Claude Code: turn a podcast into five contextual short videos

Can Claude Code help turn a long podcast into five shorts?

This workflow is possible when Claude Code can access the video and suitable media inspection, transcription and rendering tools. It does not guarantee the five best moments or viral results. An editor must review content, context, permissions and output quality.

This practical companion to BHKD’s ten video prompts provides original templates for selection, context checking and clip review. It is not a report of a test performed on a client podcast.

1. Prepare assets and define the goal

Put source-video copies, any transcript, logo and permitted assets in input, with output for drafts and finals. Define the audience, intended platform, goal and topics requiring context. Being able to download someone else’s video does not itself grant reuse rights.

Inventory prompt: Inspect input files and report readable assets, duration, dimensions and audio status using actual available tools. Do not infer content from filenames. Identify transcription tools, timestamp capability, costs and any external data transfer. Do not use paid services or upload video to third parties without my agreement.

2. Transcribe and correct factual details before selection

A transcript maps the content but does not replace source review. Check Vietnamese names, brands, figures, negations and terminology. Mark unclear speech rather than inventing a plausible sentence. Phrase-level SRT timings are not automatically accurate word-level timings.

Transcription prompt: Create a timecoded transcript from the inspected video and identify speakers only when supported; otherwise mark them unknown. Preserve meaning, flag unclear speech and list names and figures for human review. Verify existing transcripts against audio before using them. Do not build captions until I approve the text.

3. Propose eight candidates and choose five distinct clips

Do not simply extract five startling statements. Each clip should answer a specific question, preserve minimum context and offer a different idea. Candidate types include a common mistake, an explained example, a clear procedure, a story with an outcome or a qualified viewpoint. This is an editorial framework, not a platform algorithm formula.

Selection prompt: From the approved transcript, propose up to eight candidates for five 30-60-second clips. For each, provide source start/end timecodes, the question answered, main idea, proposed hook and context to retain. If only three are strong enough, report three rather than forcing five. Never invent timecodes without a timecoded source. Do not render yet.

4. Check context and write an accurate hook

A revenue-growth clip that removes rising costs can distort the message. Advice for one business model should not become a universal recommendation. Review source context before and after the selected segment, including conditions, caveats and limitations.

Context-check prompt: Compare the five selected clips with surrounding source material. Flag omissions that change meaning, context to add and clips to reject. Write two hooks each without inventing figures, results or testimonials or strengthening the speaker’s claims. Separate editorial copy from quotations; never present generated wording as a verbatim statement.

5. Test vertical framing and captions on one clip

A two-person interview may need speaker-aware cuts or split-screen rather than center cropping. Do not assume reliable automatic speaker tracking. Captions should remain readable and clear of faces and platform controls; inspect the actual publishing interface before release.

Draft prompt: Build a 1080×1920 draft of approved segment [1]. Preserve speech and context; add the approved hook, at most two caption lines, Vietnamese-capable type and restrained navy-blue-white styling. Use word-level timing only when accurate data exists; otherwise caption phrases. Export a draft, SRT and opening/middle/closing preview frames. State what has not been verified by actual playback.

6. Render five clips and deliver verifiable outputs

After approval, apply consistent typography, color and audio to the remaining clips while reviewing each separately. Name drafts clip-01-topic-draft.mp4 and approved exports clip-01-topic-final.mp4. Do not label unreviewed files final. Retain source timecodes for later verification.

Delivery prompt: Render approved clips from the template without overwriting sources. For each, provide MP4, SRT and a table with source timecodes, hook, duration, dimensions, approval status and unresolved issues. Check file existence and playability with available tools; an error-free command alone does not verify picture and sound. Isolate failed clips rather than claiming the whole batch is complete.

After publishing: measure which clips lead to further content

Track metrics actually available from the platform: watch time, completion, saves/shares and visits to full content where measurable. Compare clips with similar duration and goals; views alone are not customer acquisition. Record production time, revision rounds and cost to assess real efficiency.

Business lesson: repurposing helps a valuable idea reach the right audience in a suitable format. AI can improve productivity, while an editor remains responsible for accuracy and storytelling.

Suggested actions