Works with the AI tools you already use

CCCGGVWM+14 more

Short-video pipeline: topic → script → finished cut

Name: Short-video pipeline: topic → script → finished cut
Price: 15 USD
Availability: InStock
Author: Agensi

by Henry Zou

Four skills that take a news beat from "what should I make today" to a rendered, captioned, publishable video — with two deliberate human checkpoints.

$15

30-day refund guarantee

Secure checkout via Stripe

0 installsSecurity scanned

See it in action

You say

SpaceX just flew its first operational batch of flat-packed Starlink V3 satellites and Musk put a $200B revenue number on it. Make me a video — long cut and a short cut.

Your agent does

STAGE 1 — TITLE (checkpoint, waits for you) Three options, each committing the piece to a different angle: A. "The satellites are real. The $200 billion is arithmetic." B. "Musk's newest number, checked line by line" C. "What flat-packing a satellite actually buys you" → you pick. Everything downstream follows the angle you chose.

STAGE 2 — SCRIPT 733 words, three-part structure. Facts cross-checked against the primary source before they reach the page. On this story that meant: · two outlets gave conflicting cumulative sales figures → NEITHER went in · a projected number was labelled "planned capacity" every time, never output · the transcript of the talk was pulled and read, not an aggregator's summary of it — which is where the most useful line turned out to be

STAGE 3 — SIGN-OFF (checkpoint, waits for you)

STAGE 4 — VOICEOVER edge-tts → assets/narration.mp3 (3:52), captions.vtt, word-level cues. Completeness check: captions vs script. It has caught a real truncation — 92s of narration where 161s were expected, exit code 0, no error anywhere.

STAGE 5 — FOOTAGE First-party sources preferred, then every clip inspected across the span it will actually occupy, with the crop that will actually be applied: · one clip's title promised deployment footage; sampling its full 11 minutes showed it never shows the deployment → dropped · the best bay footage was from an earlier flight carrying SIMULATORS → used, but labelled on screen as exactly that, and only under narration about the mechanism, never under the sentence about the real launch

STAGE 6 — SCENE PLAN → plan.json, plan-short.json Boundaries taken from caption cue edges, never rounded.

STAGE 7 — RENDER (silent → audio → final, so a caption fix is not a re-encode) renders/starlink-v3-16x9-final.mp4 3:52 1920x1080 34 scenes renders/starlink-v3-short-9x16-final.mp4 2:41 1080x1920 26 scenes

STAGE 8 — QA Stills extracted and read. On this video they caught an on-screen label reading "Gen 2" while the narration discussed the third generation. ffprobe reports that file as perfectly healthy.

STAGE 9 — COVER + DISTRIBUTION COPY (+ optional YouTube upload)

Both finished cuts from this run are public: https://www.youtube.com/watch?v=RIgpSjUw_hM https://www.youtube.com/watch?v=sPMsLv1DURI

What you get

Turn a news story into a captioned shortRun a daily topic scan on your beat and get a ranked Top 6Write a fact-checked voiceover script to a target runtimeProduce 16:9 and 9:16 cuts of the same story

About this skill

Most video tooling starts after you already know what to say. This starts before that, and stops at a finished MP4.

topic-scan — scans your source list in priority order, verifies every URL resolves, scores candidates on reach, engagement, freshness and fit, and returns a ranked Top 6 with the lead flagged.

script-writing — five-step pipeline: title, frame, research, draft, sign-off. Cross-checks facts against at least three independent outlets before they reach the page. Picks one of five category templates (opinion, event, product, showcase, incident) and applies that category's cold open, analogy system and closing form. Produces a ~750-word three-part voiceover.

news-short-video — TTS, sourced b-roll, data and quote cards, burned-in subtitles, QA stills, covers, distribution copy, optional YouTube upload. Nine scripts, three render tiers, 16:9 and 9:16.

video-pipeline — chains the three and owns the seams between them. The seams are where these things actually break: facts arriving unverified, a word count line getting read aloud by the TTS, a script written to the wrong length for the platform it is going to.

WHAT YOU CONFIGURE

Three files, once. Persona and beat, your source list, 20–50 filter keywords. The method is fixed; the beat is yours. Switching from EV to biotech means editing those three files and nothing else.

THE TWO CHECKPOINTS ARE DELIBERATE

You pick the title, and you sign off the script. Everything in between — the research, the scene plan, the render, the QA stills, the cover — the agent does. Those two stay with you because the title sets the angle of the whole piece and the script carries the factual liability. This is not one-click video, and it is not trying to be.

WHAT IT IS HONEST ABOUT

It orchestrates external tools it does not bundle: ffmpeg, yt-dlp, edge-tts, Node. Run the included doctor.mjs first — one command tells you exactly what is missing and how to install it on your platform. Several of these dependencies fail silently (a missing font makes ffmpeg substitute a face and return success), which is why the preflight exists and why it exits non-zero.

Footage rights are yours. It fetches from public channels for commentary and citation and prefers first-party sources. Clearing that for commercial publication is your call.

The voiceover is synthetic. Every platform has a disclosure control.

This is the Latin-language build. TROUBLESHOOTING.md documents what breaks if you point it at CJK copy, and why.

TRACK RECORD

This runs a live channel, not a demo account. One cut produced with it has passed 130,000 views on RedNote. That one was Chinese-language; the two worked examples below are the English build, both public, both finished in a single pass through the pipeline. Watch them before you buy — they are the honest output, not a showreel.

Long cut, 16:9, 3:52 — https://www.youtube.com/watch?v=RIgpSjUw_hM Short cut, 9:16, 2:41 — https://www.youtube.com/watch?v=sPMsLv1DURI

WHAT SHIPS

34 files. Four SKILL.md definitions, nine Node scripts, reference docs for the scene-plan format, dependencies, upload and distribution copy, a TROUBLESHOOTING.md for the failure modes that return exit code 0 and pass ffprobe cleanly, and a commercial licence.