Your coding agent just landed the PR, updated the changelog, and maybe even opened the docs page. Then someone — usually you — opened CapCut to cut the launch reel by hand.
That mismatch is the whole problem. The product moved at agent speed. The video still moves at timeline speed.
SceneRok is built for the case where the same agents that write code should also write the launch video as source: VidScript or @scenerok/sdk, generative models as compile-time functions, a deterministic compiler for structure. This is a walkthrough of that path for people who already live in Claude, Cursor, Aider, OpenCode, or Codex — and why it beats both “one black-box generate_video MCP call” and another night of CapCut.
The gap after the merge
Shipping software with agents is getting boring in a good way. Shipping the announcement is not. You still need a fifteen-second cut that hits brand, timing, and a CTA on the right beat. CapCut can polish one take. It does not give you a versionable source the agent can patch when the offer changes at 11pm.
Black-box video MCPs feel closer to the agent loop until the first failure. Wrong aspect ratio. Logo too small. Brand colors off by a hair. The model gave you a clip; it did not give you a program you can fix. You regenerate and hope. That is the same lottery as one-shot campaign tools, just invoked from the terminal.
Four steps that look like software
1. Brief. Same as a ticket: audience, length, offer, must-show product shot, brand constraints. Keep it in the repo next to the feature if you can.
2. Agent authors source. For a short linear reel, VidScript stays readable — time ranges, text, VO, compile-time slots. For loops, SKU grids, or timing derived from probed media, the agent writes TypeScript with @scenerok/sdk. Both compile to the same intermediate timeline. Humans can still edit a .vid template without learning a framework; agents that need control flow stay in TS.
3. Compile. Generative plugins resolve as functions inside the script — image/video/TTS calls at compile time, not a side chat. Structure (duration, overlays, CTA placement, brand rules) stays with the compiler. Same source and same inputs should reproduce.
4. Preview, then hosted final. Preview is free; bring your own assets if you do not want model generation. You pay when you burn a hosted final render. That matches how builders already think about compute: iterate cheap, pay for the artifact that ships.
One real failure mode
An agent once emitted a clean-looking reel that was 16:9 for a Stories placement and skipped the brand lockup because the brief only said “make it punchy.” In a CapCut world you notice in export and re-edit. In a black-box MCP world you re-prompt. In a source world you fail validation or a template rule: wrong aspect for the target, missing lockup slot — then the agent patches the script and recompiles. The failure becomes a diff, not a vibe.
That is the contrast that matters. generate_video({ prompt }) returns media. VidScript / SDK returns a program. Programs get reviews, fixtures, and retries that do not erase everything that was already correct.
CapCut vs black-box MCP vs source
CapCut wins when you need one human polish pass. It loses when the same reel must come back next sprint with a new offer and the same brand shell.
A generate_video MCP wins on speed to first pixels. It loses when the failure is structural — aspect, lockup, CTA beat — because there is nothing to open except a new prompt.
Source wins when the agent is already your coworker. The brief lives in git. The script is reviewable. A bad compile is a failed check or a small patch, not a full creative restart. SceneRok’s generative-compile path is that middle: novelty in the model calls, brand and timing in the shell.
Soft next step
If your agents already ship code and you are still hand-cutting launch cuts, try the structure-first path: free account and preview on scenerok.com, BYO assets optional, pay only for hosted finals. Point Claude, Cursor, Aider, OpenCode, or Codex at the public skills/CLI and a brief in-repo. The goal is not a prettier one-shot — it is a launch reel you can fork when the feature ships again next week.
Top comments (0)