content-cascade

name	content-cascade
description	Repurpose one published video into written formats — type /content-cascade <video URL or brand channel> to turn its transcript into a blog post, LinkedIn post, and X post in the brand voice, gated through nexus-copywriter. Drafts only; publish stays founder-gated.
argument-hint	<YouTube URL, or brand name to check its channel for new uploads>
disable-model-invocation	true
allowed-tools	Read, Grep, Glob, Bash, Write, Agent, WebSearch, WebFetch, Skill

content-cascade — one video, three written formats, one voice

A published video is the estate's most expensive content unit; this skill amortises it into a blog post, a LinkedIn post, and an X post — each format-native, each in the brand's voice, each gated. Cascade direction is video→writing; it never generates video.

When to invoke

• A video just published on a brand channel (RA, CARSI, Unite, Synthex, DR/NRPG) and no written derivatives exist yet.
• Periodic check: "did publish anything new that hasn't cascaded?"
• You are about to write a from-scratch blog/social post on a topic an existing video already covers — stop, cascade the video instead.

Core procedure

1. Resolve the video. From a URL, proceed. From a brand name, check the channel for uploads newer than the last cascade (record kept in the output folder's index.md). Nothing new → report "no new uploads" and stop.
   • Done when: one target video with title, URL, publish date.
2. Fetch the transcript (yt-dlp auto-subs, or the YouTube API plane from library/connections.md). No transcript retrievable → STOP and report; never reconstruct content from the title or thumbnail.
   • Done when: full transcript text in hand.
3. Load the voice. Read the brand's persona charter from vault Personas/ and prior published examples of the SAME format. No persona charter → flag it, use nexus-copywriter defaults, and note the gap in the output.
   • Done when: voice constraints named per format (blog ≠ LinkedIn ≠ X).
4. Draft three formats through nexus-copywriter (the estate content gate — every claim sourced to a transcript timestamp or a real citation, active-voice CTAs):
   • Blog (600-1200 words): standalone value + embedded video; the transcript's argument restructured for reading, never a transcript dump.
   • LinkedIn (≤1300 chars): one insight from the video, native tone, link in comments convention.
   • X (single post or ≤5-post thread): the sharpest single claim, link on the last post.
   • Done when: three drafts pass the nexus-copywriter self-audit.
5. Package for review. Write drafts + metadata to the output folder (below) and surface for founder approval. Publishing is a separate, human-triggered act.
   • Done when: folder complete; reply names it and states PUBLISH IS NOT DONE.

Output format

~/2nd Brain/2nd Brain/Briefs/YYYY-MM-DD-cascade-<brand>-<slug>/ containing blog.md, linkedin.md, x.md, and meta.md (video URL, publish date, voice sources used, claims→ timestamp map). Update the folder's index.md line recording the cascade (this is the "last cascade" pointer step 1 reads).

Calibration

• Full run ≤ 20 minutes. Blog 600-1200 words; shorter is a LinkedIn post, longer is a rewrite nobody asked for.
• One video per invocation. A backlog of uncascaded videos = run it N times, newest first.

What this skill is NOT

• Not nexus-viral — that goes idea→video→8 video cuts; this goes published-video→writing.
• Not brand-video / video-director — no video is generated here, ever.
• Not marketing-social-content — that writes social from a campaign brief; this derives strictly from an existing video's transcript.
• Not a publisher — no posting, scheduling, or upload. Founder-gated, always.

Hard rules

1. Words ship through nexus-copywriter then brand-guardian — the cascade never bypasses the estate content gate.
2. Never fabricate from title/thumbnail when the transcript fails — fabricated derivatives are the worst-case violation of the no-false-recordings directive.
3. Drafts only. The reply must state explicitly that nothing was published.
4. Voice is per-brand AND per-format; one generic rewrite×3 is the observed naive failure. Calibrate voice the destruction-cycle way: human eviscerates drafts, skill notes update — there is no one-shot self-improving voice (source's explicit warning).

Provenance

• Source: Chase AI personal-assistant system — https://www.youtube.com/watch?v=gUv7VqcRzok
• Vault: Wiki/claude-personal-assistant-system-2026-07-15-ingest.md (cascade design, voice destruction-cycle method).