Workflow

Workflow — for the PM

How to operate the SuperStories video system day-to-day. Assumes the setup is complete. The training methodology that improves the system over time lives on the training page. The market positioning lives on the comparison page.

The system separates production into three layers, plus a review surface that runs across all of them. Your job is to drive each layer in the right order, with the right inputs.

The three production layers, briefly

1. Brand foundation. The BRANDGUIDE.md per client. Everything downstream depends on it.
2. The cut. Editorial — what makes it into the video, in what order.
3. The polish. The branded layer — graphics, motion, captions, components, channel adaptations.

Plus the review surface (our own annotation tool at videos.superstories.com) that handles internal review, team review, and client sign-off in one place.

Working with files: the uploader

All footage and renders live in Cloudflare R2. You reach them through the uploader at uploader.superstories.com. Log in with your username and password (ask Peter if you don't have one yet). Nothing to install, no keys to manage.

• Upload: open the right folder, then drag files into the window. Large videos are fine; there is no size limit.
• Browse: click a folder to open it. The breadcrumb at the top jumps back up.
• Download: click a file.
• New folder: use the New folder button.

If a file does not show up after an upload, refresh the page. If the page says you are not allowed in, your login was removed or mistyped; ask Peter.

Step 1 — Starting a new client

Before the first video, build the brand foundation. This is half a day to two days of focused work depending on the client. Without it, every video downstream is generic.

1. Open the client's folder: video/clients/<client>/.
2. Check whether a BRANDGUIDE.md already exists.
   • If yes, that is the canonical foundation. Skip to step 4.
   • If no, run the /brandguide skill in Claude Code from a folder of the client's identity assets. It scans, asks questions, and generates a structured brand guide.
3. Open the generated BRANDGUIDE.md and review it with the client in a 60-minute session. Edit until it feels accurate.
4. Add the video-specific sections (Motion Principles, Caption Styling, Video Components, Editorial Patterns). They start as DRAFT until populated.
5. Populate references/videos.md with 5-10 videos you and the client love, each with a one-line note on why. This is the taste anchor. Without this, output drifts to generic. Treat as load-bearing.
6. Run a motion exploration sprint — see motion-exploration/README.md in the client's folder. Half a day. Result: 2-3 motion patterns codified into the brand guide's Motion Principles section.
7. Commit and push to git. The brand foundation is now ready for video work.

Step 2 — Starting a new video project

For each video:

1. Create a new file in video/clients/<client>/projects/<YYYY-MM>-<slug>.md.
2. Write a structured brief in that file. Do not write "edit this video." A brief names:
   • The hook moment (the line or visual that opens)
   • 2-3 key beats to emphasize with graphics or chapter marks
   • The CTA at the end
   • Target length
   • Target channels (YouTube 16:9, Shorts 9:16, LinkedIn 1:1 — pick all that apply)
3. Create the project folder in the uploader (uploader.superstories.com) at <client>/<project-slug>/, using New folder.
4. Commit the brief to git. The project exists.

Step 3 — Recording the source

Pick the recording tool that fits the format:

• Founder-to-camera (face only). iPhone front camera on a tripod. AirDrop to Mac.
• Screen + cam (demos, walkthroughs, training modules). Screen Studio.
• Quick drafts and multi-source. Tella, browser-based.
• Multi-person interview or podcast. Riverside.fm.

Upload the raw source to <client>/<project>/sources/incoming/ via the uploader (uploader.superstories.com): drag it into the browser. The input normalizer, which runs server-side against R2, transcodes it to a clean MP4 in sources/normalized/ within a few minutes. You don't have to do anything else. Refresh the folder and the normalized file appears.

For long content (training, expert interviews, 60+ minutes), the long-form section below covers what changes.

Step 4 — The cut

Pick the editorial tool per length:

• Long-form talking-head (founder explainers, interviews). Gling. Auto-removes silences, filler, and bad takes.
• Short-form social (Reels, Shorts, TikTok). CapCut. Free, capable, designed for short-form.
• Multi-person podcasts and multi-cam. Descript. Transcript-based multi-track is strong here.
• Long-form trainings (60+ minutes). Descript. Transcript editing at length is far stronger than Gling.

Download the normalized MP4 from the uploader. Cut. Export. Upload the cleaned cut back to <client>/<project>/sources/normalized/ as cut.mp4.

If a remote editor is doing this step instead of you, see coordinating with the editor for the handoff.

Step 5 — The polish

There are two kinds of polish, produced and rendered differently. Keep them apart, because conflating them is what made earlier renders fail.

5a — Branded graphics over a cut

Lower-thirds, titles, captions, chapter cards over talking-head footage. Structured overlay-on-video work. Claude builds the graphics layer; you review a preview and iterate.

