mting314/subtitling-projects

AfterTalk launch (positioning → hardsub → publish)

The finishing pass on a _translated.ass, after the four-dimension QC text review (that's in subtitle_review_guide.md). Four steps, each backed by a deterministic script in scripts/ under the projects repo root so this skill only supplies judgment. Run everything from the projects repo root (paths below are scripts/…).

1. Positional styles — reposition subs so they clear on-screen content.
2. Line length — flag + fix any line rendering to 3+ rows (do after positional, since PiP/Side-Song lines wrap narrower and are the main offenders).
3. Hardsub segments — find the host-talk ranges to keep and burn them in.
4. YouTube title + description — write the release copy.

---

1. Positional styles (PiP / song-shift)

Repositions subtitles so they never sit on top of on-screen content. The pipeline pre-assigns the style names to the right line ranges but usually ships them as plain copies of the main style (wrong margins) with no \pos on PiP lines. The script fixes both, copying layout from a known-good reference; this skill resolves the per-episode style names to roles (they differ per character).

Role	When	Reference style (Colors of Pure Sense)	Gets \pos?
pip	host cam / card art fills the frame; subs sit mid-frame	PiP (Alignment 8, margins 100/800/50)	yes — {\pos(650,750)} per line
shift	a 2D/3D MV plays lower-right; shift subs left to clear it	DefaultOnibe - Shifted (Alignment 2, margins 100/730/60)	no — style margins do it

Discover the target's positional styles, then infer role from name tokens (PiP→pip; Shifted/Side Song/Shift→shift) and confirm with the user:

f="projects/Project Sekai/Aftertalk/<event>/<name>_translated.ass"
grep -oE "^Dialogue: [0-9]+,[^,]+,[^,]+,[^,]+" "$f" | awk -F, '{print $4}' | sort | uniq -c
grep "^Style:" "$f"

Apply (copies only Alignment + MarginL/R/V — color/font preserved; skips Comment: and already-positioned lines; idempotent, --dry-run):

python3 scripts/apply_positional_styles.py "projects/Project Sekai/Aftertalk/<event>/<name>_translated.ass" \
  --reference "projects/Project Sekai/Events/Colors of Pure Sense/Colors of Pure Sense_translated.ass" \
  --pip   "<Char> - PiP:PiP:650,750" \
  --shift "<Char> - Side Song:DefaultOnibe - Shifted" \
  --dry-run

Mapping: TARGET_STYLE:REF_STYLE[:X,Y]; --pip/--shift repeatable per character. Confirm the target and reference share the same PlayResX/Y.

Verify by rendering (defer to the aegisub-ass skill): retime one pip and one shift line to 0:00:00.00–0:00:05.00, render a frame over flat gray, and confirm the pip text sits mid-frame and the shift text clears the lower-right.

---

2. Line length (max 2 rows on screen)

3+ rows is bad practice (eye-sweep, screen real estate). Row count depends on libass's actual wrapping under each line's style, so measure by rendering. Run after positional (PiP wraps ~1020px, Side-Song ~1090px vs main ~1320px — the narrow ones are the offenders) and after any text edits (a reword changes wrapping).

Detect (renders each line under its style over black, counts rows by projection):

uv run --with pillow --with numpy python3 scripts/detect_long_lines.py \
  "projects/Project Sekai/Aftertalk/<event>/<name>_translated.ass"

For each flagged line, offer the user both fixes and apply their choice:

• Split into two events at a clause boundary (comma, and/but/so, that, sentence end) with scripts/split_subtitle_line.py --line N --before "<clause>" --transcript <t.json>. It sets the time split on a real breath: proportional-by-length estimate, then snaps to the largest word-gap within ±1s of it (rejecting distant mid-sentence pauses and any snap that leaves a half too short to read). Preserves exact wording + \pos.
• Reword shorter, preserving meaning + casual style.

Prefer split for two real clauses/sentences or a natural ... pause; reword for a single breath. Re-render every candidate fix and confirm ≤2 rows before presenting/applying, then re-run the detector — target is 0 flagged. (Requires pillow + numpy via uv run --with.)

