SCROLLUCK/dark-channel-studio-public

shorts-from-peaks

Tool order

1. search_videos(query, n, shorts) - searches YouTube, ranks results (views, virality) and stores them with their top peaks. list_sources shows what is already stored. Prefer official or permitted sources.
2. get_peaks(video_id) - heatmap peaks of a stored source. No heatmap = say so; propose another source or a manual window, never invent peaks.
3. create_clip_project(video_id, peak_rank=1, hook_text) - editor draft from a peak, or pass start/end seconds for a manual window (runs directly).
4. render_project(video_id, hook_text, custom_script) - renders the draft in the background (runs directly); poll get_job(job_id). Its checks_failed lists failing transformation checks.
5. Fix failing checks, then ask to add it to the queue with queue_add(project_id) (needs confirmation). ui_open("editor", video_id) shows it to the user.

• No heatmap, or a long upload/link the user owns: autoclip_source(url, length, max_clips, prompt, create_drafts) finds the best moments from the transcript and returns ranked clips (poll get_job); with create_drafts=true they land in the editor.

Reading peaks

• Values are normalized 0-1. A peak is a local maximum >= 0.8; peaks closer than 10 s merge into one.
• Ignore a spike in the first 5% of the video (intro artifact). Keep the top 3 by value.
• Report each peak as start-end, peak_value, rank from the tool, with the video title and URL.

Cut window

• Total length 15-30 s (hard max 30 s: the transformation gate rejects longer).
• Start 3-5 s before the peak so the viewer gets the setup; end 2-4 s after the payoff. Cut at a sentence or action boundary.
• Clamp to the video duration. Never use more than 50% of the runtime from one third-party clip: add B-roll, images or data cards.
• One moment per Short. Two peaks only if they are the same story beat.

Hook (first 3 seconds)

• A question, a stake or a surprising claim read over the first frame, plus a big on-screen headline (max 6 words).
• Use only facts present in the title, description or transcript; label rumors as "RUMOR".
• Narration: sentences <= 8 words, about 2.5 words/s, hook -> context -> payoff at the peak -> short CTA. The source's own audio stays low or off; the narrator leads.

Transformation checks (all five must pass; see docs/LEGAL.md)

• narration: original narration present.
• visual_layers: at least 2 added layers (text, images/stickers, B-roll, data card).
• text_overlay: hook or text on screen.
• duration: 30 s or less.
• source_credit: source credited in every platform caption. If one fails, fix the project and re-render; re-rendering invalidates any earlier approval.

Captions and text

• Word-by-word captions, 2-4 words per group, bold sans font, white with black outline, active word highlighted.
• Keep captions out of the bottom ~12% and sides ~6% (platform UI). Emojis only if the channel style uses them.

Title, caption, hashtags

• Title/first line = the hook, under 80 characters. One sentence of context.
• 3-5 hashtags: 1 niche, 1-2 topic, 1 broad. No banned or misleading tags.
• Always include Source: <channel> - <URL>.
• Template: references/clip-brief.md.

AI disclosure

Synthetic narration (TTS) or AI visuals require the per-post AI-disclosure flag before approval, plus the platform's own AI label on upload. Remind the user in the confirm card.

Do not

• Post a raw clip, a clip over 30 s, or a clip without narration.
• Use a source the user marked blocked or that is private/members-only/sign-in-only.
• Add to the queue, schedule or delete without the user's confirmation.