The real product is a rulebook, not a template pack
A Claude Code skill that makes AI produce genuinely decent consulting slides — by feeding it rules every session and machine-checking the output.
What it does
This is a Claude Code skill for building boardroom-quality decks. Slides are authored as 16:9 HTML (one section per slide), exported to PDF via headless Chrome, and optionally converted to an editable PPTX. The kit bundles a slide-design rulebook, a 62-part HTML slide library, a catalog PDF, and scripts that generate a starter deck and mechanically check it for rule violations.
The interesting bit
The authors are blunt that the most valuable file isn’t a template or a script — it’s references/slide-rules.md, a plain text rulebook (“put the conclusion in the title”, “no rounded corners”, “one term per deck”). Since AI forgets verbal instructions between sessions, the trick is making the model re-read the rules every time, then catching violations with check_deck.py — including leftover placeholders and title/body numbering mismatches. A final fresh-eyes review is done by a different agent that doesn’t know how the deck was made.
Key highlights
- 62 slide archetypes (27 basic + 35 advanced) as standalone HTML parts, browsable via a 62-page catalog PDF
new_deck.pycherry-picks parts by ID and merges the two libraries’ CSS without collisionscheck_layout.mjsdoes real rendering checks: footer overlaps, edge overflow, pages more than 40% empty- HTML-first workflow: PPTX conversion only on explicit request, and edits always flow back through the HTML to avoid drift
- Includes an “AI smell” lexicon to scrub telltale AI phrasing, plus self-tests for the checker itself
Caveats
- The default palette and typography are tuned for Japanese consulting decks (Mincho headings, warm earth tones); other markets will want to swap the
:roottokens - PPTX conversion replaces fonts with Yu Gothic/Yu Mincho and rasterizes SVG charts, so numbers inside charts aren’t editable
- The last page carries a “made with consulting-pptx-skill” attribution line
Verdict
Worth a look for anyone whose LLM decks keep coming out as bullet-point soup — the rules-plus-machine-check loop is the transferable idea, even if you never touch the HTML parts. If you need Western-brand styling or fully editable native charts, expect to adapt it.
Frequently asked
- What is carnot-tech/consulting-pptx-skill?
- A Claude Code skill that makes AI produce genuinely decent consulting slides — by feeding it rules every session and machine-checking the output.
- Is consulting-pptx-skill open source?
- Yes — carnot-tech/consulting-pptx-skill is open source, released under the MIT license.
- What language is consulting-pptx-skill written in?
- carnot-tech/consulting-pptx-skill is primarily written in HTML.
- How popular is consulting-pptx-skill?
- carnot-tech/consulting-pptx-skill has 1.1k stars on GitHub.
- Where can I find consulting-pptx-skill?
- carnot-tech/consulting-pptx-skill is on GitHub at https://github.com/carnot-tech/consulting-pptx-skill.