---
name: current-workflow-evidence-table
description: "Turn the screen share parts of remote user interviews into a current workflow evidence table with Vivu: which tools participants opened, window titles, why, and timestamps."
---Current workflow evidence table from interview screen shares
This skill turns a round of remote discovery interviews (one mp4 and one transcript per participant, cut to the part where the participant shares their screen to show how they do the job today) into a current workflow evidence table for the designer who draws the current state journey map. Claude measures the clips, prices them against the user's Vivu plan, indexes them in a private Vivu project, runs one search per kind of window the researcher cares about (a spreadsheet, an older internal system or another vendor's admin tool, an internal report page, an email inbox or chat), checks every hit on frames, reads the window title from a full frame, takes the participant's own words about why from the transcript, and writes current_workflow_evidence.csv with one row per checked moment and an empty "journey step" column for the designer.
The value is in what only the shared screen shows. Participants narrate as "I start here, then I paste it over there"; the transcript has no tool names and no file names. Which tool they opened, and which file or page, is on the screen and nowhere else. Every row in the table has been looked at on a full frame before it is written, the window title comes from that frame and never from Vivu's reason text, and the weakest searches (spreadsheet and older system) were measured against a test corpus full of look alikes: video call gallery grids and interviewer slides with tables.
When to use
Use when a UX researcher or product designer says "pull every tool people showed us in the interviews", "what did participants actually open when they shared their screen?", "build the current state evidence for the journey map", or "which interviews showed a spreadsheet or the old system?". For one question about one interview ("what did P07 open first?"), search Vivu directly or watch the clip. For usability test sessions where the goal is to find errors and friction in your own product, use a usability test issue workflow instead; this skill inventories the tools people use today, it does not judge them or the participant. Synthesis (themes, journey stages) stays with the researcher; the table only feeds it.
Working principles
- Report measured numbers, not estimates. When a number is an estimate, say so.
- Nothing goes into the table until its frames have been checked. Search results are candidates; a rejected candidate is counted, not written.
- Stop and tell the user when a required capability or file is missing (no transcript, no ffmpeg, no write access in Vivu). Do not guess around it.
- Ask the user before anything that is expensive to redo or that acts on their behalf: the indexing cost before upload, the sample rows before the full table.
- Window titles come from a full frame Claude has looked at; quotes come from the transcript. The reason text from Vivu paraphrases and can describe a title that is not on screen.
- Describe sensitive things on screen by kind (customer emails, an access token), never copy them, and never rate or describe the participant.
What you need before starting
Check each item at the start of the run and tell the user plainly what is missing.
| Requirement | Why | How to check |
|---|---|---|
| Vivu connector with write access | create a private project, open its upload page, search | vivu_get_account shows can_create_projects: true (tool names may carry a server prefix). A write call failing with "has not granted vivu.write" means the connection is read only; the user reconnects Vivu and allows write access |
| A shell with ffmpeg and ffprobe on the machine that holds the recordings | cut the demo parts, measure minutes, extract frames and contact sheets | ffmpeg -version and ffprobe -version |
| The interview recordings on local disk, one mp4 per participant, with the participant ID in the file name | frames come from the originals (Vivu has no export tool) and the participant column is read from the file name | ls RECORDINGS_FOLDER shows names such as interview_P03.mp4 |
| A transcript with timestamps (VTT or SRT) per recording | the quote column comes from it, never from Vivu's reason text | one .vtt or .srt next to each .mp4; the meeting tool's transcript export makes it |
| The interview guide, or the question that starts the demo part | to cut each recording to the screen share part before indexing | the user pastes the question, for example "can you show me how you do this today" |
| Consent that covers screen recording and cloud processing | the screen shows the participant's work tools and often customer data | the user confirms the consent form covers both |
| A browser that can open the Vivu upload page | the connector has no direct upload tool | the user opens the link, or Claude in Chrome is connected |
No residential IP is needed (nothing is downloaded from YouTube) and nothing is scheduled.
Inputs to collect
Ask for anything missing, most important first.
- The recordings folder and the study name. Required.
- The window kinds to look for. Default: spreadsheet, older internal system or another vendor's admin tool, internal report page, email or chat. The researcher can add one (a CRM, a ticket queue); each extra kind is one more search.
