Workflow · Media and course pipeline
One spec, every format
Real work Rebuilt from a Claude plugin I built for a content team
Write the spec once: the topic, the audience, the steps and the voice. The pipeline turns it into a script, a voice-over, captions, a video, a course lesson and a reference card, all with the same words. Nothing is released until a person has listened to every line.
- Select Run the pipeline.
- Approve the listening check.
- Open each output: the script, the captions, the video, the lesson and the card.
The spec
Everything below comes from these lines.
The outputs
Five formats from one spec, with the same words in each.
One sentence per recording, so every sentence ends cleanly. Clear pauses sit between lines, and the timings come back measured.
The pronunciation dictionary
Words the voice gets wrong are respelled in the voice script only. Captions and transcripts keep the correct spelling. No word in this script needed a respelling. These are the fixes from earlier projects.
Built from the measured timings of the voice-over, so the captions never drift. Each cue names the speaker.
Read the transcript
A short lesson for the course, with a check. The questions use the same words as the video.
Responding to a chargeback
Keystone reference card
- Read the reason code. It tells you what the cardholder is claiming.
- Gather evidence that matches that reason.
- Add a short, factual summary of what your documents show.
- Submit before the deadline on the case.
Remember: a late response resolves against the merchant. The date on the case is the one that counts.
- Whether the voice sounds right. A person listens to every line, because no check can judge tone.
- Whether the steps are still true. When the product changes, the spec changes first, then every output is rebuilt.
- How well people learn from it. That needs the course results, not the build.
Safeguards
What keeps the workflow safe to run on a real knowledge base.
- Listen before releaseThe pipeline stops after the voice-over. Nothing is released until a person has listened to every line.
- One sentence per recordingEach sentence is voiced on its own, so none is cut off, and a fix means one new recording, not a new video.
- Captions from the same timingsCaptions, the transcript and the scenes all use the measured timings, so they always match.
- Respellings stay in the voice scriptThe dictionary changes what the voice reads, never what people read.
- No copied voicesThe voice is chosen from the provider's own library. The pipeline never copies a real person's voice.
- Licensed music onlyMusic comes from a library with a known licence, kept low under the voice.
- Real data never on screenScreens are drawn in code from invented data, so no customer detail can appear.
Take it to the next level
How to get results that match the best teams, and what to build next.
- Every language, nativelyGive each language its own native voice, with its own speed and pause, rather than one voice reading every language.
- Rebuild on changeWatch the source article. When it changes, rebuild the script, captions, video and card, and flag them for a listen.
- Measure learningTrack lesson completion and check scores, and compare video with text-only versions of the same topic.
- Audio descriptionAdd a described version for blind and partially sighted people, built from the same scene list.
- A motion systemKeep scenes, timings and transitions as reusable parts, so every new video matches without hand animation.
- Spot where people drop offUse video analytics to find the second where viewers leave, and fix that scene first.
Rebuilt from a Claude plugin I built for a real content team. Keystone is an invented payments product, and every name, figure and conversation here is made up.