoro - free and offline, but English only (US & UK voices: Heart/Bella/Michael/Adam … Emma/Isabella/George/Lewis).
• ElevenLabs - the only engine with exact per-word timing, which is what animation sync uses. No ElevenLabs model is free - all bill the same 10k-credit/month quota; Flash/Turbo are ½ credit/char (2× the narration per quota; default Flash v2.5).

PulsAI

A second, standalone presentation studio lives inside P-Forge at /p-forge/pulsai - a Plus-AI-style deck builder with charts, a narrator, branding, a deck library and resources. It survives a dropped connection, waits for an upload to be ingested before using it, exports to .pptx and imports existing decks. It carries its own key and saves under <vault>/PulsAI/.

Setup

Requires Python 3 (an auto-created venv adds python-pptx + matplotlib), ffmpeg, and rsvg-convert (librsvg, for SVG infographics → PNG). You also need a voice key - ELEVENLABS_API_KEY or GEMINI_API_KEY (aliased GOOGLE_AI_STUDIO_API_KEY) - unless you rely on the free local Kokoro/Supertonic engines, which are provisioned by scripts/setup-free-engines.sh (no key, no GPU, no sudo). KIMI_API_KEY is needed only for steps left on Kimi; other step models bring their own auth (CLI logins / Ollama). Optional: OPENAI_API_KEY or the claude CLI for AI infographics. LibreOffice (soffice) is required for the .odp / .pdf export buttons; without it .pptx and SCORM still work. PulsAI’s export libraries (pptxgenjs, adm-zip) ship in node_modules.

Tips

• Draft the structure for free with Outline (Kimi) and edit it before spending on Generate.
• For non-English narration, choose Supertonic; for animation that syncs to the word, only ElevenLabs provides exact timings.
• matplotlib infographics are free and render exact text - a safe default before trying the paid image engines.

See also

• N-Forge - its Studio also builds Presentations (Markdown → PPTX/PDF).
• Local-first