- Where the demo part starts in each interview. Default: the first transcript line where the interviewer asks to be shown how the participant does it today, to the end of the screen share, at most 15 minutes.
- The Vivu project name. Default: "STUDY_NAME current workflow", private.
Files and state
Keep everything in one working folder next to the recordings:
workflow-evidence/
config.json study name, window kinds and queries, vivu project id
clips/ the cut demo parts, one mp4 and one vtt per participant
sessions.csv one row per clip: file, participant, duration seconds, vivu video id
results/ raw search results, one JSON file per window kind
frames/ contact sheets and full frames
current_workflow_evidence.csv the table
state.json uploaded clips and video ids, searches run with job ids, hits checked with verdicts, rows written
state.json is updated after every step. A rerun reads it first: clips already uploaded are not uploaded again, searches already run are not rerun, and hits already checked keep their verdicts, so an interrupted run picks up where it stopped and no moment is written twice.
Step 1: Check the Vivu connector and the setup
Goal: every row of What you need is confirmed before anything is cut or uploaded.
- Call vivu_get_account. If the tool is missing, tell the user to add the Vivu connector in Claude (https://mcp.vivu.ai/mcp) and stop. If can_create_projects is not true, or a write call later fails with "has not granted vivu.write", ask the user to reconnect Vivu with write access and stop until they have.
- Run ffmpeg -version and ffprobe -version. List the recordings folder and match each mp4 to its transcript and a participant ID in the file name (P03, P07). Rename files to letters, digits and "" only (interview_P03.mp4); Vivu replaces spaces and punctuation with "", and plain names match back without guessing.
- Ask the user to confirm the consent form covers screen recording and processing by a cloud service. If it does not for a participant, leave that recording out and say so. Asking the user about consent was not exercised in our test run.
Done when vivu_get_account shows can_create_projects: true, ffmpeg and ffprobe print versions, the user has confirmed consent, and sessions.csv has one row per recording with its transcript.
Step 2: Cut the demo parts and price the indexing
Goal: only the screen share part of each interview gets indexed, and the user approves the cost first.
- In each transcript, find the interviewer's question that starts the demo (the guide question from the inputs) and the point where the screen share ends (the participant stops sharing, or the interviewer moves to the next topic). Cut that part with a re-encode so the cut points are exact:
mkdir -p workflow-evidence/clips workflow-evidence/frames workflow-evidence/results
ffmpeg -v error -ss START -to END -i RECORDINGS_FOLDER/INTERVIEW.mp4 -c:v libx264 -c:a aac workflow-evidence/clips/interview_PID.mp4
START and END are seconds in the full recording; INTERVIEW is the recording's file name without .mp4; PID is the participant ID. Shift the transcript lines by START so their times match the clip, and save them as workflow-evidence/clips/interview_PID.vtt. 2. Measure each clip and sum the minutes:
ffprobe -v error -show_entries format=duration -of csv=p=0 workflow-evidence/clips/interview_PID.mp4
- Call vivu_get_usage for the plan and what remains this month.
- Estimate search credits: one precise search per window kind for the whole round (not per interview, because one search covers the project), so the four default kinds are 20 credits as an estimate. A precise search uses 5 credits in total and a fast search uses 1 credit. Rewording a search or adding a kind adds 5 each.
- Show one table and wait for approval:
| This round | Remaining on the plan | |
|---|---|---|
| Index minutes | measured sum of the clips | from vivu_get_usage |
| Search credits | estimate | from vivu_get_usage |
For reference, the Free plan has 20 index minutes a month and 50 search credits a month; Premium is $30 a month with 180 index minutes and 500 search credits. Indexing whole one hour interviews would rarely fit; cutting to the demo part is the main saving. If the round still does not fit, offer these levers in order: shorten each clip to the stretch where the screen is shared, index the interviews the designer needs first and the rest next month, and only then a larger plan. Never drop an interview the user asked for without saying so.
In our test run the clips were already cut to the demo part, so cutting at the guide question was not exercised in our test run, and the account was an admin account whose vivu_get_usage shows no remaining allowance, so the plan comparison and approval were not exercised in our test run either.
Done when every clip is in workflow-evidence/clips with its transcript and the user has approved the minutes and the credit estimate.
