---
name: agm-proposal-response-log
description: "Turn a proxy season's AGM webcasts into a log with Vivu: per stewardship theme, the proposal or question, the board's reply, the speaker, preliminary votes, a timed link. Use after AGMs."
---AGM proposal and response log
This skill takes the official recordings of a proxy season's annual general meetings (usually three to six companies, from each company's own YouTube channel or investor relations page) and the three or four stewardship themes an asset manager's voting team tracks (for example executive pay, climate and emissions, board independence and shareholder rights), and returns one AGM log: for each company and theme, the start and end time where a proponent presents a shareholder proposal or a shareholder asks a question, how the chair or a director answered, the sentence quoted from the captions, who was speaking as shown on the name caption or named by the chair, the preliminary vote numbers read from the results slide or announced at the meeting, the frame that shows it, a link to the official recording at that second, and an empty follow up column for the engagement team. Claude cuts each meeting down to the proposal, question and voting items, prices that against the user's Vivu plan, indexes it in a private Vivu project, runs one precise search per theme and one for the results slides, reads the captions inside every returned window to keep only the windows that are really about the theme, reads speakers and vote numbers from frames, and writes the log. After the user approves a sample, Claude posts it to the team's Slack channel or Notion page.
The value is in what only the webcast holds. Minutes and the final vote filing arrive weeks later and carry the resolution and the count, not what was said: whether the chair answered a climate lobbying question with a commitment to disclose or with "the board's response is set forth in the proxy" is only in the recording. Preliminary results are often shown on a slide or read aloud at the meeting and nowhere else that day. Automatic captions have no speaker labels, and many proponents speak over a still card, so who spoke comes only from a name caption or the chair introducing them. The hard part is that a moderator reads shareholder questions back to back (an LNG project, then climate, then dividends, then a director's conflicts of interest), so a window found for one theme can sit on the neighbouring question. Every window Vivu returns stays a candidate until the captions inside it show the theme, and quotes come from the captions, never from Vivu's reason text.
When to use
Use when someone on a stewardship, proxy voting or ESG engagement team says "log what was said on climate and pay at this season's AGMs for our top holdings", "how did the board answer the simple majority proposal, and what were the preliminary votes?", "pull every shareholder question on director independence from these five annual meetings", or "build the post AGM engagement sheet before the minutes come out".
Not for these:
- A one off question about one moment in one meeting ("when did the chair answer the lobbying question?"): search Vivu directly and read the captions around the hit.
- Final vote counts. The log records preliminary numbers shown or read at the meeting; the company's later filing is the record.
- Voting recommendations, rating a board's answer or drafting engagement letters. The log puts what was said and shown side by side and nothing more.
- Identifying people by face or voice. Speakers come only from the on screen name caption or the chair or moderator naming them.
- Earnings calls and investor days. The earnings-commentary-drift-memo skill tracks one company's commentary across quarters.
Working principles
- Report measured numbers, not estimates. When a number is an estimate, say so.
- Nothing is verified until it has been checked against the source. Every Vivu window is a candidate until the captions inside it show the theme. Quotes come from the captions; vote numbers come from a full frame or the spoken announcement; never from Vivu's reason text, which paraphrases.
- Stop and tell the user when a required capability or tool is missing. Do not guess around it.
- Ask the user before anything that is expensive to redo (the meeting list and its index minutes, before upload) and before anything that acts on their behalf (posting the log to Slack or Notion; one rendered sample first).
- Never fill a speaker from a guess. When neither a name caption nor the chair's or moderator's words identify the speaker, the speaker cell stays empty and says UNCERTAIN.
