The demo deck is 17 slides about SlideOps, built by SlideOps: open the HTML or the PDF.

Writing documentation isn't the bottleneck any more. Keeping it true is. The deck that said we run two database migrations still says two, a year after we started running ten 😅.

SlideOps is a pair of Agent Skills for Claude Code and compatible coding agents. It treats a generated document the way you'd treat generated code: it's built from a source, and it records which source it came from.

The reasoning behind it is written up in Your documentation is a build artifact. Start treating it like one.

Press Esc in any deck for the overview grid.

Install

Claude Code. This repo is its own marketplace, so two lines are the whole setup, and it's the only install that keeps itself up to date:

/plugin marketplace add glukicov/slideops
/plugin install slideops@slideops

Codex, Copilot CLI, OpenCode, or a plain checkout:

git clone https://github.com/glukicov/slideops && cd slideops
./install.sh

All four agents read SKILL.md and nothing needs porting between them. For the symlink and snapshot installs, the per-agent table and how updates reach you, see docs/install.md.

[!NOTE] Third-party marketplaces have auto-update off by default. Turn it on once in /plugin → Marketplaces, or new versions only arrive when you run /plugin marketplace update slideops by hand.

Use

Open a repository and say:

[!TIP]

💬 make slides about this repo

SlideOps scans the repo first, then asks one compact set of questions: which topic (it proposes concrete candidates it found, each with a "why now"), audience, length, theme and extras. You get an outline to approve before it writes any HTML.

If you already know what you want, skip the intake:

[!NOTE]

💬 deep dive on the auth subsystem, Ledger Dark theme, 15 slides, with a PDF

Months later, in the same repository:

[!IMPORTANT]

💬 is the architecture deck still accurate?

The agent sweeps the deck folder and triages by status. It re-quotes whatever merely moved, and flags the slides whose claim might no longer hold. It repairs what drifted instead of regenerating the deck, so the pacing and narrative you signed off on the first time survive.

Features

• Freshness checking. scripts/check.py sweeps a whole docs/slides/ folder and reports which slides cite code that has changed, moved or vanished since the deck was built, then suggests the fix or hands an agent a JSON repair brief. No model, no network, no tokens: standard library Python, and it runs in milliseconds.

• One self-contained file per deck. No build step, no CDN, works offline, attaches to an email.

• Navigation: arrow keys, click-to-advance, URL hash deep links, an Esc-toggled overview grid, and speaker notes on N (never visible in screenshots or exports).

• 13 slide patterns: title, agenda, section divider, prose + cards, reference table, before/after code, annotated snippet (half and full width), flow diagram, lane comparison, image + caption, chat bubbles, closing.

• 4 themes, one block each. Every color derives from a single :root token block via color-mix(), so switching theme is one replacement: Ledger Light (default), Ledger Dark, Midnight, Graphite. Or point it at a brand's real CSS values and map those onto the token roles.

  Ledger Light (default)	Ledger Dark
  Same deck, same markup, same content.	One :root block apart.

• Mermaid diagrams pre-rendered to inline SVG at build time and themed from the deck's own tokens, so the deck stays dependency-free.

• Verified PDF export. The companion skill renders the finished PDF back to images and checks the pages, because a PDF can have the right page count and still hand you blank images.

Documentation

Inside the skill, skills/slideops/references/ holds the specifications the agent reads: freshness, automation, style, themes, diagrams and verification.

Credits

Prior art worth knowing: frontend-slides for visual-first theme selection, and presentation-skills for pioneering the render-then-look visual QA loop that SlideOps also relies on. What SlideOps adds is the Ops half: content grounded in a repository, and a cheap way to ask later whether it still holds.

Licence

MIT. See LICENSE. Use it, fork it, ship it commercially; attribution is the only condition. The decks you generate are your own content either way.