Skip to content

Open source CapCut alternative.

TypeScript

fix(site): spell out slide direction and version asks in the tool catalog [rebuild]

David Le229fd4143d7c
diff --git a/site/scripts/lib/cut-eval/cases.ts b/site/scripts/lib/cut-eval/cases.ts
index cbbd58c5..3ddbdef8 100644
--- a/site/scripts/lib/cut-eval/cases.ts
+++ b/site/scripts/lib/cut-eval/cases.ts
@@ -204,6 +204,8 @@ export function cases(audio: { dataBase64: string; mimeType: string }): EvalCase
throw new Error("delete_item — iterating is not a delete/replace; the existing clip must stay");
if (name === "delete_asset")
throw new Error("delete_asset — iterating must not remove the existing clip's media");
+ // A reversible grade toward the asked-for look may ride along.
+ if (name === "set_color_grade") return { ok: true };
return undefined;
},
stubs: {
diff --git a/site/src/cut/components/GeneratePanel.tools.ts b/site/src/cut/components/GeneratePanel.tools.ts
index 64f35d16..96d1a1d7 100644
--- a/site/src/cut/components/GeneratePanel.tools.ts
+++ b/site/src/cut/components/GeneratePanel.tools.ts
@@ -11,7 +11,7 @@ export const VIDEO_GEN_TOOLS = [
{
name: "generate_video",
description:
- "Generate an AI video clip from a text prompt (Donkey's hosted video model: one pass, audio included, 720p, up to ~10s — the model picks the length) — the default for any video ask: \"generate/make a video of/about X\" means ONE clip from this tool, even though it sounds whole. Reach for generate_scene only when they name a narrated multi-shot production (a story, episode, or narrated short). When the user asks you to write or improve the prompt itself, put it in chat and wait for them to ask for the video. reference_asset_id anchors the render on one project asset: for an image or video, by default the tool first designs the opening frame from it with the image model (that still previews in this chat), then animates that exact frame — so tell the user it designs the frame, then renders; for audio, what it carries (speech transcribed, sound described) folds into the prompt instead. When the user wants the referenced image itself brought to life (\"animate this image/picture\"), pass animate_reference:true — the image becomes the literal first frame, unchanged, and the prompt describes only the motion. A safety-blocked or rejected anchor degrades the render on its own (identity reference, then text-only — the same ladder scene shots walk), so never resubmit a failed render with weaker inputs yourself. The render takes a minute or two and the tool RETURNS as soon as it's started — the clip previews as a live card in this chat when it finishes. Pass add_to_timeline:true (or an index) only when they asked for it in the cut; otherwise the user drags it in from the card. Needs the user signed in to Donkey (spends their credits).",
+ "Generate an AI video clip from a text prompt (Donkey's hosted video model: one pass, audio included, 720p, up to ~10s — the model picks the length) — the default for any video ask: \"generate/make a video of/about X\" means ONE clip from this tool, even though it sounds whole. \"Make me a better version / another take\" of an existing clip is also this tool — render the fresh take and leave their clip in place; a grade or effect on the original delivers a tweak, and they asked for a version. Reach for generate_scene only when they name a narrated multi-shot production (a story, episode, or narrated short). When the user asks you to write or improve the prompt itself, put it in chat and wait for them to ask for the video. reference_asset_id anchors the render on one project asset: for an image or video, by default the tool first designs the opening frame from it with the image model (that still previews in this chat), then animates that exact frame — so tell the user it designs the frame, then renders; for audio, what it carries (speech transcribed, sound described) folds into the prompt instead. When the user wants the referenced image itself brought to life (\"animate this image/picture\"), pass animate_reference:true — the image becomes the literal first frame, unchanged, and the prompt describes only the motion. A safety-blocked or rejected anchor degrades the render on its own (identity reference, then text-only — the same ladder scene shots walk), so never resubmit a failed render with weaker inputs yourself. The render takes a minute or two and the tool RETURNS as soon as it's started — the clip previews as a live card in this chat when it finishes. Pass add_to_timeline:true (or an index) only when they asked for it in the cut; otherwise the user drags it in from the card. Needs the user signed in to Donkey (spends their credits).",
inputSchema: obj({
prompt: str("The shot to generate — describe motion, subject, and mood"),
aspect: { type: "string", enum: ["16:9", "9:16"], description: "Clip shape (default: the supported shape closest to the project aspect)" },
diff --git a/site/src/cut/components/TransitionsPanel.tools.ts b/site/src/cut/components/TransitionsPanel.tools.ts
index 08d6a4d2..93aa7599 100644
--- a/site/src/cut/components/TransitionsPanel.tools.ts
+++ b/site/src/cut/components/TransitionsPanel.tools.ts
@@ -35,7 +35,8 @@ export const TRANSITIONS_TOOLS = [
style: {
type: "string",
enum: [...ANIM_STYLE_IDS, "none"],
- description: "Animation style, or 'none' to clear",
+ description:
+ "Animation style, or 'none' to clear. Slide names give the motion direction: entering from the left edge is slideright, from the right slideleft, from the top slidedown, from the bottom slideup.",
},
seconds: num("Ramp length in seconds, 0.1–2 (default 0.5)"),
}, ["clipId", "which", "style"]),