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 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 user reconnects Vivu and allows write access |
| A shell on the user's computer with yt-dlp, ffmpeg, ffprobe and node | download the captions and the meeting, cut it, measure it, extract frames | yt-dlp --version, ffmpeg -version, ffprobe -version, node --version |
| A residential IP, when the recordings are on YouTube | YouTube blocks downloads from cloud and datacenter IPs | the caption download in Step 3 |
| Captions for each meeting (YouTube's automatic captions are enough) | the captions find the agenda items, decide each window and supply the quote | Step 3 downloads them; if a meeting has none, ask for the company's transcript or a local speech to text tool such as whisper.cpp, and mark quotes NOT VERIFIED until one is available |
| An upload path | move the files into Vivu | vivu_open_upload_page plus the user's own browser, or a browser tool that can attach local files |
| Slack or Notion connector with access to the destination (only for Step 9) | post the approved log | read the channel or open the page once |
| The company list, the recording links and the theme list | themes drive the searches; the links are the source of the video and of the log's link column | the user pastes them |
This skill needs Claude Code on the user's computer (the terminal or the Code tab of Claude Desktop), because it downloads and reads local files and runs ffmpeg. Claude on the web and cloud sessions cannot reach the files, and YouTube blocks their IPs. Nothing recurs, so no scheduler is involved; run it again after the next meeting and add the new files to the same project.
Inputs to collect
Ask for anything missing, most important first.
- THEMES: three or four themes in the team's words (for example "executive pay", "climate and emissions, including climate lobbying", "board independence and shareholder voting rights"). Required.
- MEETINGS: company name, meeting date and the official recording link for each AGM. Required.
- The proposal, question and voting items in each recording as start and end times. Default: found in Step 3 from the recording's description and captions ("stockholder proposal", "agenda item", "the next question is from", "voting results", "proxies").
- DESTINATION: the Slack channel or Notion page for the log. Default: none; the log stays in the working folder.
- Optional control theme: something the team expects nobody raised, to see whether the search invents it. Default: none.
- WORK_NAME: a short name for the working folder and the Vivu project. Default: the season, such as agm-2024.
Files and state
Keep everything in one working folder:
agm-WORK_NAME/
config.json themes with their query, meetings, cut times, Vivu project id, destination
meetings.csv file, company, meeting, source_url, cut_start_s, duration_s, video_id, caption_file
video/ the cut agenda items (.mp4)
full/ full downloads used only for cutting (delete after Step 3 if space is short)
subs/ caption files (.srv1, timestamped)
results/ one JSON per search, without the result page link
frames/ frame strips the speaker and vote numbers were read from
candidates.csv every window Vivu returned, with the caption verdict and the reason
agm_log.csv one row per kept window
agm_log.md the log as it will be posted
state.json steps done, files uploaded, searches with job ids, windows checked, post status
Run the commands from inside agm-WORK_NAME/ so the relative paths resolve.
A rerun reads state.json first and skips finished steps: a downloaded file is not downloaded again, a file marked uploaded is not uploaded again, a theme with a saved result is not searched again, a window with a recorded verdict is not checked again, and a log marked posted is not posted again. A meeting added later is uploaded to the same project and every theme is searched again, keeping only windows on the new video.
Step 1: Check the Vivu connector and the setup
Goal: confirm every requirement before spending anything.
- 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 the account does not showcan_create_projects: true, or a later write call fails with "has not granted vivu.write", ask the user to reconnect Vivu and allow write access, then stop until they have. - Run
yt-dlp --version,ffmpeg -version,ffprobe -versionandnode --version. - If the user wants the log posted, check the Slack or Notion connector by reading the channel or opening the page once.
- Ask for the themes and the meeting links if they are not in the request.
In our test run vivu_get_account returned can_create_projects: true, and yt-dlp, ffmpeg and ffprobe ran in the later steps. The Slack or Notion connector check was not exercised in our test run, because nothing was posted.
Done when vivu_get_account shows can_create_projects: true, the four tools print their versions, and the user has given the themes and the meeting links.
Step 2: Turn themes into queries and confirm the rights
Goal: one search query per theme, one for the results slides, and the user's confirmation that the recordings can be used.
- Rewrite each theme as a description of someone at the meeting raising or answering it: "a shareholder, a proponent or a board member talks about THEME_IN_PLAIN_WORDS: DETAIL_A, DETAIL_B or DETAIL_C". Name the two or three concrete things that make the theme its own (for executive pay: the remuneration report, fixed and variable pay, tying pay to targets), because the search matches what is said and neighbouring questions share words.
- Say who may be speaking ("a shareholder, a proponent or a board member") instead of a name. Speaker identification by Vivu is untested, so the query names a role, not a person, and a name in the query invites a reason text that repeats it.
- Keep the team's own wording in config.json next to each query; the log shows the team's wording.