---

3. Hardsub segments (find keep-ranges → burn)

An AfterTalk interleaves host talk (keep) with watchalongs (cut): intro delay, an opening PV/digest, the in-game story watchalong, and the song/2DMV. Cut regions are non-rendering Comment: lines (or nothing), so they appear as gaps in the rendered Dialogue timeline. find_segments.py finds them and, with --transcript, reports whether each gap is silent (safe) or watchalong audio.

python3 scripts/find_segments.py "projects/Project Sekai/Aftertalk/<event>/<name>_translated.ass" \
  --transcript "projects/Project Sekai/Aftertalk/<event>/<name>_transcript.json" \
  --hardsub "projects/Project Sekai/Aftertalk/<event>/<name>.mkv:projects/Project Sekai/Aftertalk/<event>/<name>_final.mp4"

Read the CUT regions: ~0 words = silent (safe to drop); many words = a story/song watchalong you're intentionally cutting (its own audio, not host talk). Confirm the cut plan with the user before rendering — cutting is an editorial call, and the render is expensive. Record the ranges in the project's notes.md "Segments".

Fade-aware boundaries: passing --mkv (or --hardsub, which reuses its MKV) makes it run blackdetect near each boundary and pull the cut just clear of any source fade. Watchalongs/MVs fade to/from black at their edges; the pad can push a cut into a fade, and two cut-adjacent half-fades meeting at a concat join produce a black flash. It emits sub-second boundaries — no subtitles lost (the pad region is sub-free). The script reports e.g. end 0:51:00->0:50:59.69 (source fade).

Then hardsub (the --hardsub flag prints this command; must hardsub before trimming — trimming invalidates .ass timestamps):

uv run python scripts/hardsub_trim.py "<name>.mkv" "<name>_translated.ass" "<name>_hardsubbed.mp4" \
  0:01:37 0:05:27  0:10:53 0:27:40  0:36:05 0:51:00  0:52:25 1:00:09

Handles parallel GPU encoding (h264_nvenc), automatic popups.json VA/gachapin overlay card compositing, smooth 0.4s video + audio fade-in/fade-out transitions at segment boundaries, and concatenation into the final video.

---

4. YouTube title + description

Write these into the project's notes.md (## YouTube Title, ## YouTube Blurb).

Title — [ENG SUB] <Event Title> Aftertalk feat. <VA Name> (<Character>'s VA) e.g. [ENG SUB] Gazing Upon the Night Sky's Fading Stars Aftertalk feat. Karin Isobe (Saki's VA).

Blurb — two paragraphs, casual and enthusiastic, recap voice (fans who want a recap, not a teaser). Style rules:

1. No em dashes (—). Use commas, parentheses, or separate clauses. Scan the draft and replace any — before saving.
2. Open with the series framing: This episode of ProSeka AfterTalk has <VA name> (voice of <character>) covering <event/topic>.... It's a recurring series, so "this episode of," not "in the AfterTalk."
3. Weave personality into substance. The spine is the substantive coverage (story, card art, song, recording); pair each topic with a personality detail from that same topic (e.g. discussing the song and fumbling "2DMV"). Don't ghetto the silly bits into their own paragraph.
4. No hook sentences ("Bonus:", "Don't miss...", "Plus:"). Keep the descriptive recap voice throughout.

End with Original video: <url>.

Structure to mirror: paragraph 1 identifies VA + event + framing (solo/duo host, what she watches); paragraph 2 covers story/recording, card illustrations, and the song/MV, with personality woven in. See Grow Glorious Glow Aftertalk/notes.md or Gazing Upon the Night Sky's Fading Stars Aftertalk/notes.md for worked examples.

---

5. YouTube upload

After hardsubbing and writing notes.md, run scripts/youtube_upload.py to upload the finished video:

uv run --script scripts/youtube_upload.py \
  --video "projects/Project Sekai/Aftertalk/<event>/<name>_hardsubbed.mp4" \
  --notes "projects/Project Sekai/Aftertalk/<event>/notes.md"

The script automatically extracts the title and description from notes.md, verifies lengths, requests confirmation, and uploads via the YouTube Data API v3 (resumable upload).

---

Guardrails

• Positional: only Alignment + MarginL/R/V change; never recolor; never touch Comment: lines or overwrite an existing \pos.
• Segments: cutting is editorial — confirm before the (expensive) render; log the ranges to notes.md; the transcript word-count tells you why a gap is a cut.
• The scripts error out on unknown style/paths — fix the args rather than forcing.