Skip to main content
The storyteller skill teaches the agent to produce a finished short-form video from a single creative brief. End-to-end in ~6 tool calls.

​
What it does

Given a user prompt like “a sleepy fox in a bookshop discovering an old map,” the agent:
  1. Probes capabilities (get_status), which providers are wired up?
  2. Checks the budget (get_ledger), bail if available < 20 credits.
  3. Creates a workflow with the shot.
  4. Drafts a short script in a warm storyteller register.
  5. Attaches a voiceover with auto-captions.
  6. Attaches a music bed matching the mood, ducking on.
  7. Sets the audio mix to mix with native volume 0.6.
  8. Renders and polls until done.
  9. Returns the video URL.

​
The skill body

Copy this verbatim into your skill loader, or npx -y @lavendly/skills install storyteller. 
---
name: lavendly-storyteller
description: Produce a narrated short-form video from a one-line idea.
trigger: When the user asks for a video, clip, short, reel, or visual story.
---

# Lavendly · Storyteller

You have access to the Lavendly MCP. Use it to produce a narrated
short-form video from the user's brief.

## Canonical sequence

1. `get_status` → confirm voice + music providers are available. If
   not, ask the user to set them up before continuing.
2. `get_ledger` → bail with a friendly message if `available < 20`.
3. `create_workflow` with a single shot. Duration defaults to 5 s
   unless the user specifies otherwise (cap at 12 s).
4. Draft a 1-2 sentence script in a warm storyteller register. Speak
   in present tense. Avoid clichés ("once upon a time," "in a world
   where").
5. `attach_track` voiceover with the script and `subtitleStyle: "tiktok"`.
   Set `idempotency_key: "vo-${workflow_id}"`.
6. `attach_track` music. Pick a `mood` that fits the shot, cozy
   acoustic for warmth, low strings for tension, ambient pad for awe.
   `ducking: true`, `volume: 0.4`.
7. `set_clip_native_audio` on the shot: `{mode: "mix", volume: 0.6}`.
8. `create_render` with `idempotency_key: "render-${workflow_id}"`.
9. Poll `get_render` every 4 s until status is `done` or `failed`.
10. Return `result.video_url` to the user.

## Decision rules

- Captions default ON unless the user says otherwise.
- For shots without dialogue, use `mode: "off"` for native audio so
  the music bed isn't fighting baked sound effects.
- If `get_status.voice.supports_inline_tags` is true, sprinkle one or
  two emotion tags in the script (`[whispers]`, `[wistful]`). Don't
  overdo it.

## When to bail

- Available credits below the cost preview.
- The user asked for >12 s, propose splitting into multiple shots and
  switch to the multi-clip skill.
- Render failed twice in a row with the same error code.

​
Try it

Once installed:
Make me a 5-second video of a barista pulling a perfect espresso shot at dawn. Narrate it like an old French film.
The agent uses the skill verbatim, checks providers, drafts a script in a specific register, picks a low-strings music bed, renders, returns the URL. ~30 seconds wall clock.

​
Why this works better than freeform prompting

Without the skill, an agent given the same brief will often:
  • Skip the capability probe and fail mid-render when voice isn’t wired.
  • Render before attaching audio because nothing told it the order.
  • Use a different idempotency key on retry, double-billing the user.
  • Pick a generic “cinematic” music mood for every brief.
The skill is short (under 400 words) but every line removes one of those failure modes.