- Confirm the Compliance items with the user: each company's terms for its webcast, and YouTube's terms when downloading from YouTube.
In our test run we wrote the queries ourselves from a stewardship theme list; the user's approval of the queries and the rights confirmation in Step 2 were not exercised in our test run (internal testing only).
Done when the user approves the query list and confirms the Compliance items, and both are saved in config.json.
Step 3: Download the captions, find the items, cut them
Goal: only the proposal, question and voting items of each meeting, with timestamped captions, in the working folder. A whole AGM often runs two hours or more, and indexing all of it spends minutes on speeches and procedure.
Run on the user's computer, never in a cloud sandbox. Get the description and the captions first; srv1 captions put each line's start second in the file, which makes finding the items a plain text search:
yt-dlp --js-runtimes node --extractor-args "youtube:player_client=web_embedded" --skip-download --write-description --write-auto-subs --write-subs --sub-langs "en,en-orig" --sub-format srv1 -P "home:subs" -o "%(id)s.%(ext)s" "https://www.youtube.com/watch?v=VIDEO_ID"
VIDEO_ID is the 11 character ID in the recording link. Find the items with a grep for the words chairs and secretaries use (TERM is one of proposal, agenda item, next question, voting results, proxies):
grep -o -i '<text start="[0-9.]*" dur="[0-9.]*">[^<]*TERM[^<]*' subs/VIDEO_ID.en.srv1
Write down the start and end second of each item: a proposal runs from its introduction ("the next stockholder proposal is") to the chair moving on; a question from "the next question is from" or the questioner's first words to the end of the answer; the vote item from "the voting is now closed" or "the proxy results are shown on the screen" to the last result. Add 10 seconds at both ends.
Then download the full recording at 720p, which keeps name captions and result tables readable, and cut each item with re-encoding so the cut starts exactly at the second you chose:
yt-dlp --js-runtimes node --extractor-args "youtube:player_client=web_embedded" --restrict-filenames -f "bv*[height<=720]+ba/b[height<=720]" --merge-output-format mp4 --download-archive archive.txt -P "home:full" -o "%(id)s.%(ext)s" --retries 5 "https://www.youtube.com/watch?v=VIDEO_ID"
ffmpeg -v error -ss START -to END -i full/VIDEO_ID.mp4 -c:v libx264 -preset veryfast -crf 23 -c:a aac video/agm_COMPANY_ITEM_VIDEO_ID.mp4
START and END are seconds in the full recording; COMPANY and ITEM are short labels such as acme2024 and qa. A file cut at START has its own clock: file time plus START is the time in the official recording, so record cut_start_s in meetings.csv. When the company's investor relations page offers an mp4 instead, download it there and cut it the same way. Download one meeting at a time; a second download from the same video while the first runs can fail with "HTTP Error 403: Forbidden".
The youtube-competitor-watch skill (https://vivu.ai/skills/youtube-competitor-watch) covers channel verification and download troubleshooting in more depth.
In our test run the grep found the stockholder proposal block of a US carmaker's annual meeting, the question, pay and results items of a Dutch insurer's AGM, and the moderated questions and formal resolutions of an Australian energy company's virtual AGM. A short edited AGM video from another company had no question or vote item and was left out. Downloading only a section with yt-dlp's --download-sections ran slower than the full download plus the ffmpeg cut, and one section was offered only at low resolution; its proxy tables were still readable.
Done when video/ holds one mp4 per item, subs/ holds an srv1 caption file per meeting, and meetings.csv lists file, company, source_url and cut_start_s for each.
Step 4: Price, upload and index
Goal: every item indexed in a private Vivu project, with the user's approval of the cost.
- Measure each file:
ffprobe -v error -show_entries format=duration -of csv=p=0 video/FILE, where FILE is a file name in video/. Write duration_s. - Call vivu_get_usage and show one table:
| These meetings | Plan allowance | Remaining this month | |
|---|---|---|---|
| Index minutes | sum of duration_s / 60 | from vivu_get_usage | from vivu_get_usage |
| Search credits | 5 per theme plus 5 for the results slides (estimate), plus 5 for each rewording and for the control | 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 and 500 search credits. A precise search uses 5 credits in total. For example, six AGMs cut to about 20 minutes each fit in 180 index minutes; six whole meetings do not, which is why Step 3 cuts them.
