Imported from ivan-magda/blog (.agents/skills/processing-wwdc-session/SKILL.md). Install upstream with npx skills add ivan-magda/blog --skill processing-wwdc-session. Copyright stays with the author.

Processing a WWDC Session

Overview

Turns an Apple Developer video URL into a research folder under the gitignored workspace/ so a later session can draft blog posts from one synthesized file (digest.md). Two phases: A. Fetch & build runs the moment a URL is dropped. B. Screenshots runs later, when the user adds screenshots/ — slides are not available at fetch time, so the digest is written first and enriched afterward.

workspace/ is gitignored: this produces zero shippable changes. Do NOT touch src/data/blog/, public/llms.txt, src/assets/, or any SEO file — those are for published posts only.

Model policy — the digest must be Opus

Fetching, scraping, and file scaffolding (Phase A steps 1–5, Phase B steps 1–2) are mechanical and fine on any model — run the whole skill on Sonnet if you like. digest.md is the one high-judgment synthesis step and MUST be produced on Opus.

• If you (the orchestrator) are not running on Opus, spawn a subagent with the Opus model to do it (Agent/Task tool, model: opus). Give it the session dir; it reads transcript.md, meta.md, code.md (plus screenshots/ in Phase B) and writes/updates digest.md.
• If you are already on Opus, write/update the digest directly.
• Never create or update the digest on a smaller model.

This applies to BOTH the initial write (Phase A step 6) and the screenshot enrichment (Phase B step 3).

Phase A — Fetch & build the entry

Trigger: a developer.apple.com/videos/play/wwdcYYYY/NNN URL.

1. Derive paths. Collection dir is workspace/wwdc26/ (this blog's WWDC 2026 set). Session number NNN from the URL; slug = short kebab of the title (e.g. 241-foundation-models). Create workspace/wwdc26/{NNN}-{slug}/.
2. Transcript. ToolSearch → select:mcp__sosumi__fetchAppleVideoTranscript, call it with path: /videos/play/wwdcYYYY/NNN. Write transcript.md with YAML frontmatter (title, source, session, collection, duration, fetched, via: sosumi.ai) — raw Apple words, attribution intact.
3. Summary + Code tabs. These are JS-rendered tabs, not in the transcript and not reachable by WebFetch. Use Playwright: ToolSearch → select:mcp__plugin_playwright_playwright__browser_navigate,mcp__plugin_playwright_playwright__browser_evaluate. browser_navigate to the URL, then ONE browser_evaluate reading both (all tabs live in the DOM — no clicking):
   • Summary: document.querySelector('.supplement.summary')?.innerText
   • Code: Array.from(document.querySelectorAll('.supplement.sample-code pre')).map(p => p.innerText)
4. Write meta.md — description, key topics, chapter summary (from the Summary tab), a "Related sessions to fetch" checklist (talks named in this one), pointer to code.md. Write code.md — each snippet under a ## HH:MM — label heading in a fenced block. Clean any // Copy Code / // Insert code snippet placeholder text that bleeds into the extraction.
5. Write notes.md — a stub for the user's own analysis (blog angles, open questions, code to reproduce). Keep separate from transcript.md so drafts pull from the user's words, not Apple's.
6. Write digest.md (Opus only — see Model policy) — comprehensive single-file synthesis: frontmatter listing sources, TL;DR, sections following the chapters, all code blocks, an Open Questions section, and blog angles. Leave [shot: HH.MM.SS] references out until Phase B (or add them only if screenshots already exist). If you are not on Opus, spawn an Opus subagent to read steps 1–5's files and write this.
7. Update workspace/wwdc26/README.md — add the session's tracker row.

If the target folder already exists, confirm before overwriting.

Phase B — Screenshots post-completion step

Trigger: a screenshots/ folder appears in an existing session dir (the user took slide captures and dropped them in).

1. Compress in place — run, from the repo root: .agents/skills/processing-wwdc-session/compress-screenshots.sh <session-dir>/screenshots (the script is bundled with this skill). It downscales + palette-quantizes each PNG to ≤200 KB (band 100–200 KB), preserving PNG format. Keep the original Screenshot … HH.MM.SS.png filenames — do NOT rename, do NOT make .webp siblings, do NOT move into src/assets.
2. Read every screenshot (the Read tool renders images). Extract what the transcript and Code tab do NOT capture: diagram structure, on-screen API values, demo token counts/durations, the CLI surface, and any code visible on a slide that is missing from code.md.
3. Fold back into digest.md (Opus only — see Model policy) — add [shot: HH.MM.SS] references at the relevant points (the HH.MM.SS is the time in the screenshot's filename, used as a stable handle to the file — it is NOT the video timecode). Add genuinely new facts, and flag discrepancies where a slide contradicts the transcript/Code tab. Add any slide-only code to code.md with a // from screenshots/… note. Update the README row if its screenshot/digest columns change. If you are not on Opus, spawn an Opus subagent that reads the screenshots and rewrites the affected digest sections.

