Hermes Slide-Deck Capability Report

Prepared for: Jonas Krauss · Date: 2026-08-28 Inputs: Internal capability audit [A], community practice digest [C], tool landscape survey [S]

1. Executive Summary

Hermes already has a working, verified markdown-to-slide-deck pipeline on this host: Quarto + reveal.js produces interactive HTML decks, and a headless Chrome binary (from the Playwright cache) exports them to multi-page PDF — no extra installs required [A §1.1–1.2]. The broader landscape confirms that markdown-first toolchains (Marp, Slidev, Quarto) are the dominant and fastest-growing approach among AI-agent users, while SaaS generators like Gamma offer the quickest polished first draft but suffer from unreliable PPTX export [C Muster 1–4, S §1]. The strongest recommendation is a two-track strategy: (1) make Quarto reveal.js the default deck engine (already installed, already verified), and (2) add a Gamma API integration for cases where Jonas wants a polished AI-generated first draft with minimal effort. For editable native .pptx deliverables, python-pptx wrapped by the knorq-ai MCP server (with theme presets) is the only viable route, but it requires installation and ongoing maintenance [S §6]. The biggest gap is not tooling — it’s the absence of a dedicated Hermes skill that codifies the deck-building workflow; closing that gap is a 2–4 hour effort with high leverage.

2. Current State — Internal Capability Inventory

Everything below was verified by direct execution on this host [A].

What works out of the box

Existing report-publishing precedents (proven on this host)

• Whitepaper repo: Makefile + Quarto (D2 diagrams + Typst CeTZ charts + typst-pdf) → dist/*.pdf [A §1.6]
• Westbridge repo: Architecture proposals, roadmaps, analysis PDFs — all produced via Chrome print-to-pdf (Skia/PDF m148) [A §1.6]
• Proposal generator: @react-pdf/renderer + React + tsx → programmatic branded PDFs [A §1.6]

Gaps

3. What Others Do — Community Patterns

Research identified six real-world patterns for building slide decks with AI agents, sourced from 34 references with 12 verbatim-evidence citations [C].

Pattern 1 — Claude Code + Marp (dominant community approach)

The agent interviews the user (goal, audience, key points), generates a .md with Marp frontmatter, and exports via marp-cli to HTML, PDF, and PPTX. The presentation is a textfile — versionable, diffable, AI-editable. A 15-slide deck converts in ~2 seconds. Community skills like robonuggets/marp-slides (299★) bundle 22 curated example decks + SVG charts so the agent learns composition, not just building blocks.

• Strength: Fastest text-to-deck loop; huge community ecosystem [C Muster 1]
• Weakness: PPTX export renders slides as images (not editable); the --pptx-editable flag uses LibreOffice but produces narrow textboxes needing python-pptx repair [C Muster 1]
• Source: freecodecamp.org, github.com/robonuggets/marp-slides

Pattern 2 — Source-first PPTX (outline.json → script → .pptx)

“Treat a deck like code”: an outline.json is the source, a renderer (pptxgenjs or python-pptx) builds the .pptx. Design is coded as constraints (variant grammar, preset system for palette/typography/density, QA loop). Skills like siril9/presentation-skill and mpuig/agent-slides (160+ tests) implement this.

• Strength: Editable native Office objects, reproducible, QA-checkable [C Muster 2]
• Weakness: Complex setup; overkill for one-shot decks; small community (10–45★) [C Muster 2]
• Source: github.com/siril9/presentation-skill, github.com/mpuig/agent-slides

Pattern 3 — HTML / Reveal.js presentations

The agent builds a self-contained HTML file with Reveal.js, Slidev, or custom templates. “Master slides” are components; corporate identity is defined via rules + assets in AGENTS.md. Community skills include yugasun/slides-skills (32 templates, 4 themes) and broeker/reveal-deck.

• Strength: No PowerPoint dependency, animated transitions, interactive charts, “improv”-capable [C Muster 3]
• Weakness: No native Office format; PPTX export is a problem for stakeholders who need .pptx [C Muster 3]
• Source: news.ycombinator.com, github.com/godagoo/reveal-presentations

Pattern 4 — SaaS API automation (Gamma)

A POST to the Gamma API (/v1.0/generations) with text input → polished, brandable deck in ~30–60s, exportable as PDF or PPTX. OAuth for MCP server integration. Fastest first draft available.