- If the meetings do not fit, offer levers in this order: cut tighter to the proposal and question items (drop procedure, speeches and breaks), index the meetings of the largest holdings first, and only then move to a larger plan. Never drop a meeting the user asked for without saying which one, and never recompress a file to save minutes.
- Call vivu_list_projects and reuse a project named "AGM log WORK_NAME" if one exists; otherwise call vivu_create_project with that name and visibility "private". The default visibility is organization, which shows the project to everyone in the user's Vivu organization; a voting team's engagement notes belong in a private project.
- Call vivu_open_upload_page with the project ID immediately before uploading. The link expires in 180 seconds and works once, so never post or store it. Give it to the user to open in their own browser and select the files, or open it in a browser tool that can attach local files. Claude in Chrome accepts at most 10 MB per upload call; a cut item is usually larger, so those go through the user's own browser or the Vivu web app. Never split or recompress a file to fit.
- Poll vivu_list_videos until every file shows ready. Vivu replaces spaces and punctuation in file names with underscores; with --restrict-filenames and the names above they already match. Record each video_id in meetings.csv and state.json.
In our test run the six cut items measured 51.84 minutes in all and were ready about 11.5 minutes after the upload started; the names in vivu_list_videos matched the local files. The account used in our test run returns no allowance figures, so the comparison with a real plan and the user's approval were not exercised in our test run. The files went through the upload page with an automated browser, so opening the link in the user's own browser was not exercised in our test run.
Done when the user has approved the index minutes, and every file in meetings.csv shows ready and has a video_id.
Step 5: Search each theme and the results slides
Goal: a saved list of candidate windows for every theme and for the results slides.
One precise search per theme, because each theme is its own section of the log, plus one for the results slides. Every search is precise: only precise returns a time window, and fast returns whole files with an empty reason, which tells you nothing when every file in the project is an AGM. The chain behind each row: the precise search suggests windows where someone raises or answers the theme, the captions inside each window decide whether the theme is really discussed there and give the quote (Step 6), and frames from the same window give the speaker's name caption (Step 7). The results slide search is the flip from what is said to what is shown: the table on screen is where the preliminary numbers are, and Claude reads them from a full frame.
Run vivu_search_videos (project_id, query, mode "precise", maximum_results 10) for each query. Ten covers the places one theme comes up across a season's cut items (a proposal, its board response, one or two questions and their answers, the vote) with room for windows Step 6 sets aside; it is also the recall ceiling, so raise it for more meetings or longer items. vivu_search_videos 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 the results to results/KEY.json without the result page link, and show the result page link only in the live reply: it expires after four hours, so it never goes into the log, candidates.csv or a post.
The queries from our test run, on the cut items of three companies' AGMs, show the pattern:
| Field | Query | Mode | maximum_results |
|---|---|---|---|
| exec_pay | a shareholder, a proponent or a board member talks about executive pay: the remuneration report, executives' fixed and variable compensation, bonuses, or tying executive pay to performance or sustainability targets | precise | 10 |
| climate | a shareholder asks the company about climate change, greenhouse gas emissions or climate lobbying by its industry associations, or the chair or chief executive answers such a question | precise | 10 |
| board_rights | a shareholder raises a concern about how the company's own board is run: whether a sitting director is independent or has a conflict of interest with another company, how directors are elected, or replacing supermajority voting rules with a simple majority. Not a proposal about an environmental or sourcing policy. | precise | 10 |
| vote_screen | a slide or table on screen showing the voting results or proxy votes for each resolution, with numbers or percentages of votes for and against | precise | 10 |
| control (optional) | a shareholder asks about the plan for CEO succession and who will take over from the current chief executive | precise | 10 |
The board_rights wording is the second one, after 1 rewording on the same corpus; the other queries are first wordings. The last sentence of the board query keeps out proposals on other subjects that mention a director in passing.
A precise search uses 5 credits in total.
In our test run the six precise searches came to 30 credits (estimate, from the credit rule), and every search was complete on the first status call.
Done when every theme and the results slide query have a saved result file and their job ids are in state.json.
