Practical guides
Step-by-step workflows, reusable briefs, and agent guardrails
Run a Build-in-Public & Talking-Head Video Automation Pipeline
Complete a full content loop in 1–2 hours: pain-point outline, single-take screen recording, VAD+ASR retake rough-cut, screen-OCR caption correction, 16:9/3:4 covers, and draft uploads.
I'm recording a practical video on saving 90% of model budget with sub-agent teams. First generate a one-screen speaking outline (one-line conclusion, 3-step demo order, anti-tangent reminders, and personal story hooks—no verbatim script); once recorded, run VAD+ASR retake rough-cut, screen-OCR term subtitle correction, 16:9 & 3:4 cover generation, and timestamped draft metadata.
Skip robotic word-for-word scripts. Tell the Agent what real problem you solved today and have it generate a speaking outline: one core takeaway, 3 demo beats, explicit 'do not digress into' reminders, and personal story prompts.
- Each topic solves a concrete pain point and shows the outcome in the first 3 seconds
- Outlines list takeaway, sequence, and 'do not expand' reminders rather than verbatim scripts
- Repeat stumbled sentences in place during a single continuous recording
What to keepThe outline fits on one screen and supports a natural single-take screen recording.
When you stumble during recording, simply pause and repeat the sentence cleanly without stopping. Afterward, run the rough-cut pipeline: VAD trims dead air, and the Agent uses word-level ASR timestamps to drop earlier fumbled takes and keep the final clean take.
What to keepA raw 25-minute recording with retakes automatically condenses into a tight 8–10 minute rough cut.
Extract video keyframes to read visible filenames, model names, and CLI flags so the Agent fixes homophone typos in ASR captions; pair a clean speaker frame with high-contrast SVG type to render 16:9 and 3:4 covers in a local HTML preview.
- Export both 16:9 and 3:4 covers and verify legibility at thumbnail size
- Automation stops at saved platform drafts and never triggers public publishing
- Archive transcripts and performance metrics in the project folder to refine future pacing
What to keepTechnical terms in captions are spelled accurately, and both covers remain legible at mobile thumbnail size.
Generate timestamped chapters and platform tags from the transcript, upload assets into platform drafts with post-check validation, and leave the final Publish click to a human. Later, feed analytics and the text transcript back to the Agent to diagnose pacing.
What to keepPlatform drafts are populated and verified without auto-publishing, and transcripts plus metrics are archived in the project.