Most "faceless channel" advice is: paste a prompt into ChatGPT, edit in CapCut, pray. I went a different direction — a document-driven pipeline where coding agents (Devin/Claude/Codex-class) do the actual production work, and humans touch exactly two points: approving which source to produce, and deciding to upload.
130+ videos later, here's what the system looks like.
Documents are the API
Every stage produces a document the next stage consumes:
candidate card → production request → QA report → upload checklist
Agents can't drift because the docs pin the format. The candidate card carries the source URL, duration, resolution, story beats, and dedup status. The production request carries the edit decision list — frame-precision cuts, not vibes.
What agents are actually good at (measured, not vibes)
QA is where agents earn their keep. These are mechanical checks, and agents catch things I miss:
- Contact-sheet visual review — thumbnails across the timeline, catches broken renders and baked-in text
- ASR-vs-script checks — whisper output diffed against the intended script
- Loudness + A/V sync — 2-pass loudnorm, sync offset measurement
- Dedup ledger — a persistent ledger of used sources caught a duplicate before a paid generation (real money saved)
Taste stays human. Which candidate to produce, whether the joke lands — agents are bad at this and I stopped asking them.
The economics nobody talks about
Google Flow generation costs ~150 credits per 75s episode including rejected generations. The fix was a still-image storyboard gate: approve the stills before spending on motion. Rejects dropped roughly in half.
TTS assets get sha256 manifests so a render is bit-verifiable against its inputs. Every episode has a machine-readable release QA doc — pass/fail, not "looks fine I guess."
Render-script editing
Editing happens in code: EDL → ffmpeg filtergraph. Baked-text removal, caption timing, loop design (LOOP_START markers so the last beat feeds the first frame). Code-rendered video too — kinetic typography, motion graphics, image→video, Blender previz, headless-Chrome rendering.
The honest part
I packaged the whole operating system — the doc formats, QA checklists, render skeletons, Flow prompt patterns, season ops (96-episode presets, pronunciation dicts, A/B format experiments), algorithm gates, licensing notes, and a 30-day launch plan — into a 44-page guide with 8 templates and 4 sample videos rendered by the actual scripts.
It sells the operating system, not "guaranteed views." Anyone selling guaranteed views is lying to you.
→ https://wjswnsdh.gumroad.com/l/ai-agent-video-studio ($19 launch, LAUNCH50 = 50% off for the first 10)
Happy to answer architecture questions in the comments — the doc-driven-agent pattern generalizes way beyond video.
Top comments (0)