• Strength: Fastest polished first draft, 60+ languages, hosted, exportable [C Muster 4, S §2]
• Weakness: PowerPoint export is the universal complaint (fonts shift, animations lost, overlaps); 38% of Gamma threads and 31% of Tome threads cite this [C Muster 4]
• Source: developers.gamma.app, slidegmm.ai

Pattern 5 — Hermes’ bundled python-pptx skill

Hermes ships a bundled powerpoint skill with 5 helper scripts: create deck from JSON spec, read structured content, edit in-place, build from brand template, render slides to PNGs (via LibreOffice + pdftoppm). Optional finance/pptx-author for pitch decks where every number must trace to a source workbook.

• Strength: Editable native PPTX, template-driven branding, render-verify loop with vision_analyze, JSON-spec-based (agent-friendly) [C Muster 5]
• Weakness: Documented as a skill but no slide-design system (unlike robonuggets/marp-slides); users must turn to community skills for polished looks [C Muster 5]
• Source: Hermes powerpoint skill docs, Hermes pptx-author docs

Pattern 6 — Quarto RevealJS (academic/technical)

A .qmd source renders to RevealJS HTML or PPTX. Agent skills like alfredo-hs/quarto-talks encode editorial rules (intellectual spine first, overload → delete not shrink, ~23-slide limit, no invented quotes).

• Strength: Excellent for code/math/citations, Git-friendly, agent can read-edit-render-inspect-revise without GUI [C Muster 6]
• Weakness: Visual quality depends on editorial judgement; no automatic overflow detection [C Muster 6]
• Source: github.com/alfredo-hs/quarto-talks, alfredohs.com

Recurring community frustrations

1. PowerPoint export breaks — fonts substituted, animations gone, layouts overlapping. The #1 complaint about SaaS AI decks [C]
2. “Looks like AI” / same-look — after ~5 decks, clients recognize the pattern. Counter: concrete structure in prompt, name visual references, write content yourself and delegate only layout [C]
3. Hallucinations — invented statistics, fake quotes, fabricated DOIs. Affects all generators, especially dangerous in academic/regulated contexts [C]
4. Manual clipboard loop — agent writes 80–100 slides of content, but assembly in Google Slides remains manual [C]
5. Design competence — “most agent-built slide decks pass automated checks and still look bad.” Skills that encode design as constraints are the emerging solution [C]

4. Landscape — Tool Comparison & Top-3 Shortlist

Comparison table (16 tools surveyed, 49 sources [S])

Sources: [S §5] with citations from developers.gamma.app, sli.dev, github.com/marp-team/marp-cli, quarto.org, python-pptx.readthedocs.io, github.com/knorq-ai/pptx-mcp-server, github.com/smartdatabrokers/slideforge-mcp

Top-3 shortlist

1. Gamma (REST API + hosted MCP) — Best for AI-native stunning decks with minimal maintenance. The only option that is simultaneously top-of-class for AI visual design, has a GA REST API (Nov 2025) and hosted MCP server, and requires no local toolchain. Credit-based pricing on a Pro plan (~$18/mo) is predictable. Caveat: output is web-native; PPTX/PDF export can lose fidelity — acceptable if the deliverable is a link or PDF, not an editable .pptx. S §6

2. Slidev / Marp / Quarto — Best for deterministic, local, license-clean decks. Markdown-in + CLI export to PDF/PPTX/PNG in CI is the purest autonomous pipeline, fully MIT/free. Quarto is already installed on this host and verified. Choose Marp for the lightest dependency footprint; choose Slidev for the highest visual ceiling via themes/components. Note: PPTX export renders slides as images (not editable) in all three. S §6, github.com/marp-team/marp-cli, quarto.org

3. python-pptx (wrapped by knorq-ai or SlideForge MCP) — Best when the deliverable MUST be native editable .pptx. The only route with true PowerPoint-native output. Raw python-pptx is “not pretty by default” — wrap it with an MCP server that adds theme presets (McKinsey/Deloitte/Neutral/IR) and layout validation. Requires installation + LibreOffice for rendering. S §6, python-pptx docs

Why not the others: Beautiful.ai has the best visual floor but its API is early-access-only — not safe for unattended autonomous generation S §2. Tome is defunct — a cautionary tale about AI-SaaS availability risk S §2. Beamer requires a full TeX install (absent on this host) and is weaker for marketing-grade output [A §2.1, S §3].

