SkillsSkill for Claude

B-roll needs list from talking head takes

Hand over the A-roll takes for one episode and the handful of things your script mentions. Claude indexes the takes in Vivu, finds the stretches where you talk about each item, checks it against your transcript for the exact words and against frames for what is on screen, and writes a CSV with one row per mention and an empty column for how you will fill it. A transcript shows where you said it; only the footage shows whether you were holding it up, whether an insert already covers it, or whether it is just your face.

Maintained by Vivu. Updated 2026-11-19.

Download

talking-head-broll-needs-list.zip

8 KB. Unzips to talking-head-broll-needs-list/SKILL.md. Upload the zip as it is in the Claude app, or unzip it into your skills folder for Claude Code.

SHA-256 b9ba197a399f90702295c797fcff34671ca1073512df6534fb808499471f0ba5

At a glance

What the B-roll needs list from talking head takes skill does, where it runs, what it needs, and when it asks
Looks forEach stretch of your takes where you talk about an item viewers need to see, labeled talk only, shown, or already a cutaway. That label only exists in the picture, not in the transcript. Every row is checked against the transcript and frames before it goes into the list.
Runs onClaude Code on your computer (terminal or the Code tab of Claude Desktop): it needs a shell with ffmpeg and ffprobe and your takes on local disk. No residential IP and no scheduler; it runs once per episode.
Needs
  • The Vivu connector with write access, to create a project, upload the takes and search them.
  • A shell with ffmpeg and ffprobe, to measure the takes and make contact sheets from your own files.
  • The A-roll files for one episode on local disk, because frames come from the originals.
  • A local transcript per take (an editor export or a speech to text tool), because Vivu's descriptions are paraphrases.
  • A browser you can open the upload link in, or a browser tool that can attach local files.
Your Vivu planIndexing uses one index minute per minute of takes, and each item costs one precise search (5 credits), plus 5 for each reworded query. The skill shows these against your plan from vivu_get_usage before uploading.
Asks you firstConfirming the item list, renaming camera default file names, indexing the takes after showing minutes and credits against your plan, uploading takes that show anyone besides you, and writing the full CSV after you approve one sample row.

The skill does its video work through the Vivu connector. If the connector is not in your Claude yet, add it first; the skill checks that it is connected before it does anything else.

Before you run it

  • Use your own footage; anyone else on camera needs to consent, and a child needs a parent or guardian's consent or those takes are not uploaded.
  • The takes stay in your Vivu project until you delete them; the skill creates it as private because the default is visible to your whole organization.
  • Claude in Chrome uploads at most 10 MB per call; larger takes go through your own browser or the Vivu web app, never split or recompressed.
  • Search windows are wider than the mention and can run into the next item, so each row is trimmed with your transcript times.
  • A mention made over a screen recording can be missed by a speech only query; the skill rewords it and greps the transcript before calling an item absent.
  • The picture label comes from Claude looking at frames and is a draft; unclear cases such as a small object in hand are marked UNCERTAIN.
  • Retakes of the same line were not part of the test; with many takes, expect one window per take or a wide window covering two takes.
  • Vivu only finds moments; it does not cut footage, place B-roll, generate pictures, or recognize faces or logos.

Start it

Once the skill is installed, ask for the task in your own words. Naming the skill is the most reliable way to have Claude use it. For example:

Here are my A-roll takes for this week's episode in ~/Videos/EP12. I mention the app's settings page, the new dock, the charger, and the sales chart. Make me a B-roll needs list.

In Claude Code you can also type /talking-head-broll-needs-list. Claude asks for anything the request leaves out, most important first.

What is inside

  1. When to use
  2. Working principles
  3. What you need before starting
  4. Inputs to collect
  5. Files and state
  6. Step 1: Check the Vivu connector and the tools
  7. Step 2: Collect the takes, the item list and a transcript
  8. Step 3: Measure and price
  9. Step 4: Create the project and upload
  10. Step 5: Search for each item
  11. Step 6: Check every window against the transcript and the frames
  12. Step 7: Assemble the needs list and confirm it
  13. Compliance
  14. Known failure modes

The full skill

This is talking-head-broll-needs-list/SKILL.md from the download, as Claude reads it: the frontmatter first, then the instructions.