Step 6: Check every window against the captions
Goal: each window labeled "on theme" or "other", with the deciding caption line, before anything reaches the log.
- For each window, convert start_ms and end_ms to seconds and add cut_start_s from meetings.csv: the srv1 file is on the official recording's clock. Read the caption lines that start inside the window, plus 60 seconds before it, because a window can start after a question has been read:
grep -o '<text start="SECONDS_PATTERN[0-9.]*" dur="[0-9.]*">[^<]*' subs/VIDEO_ID.en.srv1
SECONDS_PATTERN is a pattern for the whole seconds in the range; for 1880 to 1999 it is 1\(8[89]\|9[0-9]\)[0-9]. Use two patterns joined with \| when the range crosses a hundred.
2. Label the window:
- on theme: someone raises the theme (a proposal, a question, a statement) or the chair or a director answers it;
- other: a neighbouring question or proposal, the same words used about something else (a sustainability proposal about workplace harassment is not climate), a procedural remark.
- Write every window with its label, the caption line that decided it and Vivu's reason into candidates.csv, and count the "other" windows as false positives for the theme. Vivu's reason text never decides the label.
- When a moderator reads questions back to back, check the line where the question is read: the theme belongs to the question, not to the window's position in the run.
- Before writing the log, grep each item's captions for the theme's own words (for example "remuneration", "compensation", "climate", "emission", "director", "majority") and read every hit that is not inside a kept window for that theme. This finds some missed statements, not all.
Worked example from our test run
The test corpus was six cut items from the official AGM recordings of a US carmaker (the stockholder proposal block), a Dutch insurer (shareholder questions, the remuneration report item, the voting results) and an Australian energy company (written questions read by a moderator, the formal resolutions): 51.84 minutes in all. Before searching we marked in the automatic captions every proposal, question and answer on the themes, passing mentions, neighbouring material that shares words with a theme (proposals on workplace harassment and on deep sea mining, questions on biodiversity, dividends and an LNG project), the result tables, and a control theme nobody raised.
On the final wordings the theme searches returned 19 windows: 18 real and 1 false, with 0 of the marked proposals, questions and answers missed; the board wording came after 1 rewording on the same corpus. Passing mentions did not count toward recall, and the board searches returned none of the ones we had marked, including the carmaker's spoken preliminary result on both of its board proposals; the climate search missed a deep sea mining sentence about carbon stores. The spoken result in the log came from reading the results announcement in the captions, not from a search. Executive pay returned 10 windows, all real: the remuneration report presentation, several windows on one proposal's statement, and the preliminary results. Climate returned 6, 5 real, including the climate question the moderator read between a question on an LNG project and one on dividends; neither neighbour came back. The 1 false was a co-filer of a harassment proposal saying her fund focuses on investing in climate solutions. The first board wording returned 4 windows with 1 false (25%), a deep sea mining statement about a former director now advising a mining company; the second returned 3, all real. The results slide search returned 3 windows, all real, covering every result and proxy table, one of them a single long window across three tables; the control theme returned 0. Median windows were 29.5 seconds for pay, 48.5 for climate and 25 for the board. The window on the chief executive's lobbying answer started after the answer had begun, and the window on the director conflict question ended before the chair answered it, which is why the caption check reads a minute before each window and past its end when an answer follows.
Done when every window in results/ has a label and a caption line in candidates.csv, and the grep has been run for each theme.
Step 7: Read speakers and vote numbers from frames
Goal: for each kept window, who was speaking and, for results windows, the numbers on screen.