5. Recommendation

Two-track strategy: local-first default + Gamma for rapid drafts

Track 1 (default): Quarto reveal.js — already working, zero install needed

This is the path of least resistance. The pipeline is verified end-to-end on this host:

1. Author content as a .qmd file (Markdown + optional mermaid code blocks)
2. quarto render deck.qmd → deck.html (interactive reveal.js deck)
3. Headless Chrome print-to-pdf → multi-page PDF
4. Diagrams via inline mermaid (bundled, zero-config) + D2 for architecture SVGs
5. Visual assets via image_generate (FAL) for cover/illustrations + PIL for composition
6. Branding via the documented palette (Dodger Blue, Indigo, Crimson Carrot) + Raleway
7. Publish via python3 -m http.server or the existing whitepaper/westbridge deploy pattern

This covers 90% of deck requests: interactive HTML for presenting, PDF for sharing. No new dependencies, no API costs, no external service risk.

Track 2 (optional): Gamma API for rapid polished first drafts

When Jonas wants a visually stunning first draft in 60 seconds without touching markup:

• Sign up for Gamma Pro (~$18/mo, 4,000 credits)
• Add the API key to 1Password (Jonas’s vault convention)
• Use the REST API (POST /v1.0/generations) or hosted MCP server for one-call deck generation
• Export as PDF for sharing (avoid PPTX export — known fidelity issues)
• Use as a starting point, then refine in Quarto if deeper control is needed

When native editable .pptx is required:

• Install python-pptx via uv pip install python-pptx (quick, one dependency)
• Activate Hermes’ bundled powerpoint skill (already shipped, 5 helper scripts)
• For themed output, consider the knorq-ai MCP server (McKinsey/Deloitte/Neutral presets + layout validation)

What to build: a dedicated slide-deck skill (Gap G8)

The single highest-leverage action is authoring a Hermes skill that codifies the deck-building workflow. This closes the gap between “the tools exist” and “the agent knows the opinionated best path.” The skill should encode:

• The Quarto .qmd template with brand palette + Raleway pre-configured
• The Chrome print-to-pdf command (with the exact binary path from Playwright cache)
• Diagram patterns (mermaid inline, D2 for architecture)
• The image_generate → PIL composition workflow for cover slides
• Editorial rules adapted from quarto-talks (intellectual spine first, ~23-slide limit, no invented quotes)
• A Gamma API fallback path for rapid drafts
• The python-pptx path for editable PPTX when required

Effort estimate:

6. Next Steps

1. Author the slide-deck skill — template .qmd with brand styling, Chrome print-to-pdf helper, editorial rules. This is the #1 priority; it makes the existing pipeline discoverable and repeatable. (2–4 h)

2. Install python-pptx — one command (uv pip install python-pptx), unlocks the bundled powerpoint skill for editable PPTX deliverables. (5 min)

3. Set up Gamma API as a fallback — register for Gamma Pro, store API key in 1Password, add a skill section for rapid AI-generated drafts. Useful when Jonas wants a polished starting point fast. (30 min)

4. Validate with a real deck — build a 10-slide deck on a real topic (e.g., the Westbridge POC summary) using the new skill, export to both HTML and PDF, and verify visual quality. This tests the full pipeline end-to-end. (1 h)

5. Consider matplotlib installation — only if data-chart-heavy decks become a regular need. D2 + mermaid cover most diagram cases today. (5 min, defer until needed)

Source Index

Internal capability audit [A]

• Deliverable: slide-deck-capabilities-inventory.md (task t_d05d3d3b)
• All findings verified by direct execution on host, 2026-08-28

Community practice digest [C]

• Deliverable: digest.md (task t_e94393f7)
• 34 sources, 12 with verbatim evidence, 51% citation coverage
• Key sources: freecodecamp.org, github.com/robonuggets/marp-slides, github.com/siril9/presentation-skill, github.com/mpuig/agent-slides, slidegmm.ai, Hermes powerpoint skill, github.com/alfredo-hs/quarto-talks

Tool landscape survey [S]

• Deliverable: slide-deck-tools-survey.md (task t_f24dcb63)
• 49 sources, citations verified
• Key sources: developers.gamma.app, gamma.app/pricing, sli.dev, github.com/marp-team/marp-cli, quarto.org, python-pptx.readthedocs.io, github.com/knorq-ai/pptx-mcp-server, docs.beautiful.ai, videosdk.live