sumdog-millionaire/claude-skills

Video Digest

Turn a recording into a durable evidence map before deciding what to do with it.

A transcript and a recording are complementary sources. Speech establishes what participants said and why. The recording establishes what was visible, including names, URLs, commands, fields, controls, documents and demonstrated results. Analyse both and link each important visible point to the clearest relevant frame.

Before running

Use AskUserQuestion before processing to settle only choices the request has not already answered:

• Emphasis: complete meeting record, detailed screen walkthrough, learning and methods, or all of these.
• Destination: beside the supplied files or in a separate run folder.
• External research: source-bound digest only, or verify and expand what the video names.

Every option still produces a detailed evidence map. Emphasis changes where the most detail goes; it does not permit a thin summary. Do not ask again about a choice the user already made.

Before the first long tool call, say in one line that you will parse or transcribe the recording, inspect the screen throughout, connect frames to the transcript and then return with the evidence map and possible next actions.

Boundaries

The evidence map is the only output approved by starting this skill. Reading, downloading, transcribing, frame extraction and requested research are preparation for it.

After the evidence map, propose downstream work and ask the user what to do next. Create tracker items, edit a project, contact services, or create a skill or agent only after the user chooses that route and approves the specific write. This prevents a recording from silently turning into work nobody agreed to.

Do not reproduce credentials, card data, private links or unnecessary personal data visible on screen. Record only what the meeting requires and describe sensitive values without copying them.

1. Identify the source

Accept either:

• a folder or explicit pair containing one recording (.mp4, .mov, .mkv or .webm) and one transcript (.docx, .txt, .srt or .vtt); or
• a public video URL.

If several recordings or transcripts are plausible, ask which pair to use. Do not guess.

Use a stable workspace:

• For a supplied meeting folder, leave originals in place and store transcript.txt, frames/, meeting-analysis/ and meeting-notes.md in that folder. Use a separate destination only when the user requests one.
• For other local files, use the supplied folder or ~/video-digest-runs/<date-or-id>--<slug>/.
• For a URL, use ~/video-digest-runs/<video-id>--<slug>/ and keep the metadata, video, captions, transcript and frames so a removed post does not erase the source.

Set <report> to meeting-notes.md for a recorded meeting and digest.md for another video. Check for an existing run with the same stable ID before fetching or processing. If <report> exists, state its outcome and ask whether to reuse it or make a fresh pass. Reuse any complete files from an abandoned run.

2. Prepare the transcript and frames

Read references/ingest.md and follow the branch for the supplied source. Completion means the workspace contains transcript.txt, frames/ and <report>. transcript.txt must preserve timestamps and speaker names when the source provides them. frames/ must cover the whole recording densely enough to account for every substantive visible change. Open <report> with source details before long processing so an interrupted run remains identifiable.

3. Build the evidence map

Read references/evidence-map.md and follow it in full.

Read every transcript turn and inspect the screen across the entire recording. At each substantive point, connect:

1. what was said;
2. what was shown or demonstrated;
3. why it mattered;
4. whether it created a decision, action, requirement, risk or open question; and
5. the timestamp and clearest supporting frame or frames.

Distinguish stated, shown, demonstrated, inferred and unclear evidence. A pointer resting over a control does not prove a click; use before-and-after frames when the action matters. Preserve caveats, disagreement and uncertainty rather than smoothing them into consensus.

Write the complete result to <report>. Analyse a short recording in the main agent. For a long recording, delegate contiguous time ranges as described in the reference, require a durable segment file from every range, and read every segment before synthesis.

Completion means every second belongs to the analysis, every substantive transcript turn is accounted for, every material screen event has been inspected, and the report passes the reference's review checklist.

4. Brief the user

Lead with a compact account of what the recording covered, its main decisions and actions, and what the screen added that the transcript alone did not reveal. State where <report> and its frames are stored and name any important uncertainty.

Then use AskUserQuestion to ask what should happen next. Offer only routes supported by the digest, which may include:

• keep the evidence map as the final output;
• turn selected findings into work in the project's issue tracker;
• make specified project or documentation changes;
• verify or expand named material through external research;
• create or update a skill or agent; or
• prepare another user-named deliverable.

Allow multiple selections when several routes can apply. Do not imply that an action belongs in an issue tracker merely because it was mentioned in a meeting.

5. Carry out the chosen next action

Read references/next-actions.md and follow only the selected branch. Show the concrete proposal before any write outside digest.md, then get approval and make exactly that change.

Finish <report> with a one-line Outcome recording what happened after analysis, including the project or tracker when relevant. If the user chooses no next action, use **Outcome:** Digest only. This line makes repeat runs cheap to identify.

Failures

Report the failed stage and its error, then stop cleanly. Do not retry blindly against rate limits, guess at transcript structure, substitute a mismatched recording, or continue with partial coverage while presenting the result as complete. Keep completed intermediate files so a later run can resume.