Step 3: Upload and index
Goal: every approved clip is ready in a private Vivu project.
- Call vivu_list_projects and reuse the project named in config.json if it exists. Otherwise call vivu_create_project with that name and visibility "private". Interview screens show internal tools and often customer data; the default visibility is the whole organization.
- Call vivu_open_upload_page with the project ID right before the upload. The link expires in 180 seconds and is a sign in link: do not paste it into any message or file. Give it to the user to open in their own browser and choose the clips, or attach them with a browser tool that can upload local files. Claude in Chrome accepts at most 10 MB per upload call; a 12 minute clip is usually larger, and the user adds those in the Vivu web app. Never split or recompress a clip to fit a tool limit.
- Poll vivu_list_videos about every 30 seconds until every clip shows ready. Match each video back to sessions.csv by file name. A clip can sit in "queued" while others index; keep polling, do not upload it again, because a duplicate doubles every count. Record the video IDs in state.json.
In our test run the clips went in through a scripted upload of the same upload page, not a user's browser or Claude in Chrome; that path was not exercised in our test run.
Done when vivu_list_videos shows every clip in sessions.csv as ready.
Step 4: Run one search per window kind and check every hit
Goal: one list of checked moments per window kind.
| Field | Query | Mode | maximum_results |
|---|---|---|---|
| spreadsheet | a participant shares their screen and it shows a spreadsheet file they work in | precise | 20 |
| legacy_system | a participant shares their screen and it shows an admin back office web page: a left navigation menu with sections such as orders, customers or settings, next to a list of records or a settings form | precise | 20 |
| report_page | a participant shares their screen and it shows an internal report or dashboard page with charts and numbers | precise | 20 |
| email_chat | a participant shares their screen and it shows their email inbox or a chat app where they look for information | precise | 20 |
Why these four: they follow the order in which people usually describe the work. The report page is where they notice something, the older system or another vendor's tool is where they look up the record, the spreadsheet is where they write it down, and mail or chat is where they chase the missing piece. Each search finds where on screen that step happened; the transcript then gives what the participant said about why, which moves each row from shown to said. All four are precise because the table needs time ranges; fast only returns whole files. maximum_results 20 is far above the handful of moments one kind has in a round of seven or eight clips, so a list that stops at exactly 20 means the cap was hit: raise it and rerun that search.
The legacy_system wording describes what is visible (a left navigation menu next to records or a form) on purpose. A wording that asked for "an older internal system" also matched a report page, and it leaned on the participant saying the place was older rather than on the screen. If the user adds a kind of their own (a CRM, a ticket queue), describe what the screen looks like, not what the tool is called; a new wording was not exercised in our test run.
Run each query with vivu_search_videos (project_id, query, mode, maximum_results). It returns a job ID. Call vivu_get_search_results until complete is true; each status call can wait up to 45 seconds, so a pending search is not a stalled one. Save each completed result as results/FIELD.json with the result_page_url field removed; show the result page link in the live reply only. It expires after four hours, so it never goes into the table or any file.
Check every hit before it goes anywhere:
- Extract three frames inside the window (start, middle, end):
ffmpeg -v error -ss SECONDS -i workflow-evidence/clips/interview_PID.mp4 -frames:v 1 -q:v 3 workflow-evidence/frames/FIELD_hN_POS.png
SECONDS is start_ms / 1000, the midpoint, or end_ms / 1000 minus 0.5; N is the hit number; POS is start, mid or end. 2. Look at the frames. The hit is real only if a frame shows the asked kind of window shared by the participant. The meeting tool's own screens are not: a gallery of participant tiles, a speaker view, or the interviewer's slides, even a slide with a table on it. A window of another kind is not either (a report page is not an admin tool). 3. Record every verdict in state.json. Rejected hits stay in results/ and are counted, not written to the table.
Vivu returns a time range that contains the moment, not an exact frame. The window tends to start a couple of seconds before the shared window appears and to end inside the next one, so the start and end frames often show something else; judge on the middle frame and set exact times in Step 5. Telling window kinds apart is a judgment about what the picture shows (visual semantics), which capabilities testing rates as weak in general, so a returned hit is a candidate until Claude has looked at its frames, and an empty list does not prove the participant never opened that kind of tool.