Group labs — transcribe from the SD video audio

Group labs (the 8xxx series) ship no sosumi transcript and no Summary/Code tab, but the session page does publish a downloadable recording. Generate the transcript locally from the video audio, then run the normal Opus digest step. This replaces the old "write a _No transcript available. stub" for any lab whose video has posted.

Tooling (already installed): whisper-cli (Homebrew), model /Users/jetbrains/Developer/whisper.cpp/models/ggml-large-v3.bin, VAD model /Users/jetbrains/Developer/whisper.cpp/models/for-tests-silero-v6.2.0-ggml.bin, ffmpeg, curl. Helper scripts in workspace/wwdc26/_tools/: process_lab.sh (chains the whole media pipeline), whisper_to_transcript.py, check_transcript.py.

Per lab:

1. Get the SD URL. WebFetch the session page (/videos/play/wwdc2026/NNNN/) and read the SD Video link from Resources — it is in static HTML, so WebFetch works and you can probe many labs in parallel (no Playwright needed). The CDN URL carries a per-session UUID, so you must fetch it per lab; you cannot construct it. "Available soon" / "Live …" pages have no download link yet — skip and revisit later.
2. Download + extract. curl -sL the SD mp4 (~240 MB), then ffmpeg -vn -ar 16000 -ac 1 -c:a pcm_s16le → 16 kHz mono WAV.
3. Transcribe — VAD + -mc 0 are MANDATORY, not optional:

   whisper-cli -m <large-v3> -f NNNN.wav -l en \
     --vad --vad-model <silero> -mc 0 -et 2.8 -oj -of NNNN -t 8
   

   Plain large-v3 (no VAD, default context) falls into a repetition hallucination loop on lab audio — in testing it emitted one phrase for 47 of 63 minutes (75 % of the session), and the output looks fine segment-by-segment. VAD skips the non-speech that triggers the loop; -mc 0 stops the decoder conditioning on its own repeats. VAD is also faster (skips silence).
4. GATE: check for loops. python3 _tools/check_transcript.py NNNN.json exits 1 if any phrase runs ≥8 consecutive segments or is ≥15 % of all segments. Never build a digest off a transcript that fails the gate. On failure, re-run whisper / inspect the audio — do not proceed.
5. Convert to Zoom-style transcript. `python3 _tools/whisper_to_transcript.py NNNN.json
6. Delete the mp4 + wav (~240/116 MB each; scratch).
7. meta.md + digest.md (Opus — see Model policy). Same as Phase A steps 4–6, but the digest must capture a live Q&A: panel roster, every question→answer exchange, and an "unconventional facts & takeaways" section for the off-the-cuff details that only surface in Q&A (corrected misconceptions, hard numbers, candid limitations). code.md stays n/a — labs show no code.
8. README row: Transcript = 🎙️ (machine transcript), Digest = ✅.

process_lab.sh NUM SLUG "TITLE" SD_URL does steps 2–6 (download → extract → whisper → gate → convert → cleanup) and exits non-zero if the gate fails.

Scaling many labs. whisper saturates the GPU, so run it serial — one lab at a time (loop process_lab.sh over the lab list in a background batch). Parallelize only the cheap parts: probe SD URLs with parallel WebFetch up front, and dispatch the Opus digest subagents in waves (each reads its own transcript.md, writes meta+digest) while the next whisper runs. Never run two whisper processes at once. A lab is ~6–9 min of whisper (faster with VAD) + ~1 min download.