---
name: talking-head-broll-needs-list
description: "Turn your talking head A-roll into a B-roll needs list with Vivu: find the spots where you mention something viewers should see, and note what is on screen there. Use before editing."
---

B-roll needs list from talking head takes

This skill turns a creator's own talking head recordings into a B-roll needs list. The creator hands over the A-roll files for one episode and names the 6 to 8 things the script mentions that viewers need to see (an app screen, a product, a chart, a place). Claude indexes the takes in Vivu, runs one search per item to find the stretches where the creator talks about it, then greps the transcript for mentions the search missed, checks each stretch against a local transcript for the exact words and against frames for what is on screen, and writes a CSV with one row per mention: item, file and take, start and end, the transcript line, the current picture, a frame, and an empty column for how the creator will fill it.

The value is in the picture column. A transcript tells you where you said "the charger"; it cannot tell you whether you were holding the charger up to the lens, whether a top down insert already shows it, or whether the shot is just your face. Only the footage shows which mentions still need a shot, so the list points the creator at the gaps instead of every mention. Every row is checked against the transcript and the frames before it goes into the list, and Vivu's own description of a moment never replaces either.

When to use

Use when a creator says things like "where do I need B-roll in this episode", "list the spots where I talk about the app but only my face is on screen", "make me a cutaway list from my A-roll", or "which mentions still need a product shot". Use it before the edit, on the raw takes.

Not for: finding shots inside a folder of B-roll footage (search that folder with Vivu directly, or use a shot pull list skill); picking one thumbnail frame; a single question about one take ("where do I mention the price?"), which is one Vivu search. Vivu here only finds moments. It does not cut, place B-roll, or generate pictures.

Working principles

  1. Report measured numbers, not estimates. When a number is an estimate, say so.
  2. Nothing is verified until it has been checked against the source. A Vivu window stays a candidate until the transcript shows the item is being talked about there and the frames have been looked at.
  3. Stop and tell the user when a required capability or tool is missing (no transcript, no ffmpeg, no write access). Do not guess around it.
  4. Ask the user before anything that is expensive to redo or acts on their behalf: the item list, indexing (it uses plan minutes), and the full CSV after one sample row.
  5. When a mention cannot be found, say so in the list. An empty search result does not prove the item is never mentioned.

What you need before starting

Check each item at the start of the run and tell the user plainly what is missing before doing anything else.

Requirement Why How to check
Vivu connector with write access create a project, open its upload page, search vivu_get_account shows can_create_projects: true (tool names may carry a server prefix). "has not granted vivu.write" means the connection is read only; the user reconnects Vivu and allows write access
A shell on the computer that holds the takes, with ffmpeg and ffprobe measure minutes, make contact sheets from the user's own files ffmpeg -version and ffprobe -version
The A-roll files for one episode, on local disk Vivu indexes copies; frames come from the originals ls AROLL_DIR shows the takes
A local transcript per take (SRT or VTT) exact words for each row; Vivu's reason text is a paraphrase the file exists next to each take, or the user's editor (Premiere, Resolve, Descript) can export one, or a local speech to text tool is installed
A browser the user can open a link in, or a browser tool that can attach local files upload into Vivu the user confirms, or the browser tool is listed

No connector writes anywhere outside the working folder, and nothing runs on a schedule.

Inputs to collect

Ask for anything missing, most important first.

  1. The folder of A-roll takes for this episode (required).
  2. The items to look for, 6 to 8 of them, each in a few words plus a hint of context ("the dock, when I talk about how it looks"). Required; offer to draft the list from the script or the transcript if the user prefers, then confirm it.
  3. How the takes are numbered. Default: the take number comes from the file name (EP12_take2.mp4 is take 2); rename before upload if the names are camera defaults and the user agrees.
  4. Whether the footage shows anyone besides the creator. Default: no; if yes, ask for their consent before upload (Compliance).
  5. Output file name. Default: broll_needs.csv in the working folder.

Files and state

Keep everything in one working folder next to the takes:

broll-needs/
  config.json        project_id, item list with wordings, takes with durations and video_ids
  state.json         completed steps, job_id per item, checked windows, rows written
  transcripts/       one SRT or VTT per take
  results/           raw search results, one JSON per item
  frames/            one contact sheet per window
  broll_needs.csv    the list

state.json lets an interrupted run resume: skip items whose job_id already has a saved result, skip windows already checked, and never write a row twice (the row key is item + file + start).

