Build a Video to Shorts Pipeline

Step 3 of 9 · About 30 minutes per recording, most of it your review

Mine the Transcript for Ideas

Read the entire transcript and extract ideas, not timestamps. This is the step that separates a pipeline worth running from an automatic clipper.

What you get

A slate of 14 to 16 candidate ideas, each with a thesis, a hook line, mapped source timestamps, and an assigned goal.

Automatic clippers slice on a sliding window and score each window for engagement. That is why their output starts mid thought and refers to things that happened off clip. We invert it: ideas first, timestamps second.

What counts as an idea

An idea is a thesis, not a moment. A claim, a story, a tip, a hot take, or a demo payoff that a cold viewer can follow with no other context. If you cannot write it as one sentence that would make a stranger stop scrolling, it is not an idea yet.

Critically, the supporting moments may be non contiguous. Setup at minute 12 stitched to the payoff at minute 38 is not merely allowed, it is usually better than either piece alone.

Write the idea miner as a skill

Make this a reusable skill file rather than a one off prompt. You will run it every week and you will want to tune it.

Paste into Claude Code
Create skills/shorts-idea-miner/SKILL.md.

Frontmatter: name shorts-idea-miner, plus a description saying it mines the
full transcript for shorts IDEAS and assembles each into a coherent clip plan
from word timestamped segments, non contiguous allowed, idea first, never
sliding window slicing.

Body, three stages:

Stage 1, idea mining. Read the ENTIRE transcript. Extract candidate ideas. An
idea is a thesis, not a timestamp: a claim, story, tip, hot take, or demo
moment that stands alone for a cold viewer in my audience, which is
[YOUR AUDIENCE]. For each idea record the thesis sentence, why it hooks, and
the mapped supporting moments as timestamps.

Stage 2, clip assembly with a hard coherence gate. For each idea pull exact
word timestamped segments. Segments may be non contiguous. Every segment must
pass: complete thought with no mid sentence entry or exit, no orphaned
pronouns, no inside references a cold viewer will not get. Ideas that cannot
be made coherent are KILLED, not padded.

Stage 3, edit spec. Per surviving idea: reorder so the strongest line lands in
the first 1 to 3 seconds, a word level caption spec, a 9:16 crop plan, text
overlays on key numbers, and a CTA end beat matching my goal split.

Output one clip plan per idea.

Run it against the transcript

Paste into Claude Code
Read work/ep-test.words.json in full and run skills/shorts-idea-miner
against it.

Write the result to episodes/<date>/shorts-slate.md, organised in tiers.
Tier 1 is what I cut first, Tier 2 is the rest.

For every idea give me exactly this shape:

### S1. "Short title in quotes"
- Thesis: one sentence, what the idea actually claims
- Segments: mm:ss-mm:ss (what happens) then mm:ss-mm:ss (what happens)
- Hook: the exact first line a cold viewer hears
- CTA: community or services
- Coherence: clean, or the specific problem to solve when cutting

Mine 14 to 16 candidates. I need 10 to survive the gate later, so do not stop
at 10. Do not pad weak ideas to reach the number. Give me the honest count and
flag the marginal ones.

Rules to apply while mining:
- Never surface an idea that criticizes or names a competitor
- Only these people may appear on camera or voice: [YOUR APPROVED LIST]. If an
  idea depends on anyone else, say so and mark it for a reshoot.
Check it worked
Open the slate and read only the Hook lines, in order, as a stranger would. Anything that does not make you want to hear the next sentence is a weak idea no matter how good the underlying moment was. Move it to Tier 2 or kill it now, before you spend render time on it.
Where this bites
The miner will happily hand you ten variations of your single best moment. Read the theses side by side and confirm each is a genuinely different claim. Ten shorts that all say the same thing is one short and nine wasted slots.
If your recording covers a specialised domain, this is where a domain brain earns its keep. We point the miner at Ambrose so it can verify claims and pull supporting data while it mines, rather than surfacing an idea that sounds great and turns out to be wrong.

Questions

Why mine ideas instead of scoring windows for engagement?

Because an engagement score cannot tell whether a clip makes sense. A window can be energetic, loud, and completely incoherent to someone who did not hear the previous two minutes. The thesis is what lets a clip survive being watched cold.

How many ideas should one hour produce?

Fourteen to sixteen candidates from a good hour long call. If you are getting five, the call was mostly logistics, and the honest answer is fewer shorts this week.

Talk it through

Stuck on this step?

Free 30 minutes. Bring the error, the transcript, or the clip that will not pass the gate, and we will work through it.