generate-deliverables

Generate Deliverables

Bundle of three small generators. Each can be requested alone or as part of "all three".

Deliverable	Tool	Output
Thumbnail	ffmpeg poster frame, optional smart-frame pick	<basename>.thumb.jpg
Description	Claude reads transcript → writes summary + chapters	<basename>.description.md
Transcription	whisper backend (delegates to burn-subtitles generator)	<basename>.srt + <basename>.txt

All deliverables land in a sibling <basename>.deliverables/ folder unless the project layout dictates otherwise (see step 1).

Procedure

1. Resolve output dir

If the source lives inside a project's renders/, write deliverables to the project's exports/<basename>/. Otherwise, write to <dirname>/<basename>.deliverables/.

DATA_DIR="${CLAUDE_USER_DATA:-${XDG_DATA_HOME:-$HOME/.local/share}/claude-plugins}/video-editing"
INDEX_FILE="$DATA_DIR/index.json"
INDEX_PATH=$(jq -r '.path // empty' "$INDEX_FILE" 2>/dev/null)

REL=$(realpath --relative-to="$INDEX_PATH" "$SRC" 2>/dev/null)
if [[ "$REL" == */renders/* ]]; then
  PROJECT_DIR="${SRC%/renders/*}"
  OUT_DIR="$PROJECT_DIR/exports/$(basename "${SRC%.*}")"
else
  OUT_DIR="$(dirname "$SRC")/$(basename "${SRC%.*}").deliverables"
fi
mkdir -p "$OUT_DIR"

2. Inputs

Field	Default
Source	required (rendered video)
Deliverables	multi-select: thumbnail, description, transcription, all
Thumbnail timestamp	auto (smart pick — see below); else HH:MM:SS
Description style	youtube (default) / linkedin / plain
Language	auto-detect for transcription

3. Thumbnail

Manual timestamp:

ffmpeg -hide_banner -y -ss "$TS" -i "$SRC" -vframes 1 -q:v 2 "$OUT_DIR/$(basename "${SRC%.*}").thumb.jpg"

Auto pick — sample N evenly-spaced frames (default 12) and pick the one with the highest "interestingness" score. ffmpeg's thumbnail filter does this internally:

ffmpeg -hide_banner -y -i "$SRC" \
  -vf "thumbnail=300,scale=1280:720:force_original_aspect_ratio=decrease" \
  -frames:v 1 -q:v 2 \
  "$OUT_DIR/$(basename "${SRC%.*}").thumb.jpg"

thumbnail=300 averages over 300-frame batches and picks the most distinctive frame. Reasonable default — fast and avoids dark/transition frames.

After generating, show the thumbnail path and ask if the user wants to redo with a manual timestamp (paired with video-timeline makes this easy: list timeline frames, pick one).

4. Description

Requires a transcription. If <basename>.txt doesn't exist alongside, run the transcription deliverable first (or ask the user to skip if they don't want one).

Read the transcript, then have Claude generate:

• Title — single line, ≤70 chars.
• Summary — 2–3 sentences, no marketing fluff.
• Chapters — extracted from the transcript at natural topic breaks. Format as YouTube chapters: 00:00 Intro, 02:14 Section name, ...
• Tags — 5–10 keywords.

Write to $OUT_DIR/$(basename "${SRC%.*}").description.md:

# Title

Summary paragraph.

## Chapters
00:00 Intro
02:14 ...

## Tags
tag1, tag2, ...

For the linkedin style, drop chapters/tags and produce a 3-paragraph post with a hook in line 1. For plain, just summary + tags.

5. Transcription

Delegate to the same backend selection logic as burn-subtitles (faster-whisper or whisper.cpp from preferences.subtitle_backend). Write both .srt and .txt (plain text from concatenated cue bodies):

awk 'BEGIN{RS="\n\n"} { n=split($0,a,"\n"); for(i=3;i<=n;i++) printf "%s ", a[i]; print "" }' \
  "$OUT_DIR/$(basename "${SRC%.*}").srt" \
  | sed 's/  */ /g' \
  > "$OUT_DIR/$(basename "${SRC%.*}").txt"

If clean-transcription is wanted, suggest running it on the SRT before re-using for the description.

6. Report

Source       : path/to/in.mp4
Out dir      : path/to/in.deliverables/
Thumbnail    : in.thumb.jpg (1280×720, 142 KB)
Transcription: in.srt + in.txt (142 cues, 9.4 min)
Description  : in.description.md (212 words, 6 chapters, 8 tags)

Notes

• This skill is a coordinator — it doesn't reimplement transcription. It calls into the same backend logic as burn-subtitles.
• Description quality depends on transcript quality. For a polished YouTube description, run clean-transcription first.
• Thumbnail "interestingness" via the thumbnail filter is heuristic — for a hero-shot thumbnail, the user should pick manually from video-timeline output.