- Cut three frames spread across the window into one strip:
ffmpeg -v error -y -ss WIN_START -t WIN_LEN -i video/FILE -vf "fps=3/WIN_LEN,scale=640:-1,tile=3x1" -frames:v 1 frames/KEY_hN.png
WIN_START is start_ms / 1000 and WIN_LEN is (end_ms minus start_ms) / 1000, both in seconds in the file; FILE is the cut file in video/; KEY is the query key and N the window's number in its result file. 2. Speaker: write the name and title exactly as a name caption shows them. Many AGMs show none: proponents often speak over a still "Shareholder Proposal" card, and a moderator reads written questions off screen. Then use the chair's or moderator's words in the captions ("the next question is from", "whose representative has prepared a pre-recorded message") and write the role (proponent's representative, shareholder, chair, chief executive). Private shareholders who ask questions are logged as "shareholder", not by name. If nothing identifies the speaker, write UNCERTAIN; never guess from the face, the voice or the seat. 3. Vote numbers: results tables are small text. For each results window, extract one full frame every half second and pick the first frame where the whole table is visible (tables fade in):
ffmpeg -v error -y -ss WIN_START -t WIN_LEN -i video/FILE -vf "fps=2" frames/KEY_hN_%03d.png
Read the for and against figures from that full frame, never from the strip or from the reason text. When the chair or secretary also reads the figures aloud, compare the frame with the captions and write vote_source "slide and spoken" when they agree. A number you cannot read clearly stays NOT VERIFIED. 4. When a meeting has no results slide, grep its captions for "preliminary", "in line with the board" and "not in line", read the announcement, and copy each proposal's spoken result into that proposal's row with vote_source spoken.
In our test run name captions appeared only at the Dutch insurer (the chief executive, the remuneration committee chair and the company secretary). The carmaker's proponents spoke over a still card and the energy company's moderator read written questions off screen, so every speaker there came from the introductions in the captions, and the person announcing the carmaker's results was never named on screen or in the captions (UNCERTAIN). The carmaker showed no results slide: its preliminary results were spoken, without figures. Every for and against figure on the three tables was readable from the frames, including a low resolution proxy table, and matched the figures read aloud wherever the captions had them right.
Done when every kept window has a frame strip in frames/ and a speaker value (or UNCERTAIN), and every results window has its numbers read from a full frame or marked NOT VERIFIED, and each proposal at a meeting without a results slide has its spoken result from the captions.
Step 8: Write the AGM log
Goal: a log the team can read in two minutes and check from.
- Show the user one sample row and the field mapping, and wait for a yes before writing the rest:
company,meeting,theme,segment,start,end,item,speaker,speaker_source,quote,response_summary,vote_result,vote_source,frame,source_link,transcript_checked,follow_up
Dutch insurer,AGM 2023-05-25,climate and emissions,answer,03:53,04:55,Answer on lobbying disclosure,chief executive,name_caption,"again no objection to disclosing the memberships but an exhaustive [list] I cannot guarantee",Will disclose association memberships; cannot guarantee an exhaustive list of their lobbying,,,frames/climate_h2_strip.png,https://www.youtube.com/watch?v=oyraKKNNxoQ&t=2673,yes,
| Field | Source | If unavailable |
|---|---|---|
| company, meeting | meetings.csv | none |
| theme | the team's wording from config.json | none |
| segment | proposal, question, answer, board response or vote, from the captions | none |
| start, end | the window in the file, as MM:SS | none |
| item | the proposal or question in a few words, from the captions | none |
| speaker, speaker_source | Step 7: name_caption or chair_call | UNCERTAIN |
| quote | copied from the captions, with obvious caption errors marked in brackets | NOT VERIFIED when there are no captions |
| response_summary | one line from the captions on how the chair or board answered; "no spoken response" when the chair only refers to the proxy statement | none |
| vote_result, vote_source | Step 7 for results rows, or the spoken result from the captions on a proposal row when the meeting showed no results slide; blank for other rows | NOT VERIFIED |
| frame | path of the strip or full frame | blank |
| source_link | the official recording link with t equal to cut_start_s plus start, in seconds | file name and time |
| transcript_checked | yes when Step 6 labeled the window from the captions | no |
| follow_up | left empty for the team | stays empty |
- Write agm_log.csv, one row per kept window, and agm_log.md, one section per company with one line per row (theme, time, speaker role, item, response, votes, link). Add a line for each theme not found in a company's items and for the control.
- Tell the user the counts: rows written, windows set aside in candidates.csv, themes not found per company, rows with speaker UNCERTAIN, vote numbers NOT VERIFIED.
In our test run the log has one row per kept window on the final wordings, with a source link on every row, a caption quote on every row except one results row where nothing on the theme was said, and the vote figures on every results row. The user's approval of the sample row was not exercised in our test run.
Done when agm_log.csv and agm_log.md exist with one row per kept window and a line for each theme not found, the user has approved the sample row, and the counts are reported.