Step 1: Check the Vivu connector and the tools

Goal: know before any work that every requirement is met.

  1. Call vivu_get_account. If the tool does not exist, tell the user to add the Vivu connector in Claude (https://mcp.vivu.ai/mcp) and stop. If it returns can_create_projects: false, or a later write fails with "has not granted vivu.write", ask the user to reconnect Vivu with write access and stop.
  2. Run ffmpeg -version and ffprobe -version. If either is missing, say so and stop.
  3. Check the remaining rows of What you need.

Done when vivu_get_account shows can_create_projects: true, both ffmpeg commands print a version, and the user has confirmed the folder and the upload path.

Step 2: Collect the takes, the item list and a transcript

Goal: a confirmed item list and one transcript per take.

  1. List the takes in AROLL_DIR. Rename camera default names to EPISODE_takeN.mp4 if the user agrees, because Vivu shows file names in results and the take number has to survive. If the creator has already started an edit with these files, renaming them breaks the media links in the editor project; ask first and rename only with a yes.
  2. Get a transcript per take. Prefer the editor's export (SRT or VTT). Otherwise use a local speech to text tool the user has; for example, with the openai whisper command line installed: whisper TAKE.mp4 --model small --output_format srt --output_dir transcripts. Running whisper was not exercised in our test run. If a take has no transcript, tell the user that rows from it will be marked UNCERTAIN.
  3. Write the item list into config.json, one entry per item: a short label for the CSV and the search wording (Step 4). Show it to the user and wait for a yes.

Done when config.json lists the confirmed items and transcripts/ holds a transcript for every take, or the user has accepted the missing ones.

Step 3: Measure and price

Goal: the user approves the index minutes and search credits before anything is uploaded.

  1. Measure each take. ffprobe reads one input per call, so loop:
for f in AROLL_DIR/*.mp4; do echo "$f"; ffprobe -v error -show_entries format=duration -of csv=p=0 "$f"; done
  1. Total the minutes. Searches: one precise search per item, plus one more for each item that needs a reworded query.
  2. Call vivu_get_usage and show one table:
This episode Plan allowance Remaining
Index minutes measured total from vivu_get_usage from vivu_get_usage
Search credits items x 5, plus 5 per rewording (estimate) from vivu_get_usage from vivu_get_usage

Plan facts: Free is $0 a month with 20 index minutes a month and 50 search credits a month; Premium is $30 a month with 180 index minutes a month and 500 search credits a month. A fast search uses 1 credit and a precise search uses 5 credits in total.

  1. If it does not fit, offer levers that shrink what gets indexed, in order: drop takes the creator already knows are unusable (false starts the creator will not use), index only the takes of the segments still being edited, then fewer items. Moving to a larger plan comes last. Do not drop takes the user asked to include without saying so.

Done when the user approves the minutes and credits in the table.

Step 4: Create the project and upload

Goal: every take is indexed in a private Vivu project. Step 4 was not exercised in our test run (the test videos were already indexed).

  1. Call vivu_list_projects and reuse a project for this episode if one exists. Otherwise call vivu_create_project with name "B-roll EPISODE" and visibility "private". The default visibility is the whole organization, and unreleased episodes usually should not be.
  2. Call vivu_open_upload_page with the project_id immediately before use. The upload link expires in 180 seconds, so hand it to the user at once to open in their own browser and pick the takes, or open it in a browser tool that can attach local files. Claude in Chrome accepts at most 10 MB per upload call; larger takes go through the user's own browser or the Vivu web app. Never split or recompress a take to fit. Never post or store the link.
  3. Poll vivu_list_videos until every take shows ready. Vivu replaces spaces and brackets in file names with underscores, so match each video back to config.json by the take name, and record its video_id.

Done when vivu_list_videos shows every take as ready and config.json has a video_id per take.

Step 5: Search for each item

Goal: one saved result per item.

Run one vivu_search_videos call per item with mode "precise" and maximum_results 10. Precise is the only mode that returns a time range; fast returns whole files, which says nothing here because every file is part of the episode. Ten leaves room for an item mentioned in several takes; it is also the ceiling on what comes back, so raise it to 20 when one item is talked about in many takes.

