Skip to main content

Video Publish Artifacts SOP

Overview​

  • What: Turn a raw recording into everything needed to publish it: an edited master (dead air out, pauses tightened, stills, end card and logo in), a 16:9 delivery, 9:16 and 4:5 social cuts, captions, a poster frame and a YouTube thumbnail.
  • When: A take has been recorded and the video is going to YouTube, the website or social.
  • Who: An agent runs every step. A person approves at two gates and publishes. A person without the toolchain asks an agent to run it.
  • Time Required: About 30 minutes of machine time per video, plus two short reviews.

Write the spec first. One video is one .video.yaml beside its script, holding the cut, stills, overlays, caption corrections, poster frame, cuts and thumbnail brief. Every step below is a command against it, so a re-run reproduces the same video. Worked example: audit-promo-c.video.yaml.

Who does what​

#StepOwnerOutput
1Record per the Teleprompter Script Preparation SOP. Walking in and out is fine; on a fluffed line, pause and restart the sentence without stoppingPerson.mov in _symphony-core-documents-inbox/
2Ingest. Move the take to the series folder as YYYY-MM-<slug>-master.mov. The master is never modified after thisAgentMaster in place, inbox empty
3Plan the edit (§A)AgentProposal, review and proxy in edit-review/
4Gate 1, content. Watch each proxy, read the review, approve or name the linePersonApproved, or notes
5Put the approved cut in the spec; add stills, overlays, logo and end card (§A, §C)AgentComplete spec
6Render and verify (§B)AgentFiles in publish/<slug>/
7Gate 2, finish. Open the contact sheets, watch the 16:9 oncePersonApproved
8Thumbnail (§D)Agent1280×720 PNG
9Publish and record the URLs (§E)PersonLive
10Retro. One line per issue into Learnings; anything recurring becomes a video-doctor check or a spec defaultAgent drafts, person acceptsA better next run

Prerequisites​

MasterThe original recording. Trimming and stills are applied by the pipeline, never by hand beforehand. Never derive a cut from a delivery file
Toolchainffmpeg, whisper-cpp with ~/.cache/whisper-models/ggml-base.en.bin, and content-conductor checked out with its .venv. video-doctor checks all of it
SC_DOCS_STORAGESet so spec paths resolve; masters, stills and logos live in the sc-docs-storage shared drive. macOS export SC_DOCS_STORAGE="$HOME/Library/CloudStorage/GoogleDrive-<account>/Shared drives/sc-docs-storage"; Windows set SC_DOCS_STORAGE=G:\Shared drives\sc-docs-storage
SC_DRIVEThe My Drive root, still used by content-conductor for its own storage (${SC_DRIVE}/content-conductor-storage/…). macOS export SC_DRIVE="$HOME/Library/CloudStorage/GoogleDrive-<account>/My Drive"; Windows set SC_DRIVE=G:\My Drive
StillsRendered from the series' insert HTML once per ratio (?r=16x9, 9x16, 4x5): the main graphic full-frame, nothing beside it. Checked as a labelled sheet before the first render
Approved copyTitle, description and social captions, in the script's publish pack
CanvaThe agent uses the Canva MCP connector, which needs no browser login. The browser login is in the Tool and Account Register

Procedure​

Commands run from the content-conductor checkout as .venv/bin/python main.py <command>. Full option reference: content-conductor docs/guides/media-processing.md.

A. Plan the edit​

video-edit-plan -s <spec>.video.yaml -o <series>/edit-review \
--script-text <teleprompter>.txt

It writes <slug>-edit-proxy.mp4 (the cut at 540p, for gate 1), <slug>-edit-review.md (kept and cut, speech versus noise, audio peak, any border, and a comparison with the script) and <slug>-edit-proposal.yaml (the edit: block for the spec).

Defaults: keep from 0.4 s before the first word to 0.6 s after the last; cut pauses over 0.6 s down to 0.3 s; crop any recorder border; with punch_in: 1.1, change framing only at sentence breaks.

Problem at gate 1Fix
A fluffed line the speaker repeated--remove START-END (raw-take seconds)
The first frames catch the mouth forming the first word--head-pad 0.14 --fade-in 0.3
Pace feels rushed--max-pause 0.4 or --pause-threshold 0.8

Read the script comparison as questions. whisper mishears. To prove nothing was clipped, transcribe the proxy itself. A peak at or above -1 dBFS is flagged: normalising cannot repair a clipped syllable, so listen to the loudest words.

