my-slidev
Deck work is image-heavy — exporting slides and reading the PNGs back — which would flood
this conversation's context. So the export-and-verify loop runs in my-slidev-agent, in its
own context. You stay here for the one upstream job the agent shouldn't do: shape the talk
into a slide breakdown. Do that, confirm the deck project exists, then spawn the agent.
Shape the talk
Skip this only when the caller already handed you a slide-by-slide outline.
A deck is a narrative, not a dump of facts. Before any markdown exists, decide the arc:
- Name the through-line. What is this talk for — what does the audience believe or be able to do by the last slide that they couldn't at the first? That goal orders everything.
- Break it into slides. One idea per slide. For each slide, write a single line: what it proves or moves forward (the claim, the demo, the diagram, the turn). A slide that doesn't advance the through-line is a cut, not a slide.
- Mark what each slide is grounded in. If a slide presents repo code, name the
file:line; if it states a fact from the prose spec, point at it. The agent pins each technical claim to this — it must not invent code or numbers. - Confirm the deck project and tooling exist. The agent exports through Slidev, so a deck
project (or at least a
slides.mdpath to write) must be in place, and export needs project-scoped playwright-chromium (npm i -D playwright-chromium). If neither the project nor the tooling is set up and you can't tell whether to scaffold it, ask before spawning.
Fixing an existing deck? Read its current slides.md first and keep its arc unless the
source clearly calls for a restructure — if you do change the structure, say so and why before
spawning.
Spawn the agent
my-slidev-agent owns the export mechanics, Slidev authoring rules, and the overflow
detect-and-fix loop — don't re-derive them. Resolve its inputs (ask only for what's genuinely
missing; never spawn without a DELIVERABLE path — it has nowhere to write otherwise), then
spawn foreground:
Agent({
subagent_type: "my-slidev-agent",
description: "deck <subject>",
prompt: `
TARGET: <new deck | fix existing slides.md at path>
OUTLINE: <the slide-by-slide breakdown you shaped above>
GROUND TRUTH: <repo code at paths, or the prose spec the talk presents>
DELIVERABLE: <path/to/slides.md>
`
})
Then relay its result — the deck path, the truth mapping, the fixes it applied, and any open questions. If it returns an open truth gap, get the user's answer and re-spawn with it.
If my-slidev-agent isn't installed, say so — don't run the loop inline (its export images are
exactly what this skill exists to keep out of the caller's context).