Write each query as the meaning plus context, in the creator's words, not as an exact phrase:

Field Query pattern Mode
any item the creator talks about THING, CONTEXT (for example: the creator talks about the dock the console sits in, describing how it looks or feels) precise
item shown on a screen recording the creator says SOMETHING ABOUT THING while the screen shows WHAT THE SCREEN SHOWS precise

The search finds where an item is said; Step 6 then turns that into what is shown, by reading the transcript inside each window and looking at its frames. The picture label is not left to Vivu, because judging what a frame lacks is where visual search is weakest.

THING is the item, CONTEXT is what the creator says about it. Use the second pattern when the creator talks about the item over a screen recording or slide, or when the first wording comes back empty and the transcript shows the item is there.

Each call returns a job_id. Call vivu_get_search_results until complete is true; each call can wait up to 45 seconds, so a pending search is not a stalled one. Save the complete result as results/ITEM.json and the job_id in state.json. The Vivu result page link in the result expires after four hours; show it only in the live reply, never in the CSV.

Done when results/ holds one complete result per item and state.json lists every job_id.

Step 6: Check every window against the transcript and the frames

Goal: every candidate window is either a confirmed mention with a picture label or dropped.

For each window (start_ms, end_ms):

  1. Transcript: read the lines between the window's start and end. Keep the window only if those lines talk about the item; copy the line that names it into the row. The reason text Vivu returns is a paraphrase, not a transcript, so never quote it. Trim the row's start and end to the lines about this item, because windows are wider than the mention and often run into the next thing the creator talks about.
  2. Frames: make a contact sheet across the window, one frame per second (two per second for windows under 10 seconds, one every two seconds for windows over 50 seconds), and look at it:
ffmpeg -v error -ss START -to END -i TAKE.mp4 -vf "fps=1,scale=192:-1,tile=8x4" -frames:v 1 frames/ITEM_hN_sheet.png

START and END are the window in seconds, TAKE.mp4 is the original take, ITEM_hN is the item and the hit number. Make the tile large enough for every frame (columns x rows at least the window length times fps).

  1. Label the current picture:
    • talk_only: the item is not in the picture, or only small in the creator's hands in a medium shot;
    • shown: the item is held toward the lens and large in frame, or a close or top down insert of it is on screen;
    • cutaway: the picture is already a product video, a web page, a slide or a screen recording about the item. When a small object is held in hand and the label is unclear, write the label you lean to plus UNCERTAIN.
  2. Coverage check: search each transcript for the item's name and synonyms (grep -n -i "ITEM_WORD" transcripts/*.srt transcripts/*.vtt). Any mention that no window covers is a missed mention: reword the query once (Step 5 second pattern), and if it is still missed, add the row from the transcript times and mark the picture NOT VERIFIED until frames are checked.

Windows whose transcript lines are about something else are false positives. Count them and tell the user the count; they do not go into the list.

Picture labels come from Claude looking at frames. Treat them as a draft for the creator: shown and cutaway mean a shot already exists, talk_only means a gap.

Worked example from our test run

In our test run on 3 public creator videos (14 minutes of edited unboxing and first impressions footage, with no retakes), we searched for the things the creators talk about (a controller grip, an HDMI cable, a charger, a dock, a camera accessory, a watch band, a power adapter and a watch face) plus one control item nobody mentions. With the first wording, every item except the watch face returned its mentions: 9 windows, 9 real, 0 false, and no captioned mention missed, including two one sentence asides. The watch face, which the creator talks about while the picture is a phone screen recording, returned 0 results for two speech only wordings and was found after 2 rewordings on the same corpus, once the query also described what the screen shows. Across the final wordings that is 10 windows, 10 real and 0 false; the control item returned 0 results. Windows ran from 9 to 64 seconds wide and every one contained the start of its mention, but the dock window (36 seconds) started inside the charge cord line and ran to the next item, so rows have to be trimmed with transcript times. Claude's picture label from the contact sheets disagreed with the label written before the search only on a small power adapter held in hand. Both sets of labels were written by the same agent, so this is not an independent check of the labels. The videos were already indexed from an earlier test, so this run measured no upload time, index minutes or time to ready. One of the test videos had no transcript, so its windows were not part of the miss count; the test videos were edited uploads rather than raw takes, and retakes of the same line were not exercised in our test run.

