Training

How We Train This Tool — for the Video Editor

This page is for the editor who works with us. It explains the methodology, what you deliver, what we deliver back, and the rhythm of the collaboration. You are not training your replacement. You are calibrating a system that captures and amplifies your judgement.

The infrastructure that makes this possible lives on the setup page. The day-to-day workflow that consumes it lives on the workflow page.

The methodology in one paragraph

Your final cut is the ground truth. The AI's attempt at the same brief is the hypothesis. The differences between your cut and the AI's cut are the training data. You never have to articulate rules in the abstract. You only react to specific moments where the AI diverged from your judgement, and that reaction becomes the rule. The system converges toward your taste over a small number of iterations.

This is closed-loop training with explainable instructions, not opaque weights. The result is a system that cuts like you do, that you continue to calibrate.

What you sell

You are not selling your time. You are selling your taste and judgement, captured in a system that you continue to calibrate.

In the old model, an editor cuts every video. Volume drives revenue, but the editor's deeper skill — what to keep, what to discard, when to break the rhythm — leaves the building the moment they do.

In this model:

• You remain the only source of truth the system references.
• Your role shifts from "cut every video" to "cut the hard cases plus calibrate the system via diffs."
• You become scarcer and more valuable per project, not less.
• Compensation is for the calibration time, not just the cuts.

This is important. We should have this conversation openly before the first session, not assume it.

Step 1 — What you deliver per project

Four artefacts. Nothing else required.

1. The brief you worked from. We provide it; you reference it as you cut.
2. The raw sources. Already on R2 in sources/normalized/ for the project.
3. Your final cut. Delivered to R2 in the project's renders/ folder or a sources/normalized/cut.mp4 slot depending on the workflow.
4. Optionally a short Loom walkthrough for unusual projects, where you explain choices that are not obvious from the cut itself.

That's it. No decision log written by hand. No documentation effort beyond the optional Loom.

Step 2 — What the AI does next

After you deliver, the system runs:

1. The AI is given the same raw sources and the same brief you worked from.
2. The AI produces its own attempt at a cut.
3. The Cut Diff Tool aligns your cut and the AI's cut by Whisper transcript (same source, same words — what stays in each version differs).
4. Differences are highlighted: green for "only in your cut," red for "only in AI's cut," neutral grey for "in both."

You don't have to do anything during this stage. It runs in the background.

Step 3 — Your 20-minute review session

This is the only documentation work you do. It is short and concrete.

1. Open the Cut Diff Tool at videos.superstories.com/diff/<project>.
2. Scroll through the highlighted differences in order.
3. At each difference, leave one short line of reasoning:
   • "That should have been cut — agree."
   • "Keep this — Connirae has a deliberate rhythm there."
   • "Chapter break because the concept shifts here."
4. Click "Copy diff comments" when done.

Your reasoning becomes the training signal. The system updates its instructions — client-specific patterns to that client's brand guide, universal patterns to the cross-client recipes library.

Total time: 20 minutes for a typical project. No prep, no after-work.

Step 4 — The iteration loop

1. AI is given: raw sources + brief
   → AI produces its cut

2. Diff tool aligns AI cut with your cut
   → lists each timestamp-divergence

3. You review the divergences (~20 minutes)
   → respond in one line per divergence

4. Claude reads the responses
   → updates the recipes:
     - client-specific patterns → client BRANDGUIDE
     - universal patterns → recipes/cut-recipes.md

5. AI re-attempts the same brief with updated recipes
   → divergence count drops

6. Iterate until divergence is "acceptable" — defined together with you
   (e.g. "I would make 8/10 of these cuts the same way")

A typical project converges in 2-3 iterations.

Step 5 — The first 6 reels (training/test split)

For the initial calibration of the system on your work, we use classic ML protocol:

• 4 reels as training set. We iterate AI until it matches your cuts within tolerance. Recipes form here.
• 2 reels as test set. We hold them back. After recipes stabilize on the training 4, we run AI on these 2 with no further instruction. We diff against your version.
  • If divergence is comparable to training → recipes generalize, ready to scale.
  • If divergence is much larger → recipes overfit, we need more variation before scaling.

This prevents the system from looking polished on the training material and failing on new content. We share the test results with you so you see the convergence in real numbers.

Step 6 — The biweekly sync (recorded)

Every two weeks, 30 minutes:

1. You walk through one recent cut explaining choices that are not obvious from diff comments alone.
2. We record the call (Whisper transcribes; Claude extracts patterns).
3. The recording becomes library material — often the deepest knowledge lives in how you think, not in any specific cut.

Optional. Useful for senior editors who enjoy teaching. Skip it if it isn't your style.

Optional supporting material

If you want to add context beyond diff comments, two formats work well — neither is required.

1. WhatsApp voice notes. Speak 5-10 minutes after a complex cut explaining decisions. Whisper transcribes, Claude formats into the cut recipes.
2. Loom walkthroughs. Scrub through a cut explaining what you did. Loom has built-in transcription. Claude reads it and extracts patterns.

Both are easy on you. We turn them into structured instructions.

Why this beats writing decision logs

The earlier version of this methodology asked editors to write decision logs by hand. We do not do that anymore.

You should never feel like you are doing documentation work. You should feel like you are giving short, specific feedback on a draft.

What this means for your career

In the short term: your hours stay roughly the same, but distributed differently. Less time cutting routine videos, more time on the hard cases and on calibration.

In the medium term: your per-hour value goes up. You are now the only source the system references for your style of work. Your skill is being captured, not commoditized.

In the long term: the cut-recipes library you helped build becomes a SuperStories asset that depends on you to evolve. You become structurally hard to replace — not because you hoard knowledge, but because you continue to update the source of truth.

This is what "you sell taste, not time" looks like in practice.

What we ask of you before we start

Three things, none of them paperwork.

1. Have the open conversation. If anything in this methodology feels off or like a trap, say so before we start. We adjust together.
2. Commit to the training/test split for the first 6 reels. Skipping it means we don't know whether the system generalizes. Worth the discipline.
3. Be honest in diff comments. If a difference doesn't have a reason, say "no strong preference." That is also a useful signal. Don't manufacture rules.

That is the whole ask.