Worked example from our test run
In our test run we used 5 synthetic interview clips, 6.33 minutes in total, made as remote meeting recordings with a REC bar, a participant tile and a synthetic voice. Four clips each showed several shared windows: spreadsheets in a desktop and a web style, older internal admin pages and another vendor's settings page with a left navigation menu, report pages with charts, a mail inbox and a team chat. The participants never named the tool; they said "I start here" or "this one" and why. Distractors were gallery grids, speaker views, interviewer slides including some with tables, and the fifth clip had no screen share at all. From the start of the upload to the last clip ready took 101 seconds.
The spreadsheet query returned 4 moments, 4 real and 0 false, with 0 missed against 4 spreadsheets in the truth list; no gallery grid or slide with a table came back. Its median window was 21.1 seconds. The first legacy_system wording (an older internal system or another vendor's admin tool) returned 4 moments, 3 real and 1 false: an internal report page. The wording in the table came after 1 rewording on the same corpus and returned 3 moments, 3 real and 0 false, with 0 missed. report_page returned 3 moments, 3 real, and email_chat returned 3 moments, 3 real, each with 0 missed. A control search for a calendar app, which no clip showed, returned 0 moments. In some clips with an admin page the participant also said the place was older; the first wording's reasons repeated that speech, while the wording in the table asks only for what is on screen and its reasons described the layout.
The test corpus was clean synthetic footage with one voice per clip and few windows per clip; real interviews, busy desktops with many overlapping windows, small shared windows and screen shares at low resolution were not tested.
Done when every hit of every search has a verdict in state.json and the user has seen the true and false counts per search.
Step 5: Set times, read titles and take quotes
Goal: each checked hit has exact times, a window title read from a frame, and the participant's words.
- Make a two frames a second contact sheet of the window (START is start_ms / 1000; DURATION is the window length in seconds):
ffmpeg -v error -ss START -t DURATION -i workflow-evidence/clips/interview_PID.mp4 -vf "fps=2,scale=160:-2,tile=8x6" workflow-evidence/frames/FIELD_hN_sheet_%02d.png
Each sheet holds 48 tiles (24 seconds); sheet S, tile K (counting from 0) is at START + ((S - 1) * 48 + K) / 2 seconds. The first tile that shows the window is start_mmss; the last tile before the next window is end_mmss. 2. Read the window title from the full resolution middle frame (FIELD_hN_mid.png), not from the contact sheet tiles and never from the reason text. If the title bar is cut off or blurred, write UNCERTAIN and the visible part. 3. Take the participant's words from the transcript lines whose times fall inside start_mmss to end_mmss. Copy them as written. If nobody says why in that stretch, leave the quote blank rather than borrowing a line from elsewhere. 4. Look at the frame for customer data, email addresses, account numbers, access tokens or passwords. Write only the kind in sensitive_on_screen (for example "customer email addresses"), never the value.
In our test run the first and last tiles of all 13 checked windows matched the scene changes in the truth list, every window title read from the full frames matched the title on screen, and one reason misread a vendor name that the frame showed correctly. Quotes were taken from transcripts built from the clips' own timing, so local speech to text was not exercised in our test run.
Done when every checked hit has start_mmss, end_mmss, window_title_read (or UNCERTAIN), quote_local and sensitive_on_screen in state.json.
Step 6: Assemble the table and confirm the sample
Goal: the designer gets one table in participant order.
- Build one row per checked hit, ordered by participant and then time:
participant,category,session_file,start_mmss,end_mmss,window_title_read,quote_local,frame,sensitive_on_screen,journey_step,status
P03,spreadsheet,interview_P03.mp4,00:16,00:33,Weekly returns tracker.xlsx,"So every Monday I start here. I copy last week's numbers down, because nobody else keeps the history, and my manager asks for it.",frames/spreadsheet_h1_mid.png,no,,checked
P03,legacy_system,interview_P03.mp4,00:34,00:51,OrderDesk Classic - Returns queue,"Then I look up each order in this one, one at a time.",frames/legacy_system_h2_mid.png,customer email addresses,,checked
- Field mapping:
| Field | Source | If unavailable |
|---|---|---|
| participant | the ID in the file name | ask the user |
| category | the search the hit came from | none |
| session_file | the clip | none |
| start_mmss, end_mmss | the contact sheet in Step 5, MM:SS in the clip | the window start and end, marked approximate |
| window_title_read | the full frame Claude looked at | UNCERTAIN |
| quote_local | transcript lines inside the window | blank |
| frame | path of the middle frame | none |
| sensitive_on_screen | what kind of private data the frame shows | no |
| journey_step | left empty for the designer | blank |
| status | checked, or candidate when a frame could not settle it | candidate |
- Show the user two sample rows and the mapping, and wait for a yes before writing the full table. Say which fields are inferred (category and window_title_read) and what changes if Claude read the frame wrong: the designer draws a step in the wrong tool.