Done when every window in results/ is either a row with a transcript line, a frame and a picture label, or counted as a false positive, and every transcript mention of every item is covered or listed as missed.

Step 7: Assemble the needs list and confirm it

Goal: broll_needs.csv the creator can work from.

  1. Show one sample row and the field mapping, and wait for a yes before writing the whole file:
item,file,take,start_mmss,end_mmss,transcript_line,current_picture,frame,fill_with
power adapter block,EP12_take1.mp4,1,04:14,04:30,"I also got a block because I only have like one for my iPhone",talk_only (UNCERTAIN: small adapter held at chest height),frames/power_block_h1_sheet.png,
Field Source If unavailable
item config.json label none
file, take Vivu file name matched to config.json; take from the file name UNCERTAIN
start_mmss, end_mmss window start_ms / end_ms, trimmed to the transcript lines window as returned, marked untrimmed
transcript_line local transcript NOT VERIFIED
current_picture Claude's label from the contact sheet NOT VERIFIED
frame contact sheet path blank
fill_with left empty for the creator (reshoot, screen record, or pull from old footage) blank
  1. Sort rows by take and start time. An item mentioned in several takes gets one row per take, so the creator can pick the take they keep.
  2. Write broll_needs.csv, then add a short summary in the reply: rows per picture label, false positives dropped, items with no mention found, and the Vivu result page links (expires after four hours) for a quick look.
  3. Step 7's user confirmation of the sample row was not exercised in our test run; the CSV itself was assembled.

Done when the user has approved the sample row and broll_needs.csv holds one row per confirmed mention.

Compliance

  1. Use the creator's own footage. If anyone else appears or speaks in the takes (a guest, a family member), ask for their consent before Step 4.
  2. Minors: if a child appears, the skill needs the explicit consent of a parent or guardian before upload; without it, do not upload those takes.
  3. The takes stay in the user's Vivu project until they delete them. Create the project as private (the default is visible to the whole organization).
  4. Vivu finds moments only. The skill does not cut footage, place B-roll, generate pictures, or recognize faces or logos.

Known failure modes

Symptom Cause Fix
an item talked about over a screen recording returns 0 results (observed) in our test run the watch face came back empty for two speech only wordings while the picture was a phone screen reword with the second query pattern, and grep the transcript for the item before saying it is never mentioned
one window covers two neighboring items (observed) precise windows are wider than the mention; in our test run the dock window started inside the charge cord line trim each row to the transcript lines about that item
"503 Stream removed (recvmsg:Connection reset by peer (104))" when starting a search (observed) a transient connector error retry the same call once; it went through on retry in our test run
ffprobe stops with "provided as input filename, but" ... "was already specified." (observed) ffprobe reads one input file per call measure each take separately with the Step 3 loop
the picture label for a small object in hand is unclear (observed) an adapter held at chest height in a medium shot reads as shown or as talk_only write the leaning label plus UNCERTAIN and let the creator decide
a take has no transcript no editor export and no speech to text tool export one or install one; otherwise mark its rows UNCERTAIN and use vivu_get_video_summary with include_segments, start_ms and end_ms around the window as a second signal
the same line in two takes comes back as one wide window windows are wider than the moment, and takes recorded back to back sit close together split the window with the transcript and write one row per take
an item has no window and no transcript match the creator never said it, or said it in other words ask the creator; an empty result does not prove the item is absent
the transcript line in a row does not match what was said the line was copied from Vivu's reason, which is a paraphrase copy lines only from the local transcript
"has not granted vivu.write" Vivu connected read only reconnect Vivu with write access
upload page asks to sign in or shows an error the one time upload link expires after 180 seconds request a new link right before opening it
a take is rejected by the browser upload tool the file is above the tool's limit (Claude in Chrome accepts at most 10 MB per upload call) the user opens the upload link in their own browser or adds the take in the Vivu web app
a result's file name does not match config.json Vivu replaced spaces or brackets with underscores match on the take name

Connect Vivu, then add the skill.

The skill runs through the Vivu connector. Add it to Claude from the connector directory, then install the skill.