Then add the stills, logo and end card to the spec, anchored to what was said, not to seconds:

inserts:
- image: "${SC_DOCS_STORAGE}/…/insert-07-step-1.png" # landscape
cuts:
vertical: "${SC_DOCS_STORAGE}/…/insert-07-step-1-9x16.png"
feed: "${SC_DOCS_STORAGE}/…/insert-07-step-1-4x5.png"
start: {say: "Step one"}
end: {say: "Step two"}
end_card:
image: "${SC_DOCS_STORAGE}/…/insert-10-close.png"
cuts: {vertical: "${SC_DOCS_STORAGE}/…/insert-10-close-9x16.png", feed: "${SC_DOCS_STORAGE}/…/insert-10-close-4x5.png"}
hold: 3.0
branding:
logo:
image: "${SC_DOCS_STORAGE}/…/symphony-core_web_wide_trans-reverse_2000x547.png"
position: top-right
width: 0.16

A phrase that was not said, or was said twice without occurrence, stops the run. Where whisper mishears a word, anchor on the line before it. Text never sits over a still; the pipeline refuses a spec where it would.

B. Render and verify​

video-doctor  -s <spec>.video.yaml            # toolchain, master, stills
video-publish -s <spec>.video.yaml -o <out> --dry-run
video-publish -s <spec>.video.yaml -o <out>
video-verify -m <out> # measurements, contact sheets

Run video-doctor first, every time. It finds a missing filter, font or still in seconds rather than part-way through a render.

Leave the picture alone. Audio is always normalised to -14 LUFS. Exposure or colour correction is applied only when a side-by-side comparison proves it better, per video-correction-measured-not-better.

Correct captions in the spec, not the file. Read the SRT against the script and add each fix to captions.corrections; a correction that matches nothing stops the run. Captions ship as an .srt sidecar, because Homebrew's ffmpeg cannot burn them in (video-doctor reports it).

Pick the poster frame deliberately: eyes open, mouth nearly closed, facing camera.

C. Overlays​

Text in the spec is what a sound-off viewer gets, so it is not optional. Map each storyboard line to a phrase that was said:

overlays:
- style: statement
text: "Customers look you up first"
start: 0.3
end: {say: "Google listing", offset: -1.0}
- style: lower_third
text: "<Presenter name>"
subtitle: "<Role>, Symphony Core"
start: {say: "Our free marketing audit"}
end: {say: "One your Google profile", offset: -0.3}
- style: url
text: "symphonycore.com/audit"
start: {say: "Run it free"}
end: end

Styles are statement, lower_third and url; type is fitted per aspect ratio automatically. Overlapping cues are rejected. Open the video-verify contact sheets: they sample every overlay, every still and the frame after every cut, in every ratio, and are the only check for type that overflows a narrow crop.

D. Thumbnail​

Copy the master; never generate. The vetted 1280×720 master and its rules are registered in the Static Creative Standard §9; the spec's thumbnail.canva_master carries its id.

  1. copy-design from the master. Element ids survive the copy.
  2. read-design with open_transaction: true to get the locator_ids.
  3. edit-design with find_and_replace_text for the headline (not replace_text, which flattens mixed formatting); insert_fill if the photo changes.
  4. Show the person the preview. On approval, edit-design with finalize: "commit", as a separate call.
  5. export-design as PNG at 1280×720, immediately. A Canva design keeps its id after edits, so an export is the record of what was approved.

Headlines stay on one line: about 20 characters at 108 pt. Test at phone size: scale to 168×94 and read it. If it does not read, cut words; do not shrink the type.

E. Publish​

A person publishes; the agent stages, per draft-means-draft-only.

  1. YouTube: the YouTube Video Publishing SOP.
  2. Website: embed as a lazy-loaded poster frame with a play button, no autoplay.
  3. Social: the GHL Social Planner Posting SOP. 9:16 to Reels, Shorts and Stories; 4:5 to the feed; links in the first comment.
  4. Record the URLs in the script's published_urls frontmatter.