Bulk mode — many sessions at once (parallel subagents)

When a whole catalog is dropped (dozens of URLs), do NOT run Phase A serially per session. Split the work so the one shared resource (the Playwright browser) never races and the 20–60 KB transcripts stay OUT of the orchestrator's context:

1. Build a manifest first. Parse the link list into {num, title, url, folder} (clean, stopword-stripped slugs) and create all folders. Always dispatch using the manifest's folder names — hand-typed slugs drift from the created dirs and make the browser_evaluate filename write fail with ENOENT.
2. One fetcher subagent at a time (serial, OWNS the browser): for each session browser_navigate + ONE browser_evaluate that saves Summary/Code to <folder>/_supplement.json via the filename argument (keeps the payload out of your context). Batches of ~16. Never run two fetchers at once — the single browser instance will race and cross-contaminate pages.
3. Parallel synthesis subagents (Opus — see Model policy) — one per session, waves of ~8. Each fetches its OWN sosumi transcript, reads _supplement.json, and writes all five files. They use sosumi + file tools ONLY (no browser), so a synthesis wave can run concurrently with the NEXT fetcher batch running in the background.
4. Normalize after each fetch batch. browser_evaluate saves its result double-JSON-encoded (a JSON string wrapping the object) — decode twice (json.loads(json.loads(...))) once so synthesis reads clean JSON.
5. Sync trackers from disk, don't hand-maintain. A script that scans folders (digest present? supplement summary/code non-empty?) idempotently regenerates the README table and the link-list ✅ markers — survives context summarization.
6. Labs / keynotes have no sosumi transcript. Group labs return "Transcript not found" from sosumi and have no Summary/Code tab — if the lab's video has posted, transcribe it from the SD audio (see Group labs — transcribe from the SD video audio above), not a _No transcript available._ stub. Only stub a lab whose video is still "Available soon." Some keynotes genuinely have no transcript — stub those.

Give each synthesis subagent a one-line dispatch pointing at a shared instruction file plus NUM/TITLE/URL/FOLDER, so prompts stay tiny and consistent across the fleet. Overlap rule: synthesis (sosumi) + the next fetcher (browser) may run concurrently; two browser users may not.

Quick reference

Per-session files: transcript.md, meta.md, code.md, notes.md, digest.md, screenshots/.

Common mistakes

• Producing or updating the digest on a smaller model — the digest is the deliverable and the only real synthesis step; it must be Opus. Steps 1–5 (and screenshot compression) can be any model. If you're on Sonnet, spawn an Opus subagent for the digest — don't write it yourself.
• Skipping screenshot compression — raw Retina captures are ~2.5 MB each. Always run the script; preserve PNG.
• Renaming screenshots — the [shot: HH.MM.SS] handle is the filename's time. Renaming breaks every reference. Keep the originals.
• Not updating the digest after reading screenshots — Phase B exists to enrich the digest; reading the images without folding new facts (and discrepancies) back in defeats the point.
• Treating [shot:] as a video timecode — it is the screenshot filename's clock time, just a file handle.
• Touching shippable files — workspace/ is gitignored scratch. No llms.txt, src/data/blog/, or src/assets/ edits when processing a session.
• Trusting WebFetch for the Summary/Code tabs — they are JS-rendered; use Playwright. (The download/SD links are static HTML, so WebFetch is fine for those.)
• (Group lab) Transcribing without VAD + -mc 0 — plain large-v3 hallucination-loops on lab audio (one phrase for most of the session) and looks fine per-segment. Always pass --vad --vad-model <silero> -mc 0, then run the check_transcript.py gate before any digest.
• (Group lab) Inventing speaker labels — large-v3 has no diarization. Keep cues unattributed; capture the panel roster from the intro.
• (Group lab) Running two whisper processes at once — whisper saturates the GPU; concurrency just contends. Serial whisper; parallelize SD-URL probes and Opus digests only.
• Leaving // Copy Code placeholder text in code.md from the pre extraction.
• (Bulk) Running two browser fetchers at once — the Playwright instance is shared and single; concurrent navigations race. Exactly one fetcher; parallelize only the sosumi synthesis.
• (Bulk) Hand-typing folder slugs per dispatch — they drift from the created dirs and the filename write fails (ENOENT), silently dropping that supplement. Always read folders from the manifest.
• (Bulk) Forgetting the double-encode — browser_evaluate's saved _supplement.json is a JSON string wrapping the object; parse twice or synthesis sees a string, not {summary, code}.

Verification

• Folder has transcript.md, meta.md, code.md, notes.md, digest.md.
• Every screenshot is valid PNG and ≤200 KB (find screenshots -name '*.png' -size +204800c returns nothing).
• Screenshot filenames unchanged; each [shot:] in digest.md resolves to a real file.
• digest.md reflects slide-only facts and flags any slide↔transcript discrepancies.
• README.md tracker row is current.