Step 9: Post the log to Slack or Notion
Goal: the approved log in the team's channel or page, posted once.
- Name the destination (the Slack channel or the Notion page) and show the full rendered message or page exactly as it will be posted. Posting acts as the user, so wait for an explicit yes.
- Slack: one message, one section per company, one line per row with the theme, time, speaker role, the response in a few words, the preliminary votes and the official recording link with its time. Notion: a page with agm_log.md as the body and the CSV rows as a table.
- Use the official recording links, never a Vivu result page link (it expires after four hours) or a local file path.
- Record the post in state.json so a rerun does not post it again.
Posting was not exercised in our test run; Step 9 rendered the Slack message and the Notion page as files and sent nothing.
Done when the user has approved the rendered sample, the log is posted to the named destination, and state.json records it as posted.
Compliance
- Use only recordings the company itself published (its investor relations page or official channel). Webcast terms differ by company and some forbid downloading; check them before Step 3. Keep downloads for the team's internal analysis and do not redistribute the video or clips.
- Downloading from YouTube may conflict with YouTube's Terms of Service. Have the user confirm this is acceptable for their organization before Step 3, or use the company's own mp4 when it offers one.
- Download to the user's computer and upload from there. The skill does not hand video links to a third party importer.
- The people in the recording are directors, executives, proponents and shareholders speaking at a shareholder meeting. Speakers come only from the on screen name caption or the chair's or moderator's words; the skill does no face or voice recognition. Private shareholders who ask questions are logged by role, not by name.
- The log lists what was said and shown, side by side. It does not recommend votes, rate a board's answer or draft engagement messages.
- The videos stay in the user's Vivu project until the user deletes them. The project is created private because the default visibility shows it to the whole Vivu organization. Delete a project or video only when the user asks, and confirm first.
- Posting to Slack or Notion acts as the user and needs an approved sample first (Step 9).
Known failure modes
| Symptom | Cause | Fix |
|---|---|---|
| (observed) yt-dlp prints "ERROR: unable to download video data: HTTP Error 403: Forbidden" or "Error opening input files: Server returned 403 Forbidden (access denied)" | downloads from YouTube ran at the same time | download the meetings in turn with the web_embedded player client, as in Step 3; retried alone, it worked |
| (observed) a climate window sits on a proposal about something else | one phrase in a workplace harassment proposal, "investing in climate Solutions", matched the theme | Step 6 labels it "other" because the rest of the statement is about harassment; count it as a false positive |
| (observed) a board window sits on a deep sea mining proposal | the statement mentions a former director advising a mining company | the second board wording in Step 5 keeps out proposals on other subjects; Step 6 still checks every window |
| (observed) the captions show a vote figure as "99 0.04 percent" | automatic captions garble spoken decimals, and secretaries misspeak and correct themselves | read vote figures from a full frame of the results table; use the captions only to cross-check |
| (observed) no name caption on any frame of a window | proponents speak over a still "Shareholder Proposal" card, and moderators read written questions off screen | take the speaker from the introduction in the captions; otherwise UNCERTAIN |
| (observed) a results window covers several tables | a meeting's result tables follow each other on screen | extract frames every half second across the window and read each table from its own frame |
| (observed) a meeting has no results window | the chair announced the preliminary results aloud, without a slide or figures | log the spoken result with vote_source spoken and leave the figures NOT VERIFIED until the filing |
| a window for one theme sits on the neighbouring question | a moderator reads questions back to back with shared words | Step 6 labels it "other" from the line where the question is read; reword the query with the theme's own details |
| a reason text names a speaker or states a vote figure | the reason paraphrases and can be wrong | never copy it; quote the captions, read the speaker from the caption or the chair's words, read votes from a full frame |
| a theme gets no window for a company | the theme did not come up, or people used other words | grep the captions for the theme's words; an empty result does not prove it was never raised |
| "Sign in to confirm you're not a bot" | download attempted from a cloud IP | run Step 3 on the user's computer |
| 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 cut file is rejected by the browser upload tool | Claude in Chrome accepts at most 10 MB per upload call | the user opens the upload link in their own browser or adds the file in the Vivu web app |
| "has not granted vivu.write" | Vivu connected read only | the user reconnects Vivu with write access |