Verification​

video-verify checks size, duration and loudness on every file and builds the contact sheets. A person checks the rest:

WhatHowTarget
Loudnessvideo-verifyEach file within ±0.5 LUFS of -14
Stills and overlays rightContact sheets, every ratioRight card on each beat, no clipped type
Joins cleanContact sheets, join frames; watch the 16:9 onceNo flash frames, no audible clicks
Captions accurateSRT against the scriptNames, numbers and URL correct
Master untouchedls -l mtimeUnchanged
Thumbnail168×94 test; compare with the masterReadable; only headline and photo differ
Spec committedgit status.video.yaml tracked
PublishedEach URL, logged outPlays, captions present

Learnings​

One line per issue found in a run. When an issue recurs, it becomes a video-doctor check or a spec default, and the line says so.

DateVideoIssueResolution
2026-10-09Audit promo A-COBS left a 16 px right and 10 px bottom black borderThe edit crops it. In OBS: right-click the source, Transform › Fit to screen
2026-10-09Audit promo AThe first frames caught the mouth forming the first word--head-pad 0.14 --fade-in 0.3
2026-10-09Audit promo A-Cwhisper dated sentence-final words into the pause, so the review reported kept words as missingThe review now tests overlap. Transcribe the proxy to prove nothing was clipped
2026-10-09Audit promo BA push-in would have changed framing mid-phrasePush-in changes only at joins that removed a second or more
2026-10-09Audit promo A, BAudio peaked at -0.5 and 0.0 dBFSFlagged at gate 1. Lower the mic gain 3 dB next session
2026-10-09Audit promoThe closing card's URL touched the panel edgeType size reduced in the insert HTML
2026-10-12Audit promo A-CStills carried a logo, URL and progress grid beside the graphic, so one image could survive the vertical cropOne still per ratio, the graphic full-frame (cuts: on each still)
2026-10-09Audit promo A-CEvery still was saved under the previous slide's name (a zsh array counts from 1), so the first render showed the wrong card on every beatRe-rendered under bash. Check the stills as a labelled sheet before rendering. Gate 2 caught it; the measurements all passed

Revision History​

DateVersionChangeAuthor
2026-11-173.4Spec paths resolve through SC_DOCS_STORAGE, the sc-docs-storage shared drive, after storage moved off My Drive (issue #593). SC_DRIVE stays the My Drive root for content-conductor storage, and Windows paths use G:.Symphony Core
2026-11-123.3The logo overlay example now names the current white logo master. The file it named was an older, retired wordmark.Symphony Core Systems Team
2026-10-123.2Stills are rendered once per ratio, showing only the main graphic full-frame; the spec gives each still its cuts: (content-conductor PR #12).Symphony Core Systems Team
2026-10-093.1Cut from 387 to about 200 lines to meet the SOP Standard. The hand-typed "human path" ffmpeg and whisper commands were dropped; a person without the toolchain asks an agent. The thumbnail master id and Canva login now point to the Static Creative Standard and the Tool and Account Register; the overlay example uses a placeholder name. Caveats now enforced by video-doctor (libass, fonts, subtitle paths) were reduced to one line each. History moved here from the body: the first video's parameters lived in a temp directory and were nearly lost, which is why the spec comes first; video-doctor exists because the first run found a missing filter and a missing font mid-render; one cut shipped reading "re not short on l" before contact sheets were added; generate-design needed three correction rounds on its first run, and copying the master needed none; and the first thumbnail was exported at 14:31 with its design edited at 15:05, unnoticed for a day, which is why export follows approval immediately.Symphony Core Systems Team
2026-10-093.0Starts at the raw take. Adds the edit (video-edit-plan, two review gates, stills, end card and logo with phrase-anchored cues), the who-does-what table and the Learnings table.Symphony Core Systems Team
2026-09-152.1§D's YouTube step became a pointer to the new YouTube Video Publishing SOP. Added the re-export rule and the one-line headline budget.Symphony Core Systems Team
2026-09-152.0Steps became commands against a versioned .video.yaml. Caption corrections moved into the spec. The thumbnail switched from generating to copying a vetted master.Symphony Core Systems Team
2026-09-151.1Added on-screen text overlays, rendered per aspect ratio, with overlay-free -clean copies kept.Symphony Core Systems Team
2026-09-151.0Initial SOP, from the first end-to-end run on the SC explainer V1.Symphony Core Systems Team