dena-video-editing-workflow

name	dena-video-editing-workflow
description	Use when planning, editing, assembling, reviewing, or rendering Dena Meidina social videos in this repository, especially raw videos, reference-video adaptation, captions, overlays, HyperFrames compositions, QA, render readiness, or platform publishing checks.

Dena Video Editing Workflow

Overview

Use this skill as the router for Dena Meidina social-video work in this project. It enforces the docs/agents/ sequence, Dena style guide, HyperFrames composition rules, handoff artifacts, and optional QA review gate.

Start Here

Before any Dena video task:

1. Read AGENTS.md.
2. Read docs/dena-social-video-style-guide.md.
3. Identify the current workflow stage.
4. Read the matching agent file in docs/agents/ completely.
5. Read required upstream artifacts listed by that agent.
6. Produce the expected handoff files in videos/<slug>/.

For the full chain and routing table, read references/agent-chain.md.

For validation, QA, and render gates, read references/quality-gates.md.

Stage Router

Use exactly the agent that owns the current decision:

User need	Agent
New raw/reference video, angle, hook, format, style adaptation	docs/agents/01-creative-director.md
Transcript, silence/filler cuts, pacing, processed base video	docs/agents/02-transcript-cut-agent.md
Captions, subtitles, hook text, phrase grouping, ASR correction	docs/agents/03-caption-subtitle-agent.md
Screenshots, b-roll, generated images/video, diagrams, stickers	docs/agents/04-asset-generation-agent.md
Overlay timing, pattern interrupts, zooms, effects, transitions	docs/agents/05-motion-overlay-agent.md
index.html, compositions/*.html, timed clips, GSAP, HyperFrames assembly	docs/agents/06-hyperframes-assembly-agent.md
Optional QA, punch list, render/platform readiness review, regression review	docs/agents/07-qa-review-agent.md
R2/Repliz auto publish after explicit user approval	docs/repliz/integration-spec.md

Do not skip ahead unless the user explicitly requests a narrow technical fix and upstream decisions already exist.

Lean Fixing Defaults

Use this when fixing a previous Dena edit, or when the user says the result was over-cut, captions drifted, or the process burned too much time/token.

• Do not rerun the full agent chain by default. Reuse existing artifacts, route only the broken part, and update downstream files that depend on it.
• Treat "boring motion", "too much transcript was cut", "needs non-slop assets", and "no sound effects" as one targeted revision path: keep the compact cut as the baseline when it feels denser, Agent 02 restores only missing context, Agent 04 creates local/manual explanatory assets when useful, Agent 05 adds purposeful motion plus SFX timing, Agent 06 assembles, and Agent 07 re-verifies.
• Treat "captions skipped words" as a Caption Agent bug by default. Regenerate or realign from the locked processed word-level transcript so every surviving spoken word is represented by active running captions.
• Default recut intensity is light-medium: remove long silence, heavy filler, and clear repetition only. Do not force a viral hard-cut unless the user asks for aggressive pacing.
• Lock the base cut before final captions. If processed.mp4 changes, regenerate or realign caption-beats.json from the processed video instead of manually nudging old raw-timeline captions.
• If captions, assets, motion, or SFX feel out of sync after a compact recut, transcribe the locked processed audio and retime caption-beats.json, overlay-timeline.json, HTML clips, and SFX cues from the same cue map.
• If the user wants source code/editor footage visible as proof, keep overlays as compact callouts in unused top/side space. Do not use large privacy/context masks that hide the proof moment.
• If the user provides a link or the transcript names a tool/product/site, route to Agent 04 for web research/inspection and local screenshot or screen-record capture before motion planning.
• If real captures do not explain the point well enough, use Codex/image generation for grounded bitmap stills or short support visuals instead of stiff filler cards.
• Preview with frame grabs/contact sheets first. Run full render only after cut, captions, and overlays look correct, then rerender only for blocker/major fixes.
• For Dena raw talking-head clips, preserve natural context without killing density. A compact 55-70s cut can beat a fuller 80-100s cut when the lesson still lands.
• If the user says SFX has no sound, verify both the solo SFX stem and the final render audio with a level scan. File existence is not enough; cues that peak too low are effectively missing.
• For transcript quality, use the project-local Whisper install in vendor/whisper.cpp with ggml-large-v3-turbo.bin; convert source audio to WAV first and pass an Indonesian/domain-term prompt.
• If a longer HyperFrames render stalls in default low-memory mode after passing check/snapshot review, retry the final render with PRODUCER_LOW_MEMORY_MODE=false and document the workaround in render-review.md.
• When a session produces a workflow learning, update the relevant docs in the same turn; the project-local .codex Stop hook enforces this for learning prompts and workflow/config changes.
• If the user asks whether session learnings are already documented, treat that as a docs reconciliation task: verify the covered learnings and add or refine a concise docs note in the same turn before answering.

Non-Negotiables

• Use Indonesian by default and preserve Dena's natural register.
• Treat reference videos as ingredients, not costumes.
• Keep Dena positioned as a credible AI systems builder, senior developer, founder/operator.
• Default Dena storytelling/talking-head videos require running captions that cover every spoken word surviving the cut, not only highlight phrases.
• Default processed speed is 1.2x; any lower speed needs a documented clarity/emotion exception.
• Keep captions readable on phone: short phrases, white/black base, selective yellow emphasis.
• Keep overlays purposeful: clarify, prove, reset attention, or transition. URL/tool/product mentions need researched/captured or generated context assets when they help viewer understanding.
• Motion plans for designed recuts must include purposeful, audible-but-under-speech SFX cues when transitions, proof reveals, or title hits need impact.
• CTA must be non-promissory unless the user explicitly approves a promise.
• Keep every layer editable until final render.
• For HyperFrames work, read /hyperframes and the routed HyperFrames skill before editing .html.
• After any .html edit, run npm run check and fix errors before handoff.
• After final render, stop for user review. Offer publish as-is, QA first, or revisions. Do not upload to R2 or schedule Repliz until the user explicitly approves.
• Repliz publish must use --approved; R2 uses Wrangler with CLOUDFLARE_ACCOUNT_ID, bucket from R2_BUCKET, and https://<r2-public-domain>.

Handoff Contract

For a complete edit, the expected artifact chain is:

videos/<slug>/creative-brief.md
videos/<slug>/edit-decision-notes.md
videos/<slug>/caption-plan.md
videos/<slug>/caption-beats.json
videos/<slug>/publish-captions.md
videos/<slug>/asset-plan.md
videos/<slug>/asset-manifest.json
videos/<slug>/motion-plan.md
videos/<slug>/overlay-timeline.json
videos/<slug>/assembly-notes.md
videos/<slug>/assembly-checklist.md

Only create asset artifacts when assets are needed. Only create QA artifacts when the user chooses QA first or explicitly asks for QA. Only create final-approval.md after QA passes.

If Inputs Are Missing

Do not invent missing upstream decisions.

• Missing direction: run Agent 01.
• Missing transcript/cut: run Agent 02.
• Missing caption timing: run Agent 03.
• Missing assets: run Agent 04 or mark assets unnecessary.
• Missing motion plan: run Agent 05.
• Missing assembly notes/checklist: run Agent 06.
• User chose QA first and QA evidence is missing: run Agent 07.

If the user asks for a narrow fix, document which upstream assumptions are being reused.