Local-first AI research stack

The working idea

The deck is not a separate phase. It is another view over the research.

Most AI presentation workflows ask you to leave the research environment, prompt a deck generator, then clean up the result in an isolated app. That can be useful for a first draft, but it is a weak place to do serious iteration.

This setup keeps the source material, source links, outlines, speaker notes, visual ideas, and deck source in the same archive. The AI agent can read it. The human can browse it. The deck can be rebuilt without losing the trail of why each slide exists.

PPTX may still be the delivery artifact. It does not have to be where the thinking lives.

Infrastructure map

One archive, multiple interfaces

Sync layer

Nextcloud archive

Markdown files, images, source extracts, project folders, daily notes, and reusable context files.

Human UI

Obsidian

Readable notes, backlinks, search, quick editing, and enough structure without turning the archive into a database.

Agent bench

CLI agents

Claude Code, Codex, Gemini, and similar tools can search, edit, summarize, compare, and assemble drafts from local files.

Working output

Research pack

Source notes become synthesis, decision logs, slide outlines, critique, speaker notes, and reusable fragments.

Presentation layer

Slidev

Markdown slides, local theme, assets, browser preview, static build, and optional PDF or PPTX export.

Iteration loop: research -> outline -> deck source -> browser review -> agent edits -> human judgment

System layers

What each layer is responsible for

Stores the thinking

The archive is plain files: notes, links, drafts, exported snippets, diagrams, and source images. It survives tool churn because it is not locked inside one AI product.

Obsidian

Keeps it readable

Obsidian is the human interface, not the system of record. It makes the archive navigable without forcing every artifact into a publishing format too early.

CLI agents

Do the heavy lifting

Agents can inspect the same folders the human uses, update notes, compare versions, draft sections, extract arguments, and keep context close to the work.

Research pack

Turns files into judgment

The useful artifact is not a pile of links. It is a structured pack: what was found, what it means, what remains uncertain, and what deserves a slide.

Slidev

Compiles the story

Slidev keeps the deck source as Markdown. Layouts, theme, visuals, and speaker notes can evolve together with the research instead of being rebuilt in a binary editor.

Review

Closes the loop

The browser preview, screenshots, layout checks, and exports make the deck testable. The result can still end up as a PDF or PPTX when distribution requires it.

Why not just prompt a deck generator?

The point is not avoiding presentation tools. It is avoiding a broken handoff.

AI deck generator first

xResearch and deck generation happen in a separate tool session, often with weak memory of the source material.
xThe output is usually a binary or semi-opaque artifact that agents cannot inspect and revise cleanly.
xIteration becomes prompt, export, fix, repeat. The reasoning trail fades while the slide polish gets louder.
xThe deck looks finished before the argument is finished, which is exactly backwards for research-heavy content.

Archive and Slidev first

+Research notes, source links, outline, visuals, and deck source live in the same project folder.
+Agents work on plain text and assets they can search, diff, summarize, and rewrite with much less ceremony.
+Obsidian remains the review surface for thinking, while Slidev becomes the review surface for presentation.
+The story can move from memo to talk to article without throwing away the intermediate work.

Operating pattern

A repeatable four-step loop

Collect

Save sources, transcripts, screenshots, rough thoughts, and prompts into the project folder as Markdown and assets.

Synthesize

Use agents to compress the raw material into claims, tensions, evidence, gaps, and reusable narrative blocks.

Author

Turn the outline into Slidev Markdown, with speaker notes and visuals still close to the research files that created them.

Review

Preview in the browser, check layout, adjust the argument, then export only when the content has stopped moving.

Folder blueprint

Make the file system carry the workflow

The important move is boring: keep related materials together. A project folder should be readable by a human in Obsidian and workable by an agent in the terminal.

Context before commands Put goals, audience, constraints, and working style into notes that agents can read before producing output.

Plain files over tool memory Do not rely on one chat session remembering the project. Persist the useful state as Markdown.

Assets next to source Images, screenshots, exports, and deck files should be close enough that references stay obvious.

Presentation as code Decks become reviewable source files instead of opaque final artifacts.

project-topic/
  context.md              # goal, audience, constraints
  source-notes.md         # links, quotes, extracts
  research-pack.md       # synthesis and uncertainty
  outline.md             # story structure
  images/                  # screenshots and visual assets
  slidev/
    slides.md             # presentation source
    theme/                # reusable layouts
    public/               # deck assets
    screens/              # review screenshots
    dist/                 # static build output

The useful part

This is less about Slidev and more about keeping the content supply chain intact.

The same infrastructure can produce notes, articles, newsletter sections, advisory analysis, demos, or presentations. The advantage is not that Markdown is fashionable. It is that the research remains inspectable, portable, and reusable while AI agents are helping to shape it.

Back to vibes