CtrlK
BlogDocsLog inGet started
Tessl Logo
PluginPublic

jbaruch/speaker-toolkit

Twelve skills for conference speakers covering rhetoric analysis, talk authoring, deck building and export, CFP submissions, publishing, illustrations, shownotes, screencast production, and vertical shorts, informed by a 113-entry Presentation Patterns taxonomy (83 observable: 64 patterns + 19 antipatterns; 30 unobservable: 21 patterns + 9 antipatterns).

Quality

91%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide
SecuritybySnyk

Low

Low-risk findings worth noting

Overview
Quality
Evals
Security
Files

SKILL.mdskills/shorts-producer/

name:
shorts-producer
description:
Turn a processed talk into vertical shorts: pick its most shareable takes from the finished vault analysis, then render each as a 1080x1920 video with the hook headline, the deck slide timed to the talk, a face-tracked crop of the speaker, a shownotes QR code, and live captions that highlight each word as it is spoken, lightly corrected while still matching the speech. Use when the user asks for shorts, reels, TikToks, vertical clips, social clips or viral moments from a talk, or wants captions fixed on one.
user_invocable:
Yes

Shorts Producer

Process steps in order. Do not skip ahead.

Resolve the absolute path of this loaded SKILL.md, then set speaker_toolkit_root to the plugin root two directories above the directory containing this file. Never derive it from the consumer working directory. Treat {speaker_toolkit_root} as absolute in every toolkit-owned command; vault, recording, and output paths remain consumer-owned. In an ACR install this skill's directory is named acr__jbaruch__speaker-toolkit__<skill>; there, read every {speaker_toolkit_root}/skills/<name>/ path and every relative link into another skill's <name>/ as sibling acr__jbaruch__speaker-toolkit__<name>/.

This skill runs on macOS: it needs ffmpeg, pdftoppm, swift (Apple Vision) and the vault venv's mlx-whisper. Every script prints diagnostics on stderr, exits 2 on a usage error, and on a refusal prints {"ok": false, "error"} and exits 1; the error says what to change.

Step 1 — Resolve the interpreter

python3 "{speaker_toolkit_root}/skills/screencast-recorder/scripts/resolve-interpreter.py" <vault_root>

Stdout: {"ok": true, "python_path", "vault_root", "database"}. Use python_path for every command below. On exit 1, repair through Skill(skill: "vault-ingress"). Proceed immediately to Step 2.

Step 2 — Read the talk

{python_path} "{speaker_toolkit_root}/skills/vault-ingress/scripts/read-tracking-database.py" <database>

Find the talk the user named in database.talks[]. It must have status processed; otherwise finish here and tell the user to process it with Skill(skill: "vault-ingress") first. Note filename, title, conference, video_url, shownotes_url, slides_local_path, pattern_observations, humor_postmortem and verbatim_examples, the matching qr_codes[] entry (talk_slug, short_url), and the speaker profile's speaker.handle and speaker.website from <vault_root>/speaker-profile.json. A talk without a deck PDF at slides_local_path cannot be rendered; finish here and say so. Proceed immediately to Step 3.

Step 3 — Get the recording

Pick a local work folder, ~/Library/Caches/speaker-toolkit/shorts/<talk filename stem>/, never a cloud-synced one. When the user has a local recording of the talk, use it. Otherwise download the best stream up to 2160p with the vault's yt-dlp:

{python_path} -m yt_dlp -f "bv*[height<=2160]+ba/b[height<=2160]" --merge-output-format mkv -o "<work_dir>/source.%(ext)s" <video_url>

The face crop is cut from this file, so take the highest resolution offered. Proceed immediately to Step 4.

Step 4 — Pick the takes

Read the talk's analysis at <vault_root>/analyses/<filename> and its timed transcript at <vault_root>/transcripts/<youtube_id>.segments.json, then choose takes as skills/shorts-producer/references/clip-selection.md describes. Write plan.json in a shorts/ folder inside the talk's own working folder (the folder that holds its deck build); ask the user where that is when the record does not make it clear. Its shape is the docstring of {speaker_toolkit_root}/skills/shorts-producer/scripts/shorts_plan.py. Fill the plan's top level and each clip's id, title, hook, start, end and rationale now. Set qr_url to the qr_codes[] short_url, else shownotes_url, else null. Proceed immediately to Step 5.

Step 5 — Prepare each clip

{python_path} "{speaker_toolkit_root}/skills/shorts-producer/scripts/prepare-clip.py" --plan <plan.json> --clip <id>

Run it once per clip. Stdout names words (measured word timings in source seconds), frames, faces, contact (a frame every two seconds with a 10% grid and red boxes on detected faces) and deck_contact (numbered deck pages). Proceed immediately to Step 6.

Step 6 — Finish each plan entry

For every clip, look at its contact.jpg and the deck's contact.jpg, read its words.json, and fill in:

  • start and end on word boundaries from words.json, so the clip opens on its first word and ends after its payoff. A boundary outside the prepared window means re-running Step 5 for that clip.
  • speaker_region: the normalized [x, y, w, h] box the speaker stays inside for the whole clip, leaving out faces on the projected slides and in the audience.
  • slides: [clip offset seconds, deck page] pairs, starting at 0, for each slide shown on the stage screen. Open individual frames/<seconds>.jpg to pin a change to the half second.
  • captions: the spoken words, corrected as skills/shorts-producer/references/caption-corrections.md allows.

Proceed immediately to Step 7.

Step 7 — Render each clip

{python_path} "{speaker_toolkit_root}/skills/shorts-producer/scripts/render-short.py" --plan <plan.json> --clip <id>

Stdout: {"ok": true, "video", "srt", "txt", "duration", "caption_edits", "crop"}, the files landing in the plan's out_dir. Read caption_edits: every entry must be a correction you meant. A refusal that says the captions stray from the speech means correcting less, then re-running this step. Proceed immediately to Step 8.

Step 8 — Verify each clip

{python_path} "{speaker_toolkit_root}/skills/shorts-producer/scripts/verify-short.py" --plan <plan.json> --clip <id>

Exit 0 means every check passed; exit 1 lists failures. Open review (six frames) and check that the speaker stays framed, the slide matches what was on stage, and nothing overlaps the captions. Fix the plan entry and return to Step 7 for any clip that fails or looks wrong. Proceed immediately to Step 9.

Step 9 — Present the shorts

List each short with its file path, hook, duration and caption edits, and the review.jpg path. Uploading and posting stay with the user. Finish here.

skills

README.md

tile.json