5b — Generative motion backgrounds

The brand motion pieces (audio-reactive WebGL / canvas, like the podcast animations). Claude generates the look; they are not editable in place — you do not edit text inside them.

• They are driven by the live audio, so they cannot be frame-rendered (HyperFrames / timecut give you only the intro and music, no captions). Render them by playing and recording the page: render/render.mjs (runs headed on a Mac, in real time). See render/README.md.
• Render the background text-free (--no-text). The captions are pulled out into an editable captions.srt next to the animation.

The rule for the producer

Claude's job ends at the motion background. Text, captions, sizes, line breaks, trims and the final assembly happen in the editor (Descript / CapCut / Premiere), with the rendered background as a clip and the .srt imported as captions. A one-word fix is an srt edit, not a re-render and not "ask Claude to rebuild the whole thing". Don't try to make Claude Code an editor; we are not rebuilding Canva or After Effects. The full producer hand-off is in render/FOR-THE-PRODUCER.md.

Step 6 — The review loop

1. Publish the draft to the review hub:

   ss-publish ./renders/16x9.mp4 --client=<x> --project=<y> --status=review
   

2. The CLI uploads to R2, generates a thumbnail, registers in the hub, and prints the review URL.
3. Share the URL with whoever needs to review. The annotation tool is on every review page.
4. Reviewers scrub through, drop pins at timestamps with notes, click "Copy notes."
5. The notes paste into a Claude conversation. Claude reads them, revises the HyperFrames source, re-renders.
6. Re-publish with ss-publish. Loop until approved.

The same surface handles internal review, team review, and client sign-off. No Frame.io.

Step 7 — Long-form trainings (60+ minutes)

Trainings are not just "longer videos." The balance between layers shifts.

• The cut is the heavy lift. Use Descript, not Gling. Transcript-based editing at length is the right tool. Read the transcript like a Google Doc, cut paragraphs, video follows.
• HyperFrames does less motion but more structure. Title cards, chapter markers, terminology lower-thirds, slide overlays, captions throughout, end cards. Predictable graphical support over the whole duration.
• Summaries, key takeaways, and recap cards become first-class graphical interludes every 8-12 minutes. They are auto-generated from the transcript (Whisper + Claude) and styled by the brand foundation. This is genuinely powerful for trainings and not feasible to do by hand at length.
• Whisper for captions and a deliverable transcript. Trainings need closed captions for accessibility and a searchable transcript for the LMS. Generate once, use everywhere.
• Render time scales linearly. A 90-minute render takes 30-90 minutes. Render overnight or split into chapters and stitch.
• One source, multiple outputs. Full training + per-chapter modules + 9:16 promo clips — all from the same HyperFrames source.
• Multi-language is cheap. Add a second SRT, re-render. Branding stays identical.

Step 8 — Coordinating with the editor

When a remote editor does the cut instead of you (the default for senior editor work):

What you give the editor:

1. The brief from projects/<slug>.md.
2. A login to the uploader (uploader.superstories.com), see setup Step 1d, so they can download the normalized source and upload their cut. Nothing to install.

What the editor delivers:

1. The cleaned cut uploaded as cut.mp4 in <client>/<project>/sources/normalized/.
2. Optionally a Loom walkthrough of unusual decisions (not required, useful for complex projects).

What you do next: Continue with step 5 (polish). The cut is now your input.

For the training loop where the editor's choices are compared against the AI's choices, see the training page. That's a different track that runs in parallel.

Step 9 — Multi-channel output from one source

For every video, default to producing the relevant channel formats from the same HyperFrames source, not as separate edits:

The brief in step 2 should name target channels. HyperFrames renders each.

Step 10 — Thumbnails

For YouTube and LinkedIn, thumbnails ship alongside the video.

1. Ask Claude: "Generate thumbnail variations for <project>. Reference high-performing thumbnails in <niche> via Apify."
2. Claude analyses competitor thumbnails, identifies patterns, generates branded variations using the brand foundation.
3. Pick two for A/B testing. Publish.

First-time path — your first brand guide and first video

Expect the first video to take longer than later ones. The brand guide and project setup are the heavy lifts. After that, per-video time drops to 30-60 minutes.

What we don't do

• We don't write "edit this video" as a prompt. We write a brief.
• We don't review by writing timestamps in a Google Doc. We use the annotation tool.
• We don't default to Frame.io. Our own surface handles internal and client review.
• We don't try to do cinematic motion graphics inside this workflow. For a hero campaign piece, commission a motion designer.
• We don't ship without the taste anchor referenced. If references/videos.md is empty, the foundation is incomplete.