- Write current_workflow_evidence.csv. Add a short note under it with the counts per kind (checked and rejected) and the cells that are empty (a participant with no row for a kind). An empty cell does not prove the participant never uses that kind of tool; it only means it was not on screen in this clip.
In our test run the table had 13 checked rows from the searches and the note listed the empty cells; the user's approval of the sample rows was not exercised in our test run.
Hand the table to the user. Sharing it with the designer (a shared drive, a doc, a chat) is the user's step; if they ask Claude to send it, name the destination, show what will be sent, and wait for a yes. Frames with sensitive_on_screen set stay in the working folder unless the user blurs them first. Sending was not exercised in our test run.
Done when current_workflow_evidence.csv exists with one row per checked hit plus the count note, and the user has approved the sample rows.
Compliance
- Use only interviews whose consent covers screen recording and processing by a cloud service. Confirm this in Step 1, before anything is uploaded.
- Participants are listed by ID only. The table describes tools and windows, never the participant (no "slow", "confused" or similar), and the skill does no face, voice or identity recognition.
- Customer data, email addresses, account numbers and tokens on the participant's screen are described by kind, never copied. Frames stay in the working folder; blur them before they go into any report.
- Another vendor's screens shown by a participant are for internal research only; do not put those frames in anything public.
- Minors: B2B interviews rarely include them; if one does, use it only with guardian consent, otherwise leave it out.
- The clips stay in the user's Vivu project until the user deletes them. The skill never calls vivu_delete_video or vivu_delete_project unless the user asks, and confirms first.
- Nothing is posted or sent. The table stays in the working folder until the user hands it on.
Known failure modes
| Symptom | Cause | Fix |
|---|---|---|
| (observed) an internal report page comes back as an admin tool; in our test run the first legacy_system wording returned 1 false of 4, with a reason "showing the Ops Insights fulfillment dashboard used to monitor regional metrics" | a wording that asks for an older system matches on what the participant says about the place, not on the screen | describe the visible layout (left navigation menu next to records or a form), as in the table, and check the middle frame |
| (observed) a reason gives a wrong window title; in our test run one reason said "Shipland Admin portal" for a page titled Shiplane Admin | the reason text is a model description, not a reading of the title bar | read every title from the full resolution middle frame |
| (observed) the start frame shows the gallery or the previous window, and the end frame shows the next one | the window starts a couple of seconds early and runs into the next window; in our test run the median spreadsheet window was 21.1 seconds | judge on the middle frame and set times on the Step 5 contact sheet |
| (observed) a clip stays "queued" while the others index | indexing waits in a queue | keep polling vivu_list_videos; do not upload the file again |
| the video call gallery or a slide with a table is listed as a spreadsheet | grids of tiles and tables look alike at a glance | reject hits whose frames show the meeting tool's own screens; count them |
| a quote does not match what was said | it was copied from the reason text, which paraphrases | take every quote from the transcript lines inside the window |
| a kind has no rows for a participant who surely uses it | the window was too short or too small on screen, or the search missed it | say so in the note; look through that clip's demo part on a contact sheet before telling the designer the tool is absent |
| a list stops at exactly maximum_results | the cap cut the list short | raise maximum_results and rerun that search |
| "has not granted vivu.write" | Vivu connected read only | the user reconnects 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 clip is rejected by the browser upload tool | Claude in Chrome takes at most 10 MB per upload call | the user adds that file in the Vivu web app; never split or recompress it |
| a result's file name does not match sessions.csv | Vivu replaced spaces or punctuation in the name | rename to letters, digits and "_" before upload, and match on the name |