1 asset -> 20 derivatives

The math of repurposing

One real recording of yours goes in; 15+ reviewed derivatives come out, sitting in approved/ with publish dates. Here's the math that makes that hour worth it. Your hardest content to produce is the long stuff: the podcast episode, the webinar, the deep-dive post. Your highest-volume need is the short stuff: daily social, newsletter sections, clip scripts. Repurposing converts one into the other - one 40-minute recording becomes a LinkedIn series, a batch of X posts, a newsletter section, a blog post, and five clip scripts with timestamps. Twenty-plus assets from one recording session.

Done by hand, that conversion costs more than the original recording, which is why most teams never do it. Done as a skill the team runs on every new recording - not a prompt someone reconstructs each time - it's an hour of review. Anthropic's marketing team reports influencer-script workflows alone freeing 100+ hours a month; repurposing is the same shape of win.

Transcripts in: the ingestion step

Everything starts from a clean transcript, and you almost certainly already have one. Every recording surface a team uses - the meeting recorder, the podcast host, the webinar platform, the video editor - exports a transcript or at least captions. The pipeline's input is that exported file, dropped into the asset folder. Claude's work starts at the file; no integration required.

1. Create derivatives/<date>_<asset-slug>/ for the source asset - everything derived from it lives in this one folder.
2. Export the transcript from wherever the recording lives and save it as transcript-raw.md in the asset folder. Raw captions work too - paste them in; cleanup is the next step's job.
3. Have Claude clean it into transcript.md: fix speaker labels, mark timestamps every minute or so, strip filler. Keep the raw version untouched - the fact-checker string-matches quotes against it later.

The repurpose skill: five stages

The pipeline is: record -> transcribe and clean -> extract moments -> generate written assets -> distribute. Stages one and two you just built. The repurpose skill handles three and four; distribution comes in module 2. Repurposing skills are common in the open catalogs - install the library's version if one fits, then adapt it to your voice file and the quote-check rule below.

1. Extract: Claude reads the clean transcript and proposes 3-5 standalone moments, each with a timestamp, a one-line 'why this lands', and the verbatim passage it comes from.
2. You pick. Thirty seconds of human judgment here beats any amount of generation downstream - kill the weak moments before they spawn derivatives.
3. Generate: for each chosen moment, produce the derivative set - each one as its own file in the asset folder, drafted against voice.md and the channel formats below.
4. Quote check: every direct quote in every derivative gets string-matched against the raw transcript. People remember what they said on their own podcast.
5. Encode it all as .claude/skills/repurpose/SKILL.md taking the asset folder as its argument.

derivatives/2026-06-02_podcast-ep14/
├── transcript.md            # cleaned, timestamped
├── transcript-raw.md        # untouched, for quote verification
├── moments.md               # the 3-5 extracted moments + your picks
├── linkedin/                # one post per file, 3-5 posts
├── x/                       # short posts, threads
├── newsletter/              # one section for this week's issue
├── blog/                    # one long-form draft if the episode earns it
└── clips/                   # 5 clip scripts with start/end timestamps

One folder per source asset keeps provenance obvious: six months later you can see exactly what came from where, which derivatives shipped, and which formats actually performed for that kind of source.

Per-channel formatting that respects the channel

A derivative that ignores its channel's format is just cross-posting, and audiences smell it. Encode the format rules per channel in the repurpose skill, alongside the voice file.

• LinkedIn: a hook line that survives the 'see more' fold, short paragraphs, one idea per post, no link in the body of the first draft (decide link strategy at review).
• X: a different room, not a shorter LinkedIn. One sharp line or one long post; threads only when a moment genuinely has 3+ beats. The register is more personal and more openly proud of results: show receipts (numbers, what you built, a screenshot) rather than lessons. It gets its own voice file (brand/voice-x.md) learned from real top X posts in your niche, never from your LinkedIn corpus.
• Newsletter section: assumes warm readers - more context, a personal aside, one clear link out.
• Blog: only when a moment can carry 800+ words without padding. Run it through the full write-post pipeline instead - it deserves the fact-checker.
• Clip scripts: verbatim transcript excerpts with start/end timestamps and a suggested on-screen hook line - ready to hand to whoever edits video.

The same rule holds when you cross-post. Our own engine sends every LinkedIn post to X as well, but never as pasted text: a separate content-x skill rewrites it against an X voice file built from about 1,200 real top-performing X posts in our niche. The rewrite keeps the substance and changes the shape, the hook, and the register. If the LinkedIn version and the X version read the same, one of them is wrong for its room.

Batch review and spacing

Twenty derivatives reviewed one by one across a week is a chore that dies quietly. Reviewed in one 30-40 minute batch, it's a routine that survives. Open the asset folder, read everything, edit fast, delete the duds without guilt, and move survivors to content/approved/.

Then space the survivors over two weeks rather than flooding three days. Spacing is a scheduling problem, and module 2 turns it into a queue of dated files. For now, note the intended publish date in each derivative's filename or frontmatter.

Do this now

Sources and further reading

• How Anthropic teams use Claude Code
• Claude Code skills docs