Felo YouTube Subtitling on Doubao WorkDoubao Work

A recording is not a document until somebody writes it down

Briefings, earnings calls, and policy sessions arrive as video, and the deliverable is still a document. This pulls the captions as timestamped text, so the document can quote the minute instead of paraphrasing the hour.

  • Bundlefelo-skills/felo-youtube-subtitling
  • InputA YouTube URL
  • Comes back asTimestamped lines
  • Next stepDocument or spreadsheet
Built for Doubao WorkTimestampedFeeds a document
Get an API keySee the platform guide
The command underneathfelo youtube "<video url>"
Doubao Work logo
This skill runs on

Doubao Work

ByteDance's Doubao Work · agent work platform

Read the Doubao Work guide

Doubao Work turns a request into a finished file and can run it on a schedule. A Felo skill decides what goes into that file and keeps the sources behind every conclusion.

Runs on
Personal, team, and enterprise builds
Skills arrive by
Skill panel import, or the Felo CLI
Automation
Scheduled tasks and computer use
What you get back

Timestamped lines, so a quotation can be checked against the minute

The timestamps are not decoration. They are what lets a reader verify a figure later, and what makes a quotation usable in a document that somebody else will review.

08:40CFO

Revenue in the region grew 14 percent year over year, and about half of that came from the two new markets.

14:02Analyst

Is that growth from volume, or is it pricing?

14:31CFO

Volume, mostly. Pricing contributed about two points.

22:10CFO

We expect the same cadence next quarter, with the caveat that the comparison base gets harder from here.

31:55Host

That is the last question. Thanks everyone for joining.

Volume, mostly. Pricing contributed about two points.
Ask for figures to be pulled with their timestamps, and the number that ends up in the report can be traced back to the second it was said.
The routine

From a shared recording to a paragraph somebody can sign off

Four steps, and the order is the whole method: text first, judgement second, writing third.

Runs when

A recording lands in the chat — an earnings call, a briefing, a webinar — and the deliverable is a document.

  1. 01

    Get the transcript before the summary

    A summary written from a video is an interpretation. A summary written from text can be checked against the text, which is what makes review possible.

  2. 02

    Extract every figure with its timestamp

    Ask for each number and the minute it was said. This is the step that turns a recording into something a finance or legal reviewer will accept.

  3. 03

    Write the document against the quotes

    Have the task draft the section using the extracted lines, and keep the timestamps in a footnote or an appendix rather than deleting them.

  4. 04

    File it where the team reads it

    A transcript that stays inside one task is not a source. Put it in the workspace next to the document it supports.

Leaves behindA document section with traceable quotations, plus the extracted figures with timestamps and a note of what the speaker never addressed.

Add it on Doubao Work

Same two routes, with one extra step for the output

The install is the usual pair. The difference on this page is what happens after: the transcript is raw material for a document, not a deliverable on its own.

Path A

Import the skill bundle

The shorter route, when your build exposes a skill or connector panel. Personal, team, and enterprise builds differ here, so check what your own account shows before planning around it.

  1. Clone the repositorygit clone https://github.com/Felo-Inc/felo-skills.git. Every bundle is a folder with a SKILL.md, and you only need the one you are installing.
  2. Import the folderBring that folder in through the skill or connector panel so video transcripts inside a document task is available while the task runs, not only afterwards.
  3. Confirm it is callableAsk the agent which skills it can use before you build a scheduled task around one. An imported skill that never gets called is the usual failure here.
Path B

Drive the Felo CLI

The route that works on every build. Install it once, then either run it yourself during setup or let computer use invoke it as part of a task on your machine.

  1. Install and store the keynpm install -g felo-ai felo config set FELO_API_KEY your-api-key-here
  2. Run the command by hand oncefelo youtube "https://youtube.com/watch?v=..."
  3. Then fold it into the taskAsk the agent to call the CLI when it needs current information or a deliverable, so the same command runs inside the task instead of beside it.
Where it goes

Three places the transcript ends up earning its keep

The transcript is an intermediate artifact, and it is worth deciding its destination before generating it.

Report appendix

Quoted lines

Keep the timestamps. A reviewer who can jump to the minute stops arguing with your summary and starts checking the source.

Spreadsheet

Figure table

For anything with numbers: one row per figure, with the timestamp and the speaker, so a change in the model can be traced to a sentence.

Workspace document

Full transcript

File the whole thing where the team can search it. The next question about this recording should not need a new fetch.

In practice

Two prompts for a recording you have not watched

Both treat the video as a source rather than as content, which is what makes the output usable in a report.

Type thisGet the transcript of this earnings call and list every financial figure it gives, with the timestamp for each.What comes back

A table of figures anchored to minutes, ready to be checked before anything is published.

Type thisDraft the market section of the report from this briefing. Quote the parts that matter, and keep the timestamps.What comes back

A draft whose claims point back at the recording, with the speaker's caveats preserved rather than smoothed away.

Where it stops

Three limits worth stating in the document itself

Transcripts have known failure modes, and naming them is cheaper than defending a number later.

Auto-captions get names and figures wrong

If a number is going into a report, check it against the recording. Captions are good enough to search, not good enough to quote blind.

No captions means nothing to fetch

A recording without subtitles gives the skill no text. That is a source problem, and no prompt fixes it.

A transcript is not a fact-check

It tells you what was said, accurately to the captions. Whether the speaker was right is a separate question and a separate source.

FAQ

What Doubao Work users ask about transcripts

Can it handle a recording in another language?
It returns whatever captions the video carries. If those are auto-generated in the original language, translation is a separate step you can then run on the text.
Does this work for internal recordings?
Public YouTube links are the supported input. Internal calls are usually behind a login, which puts them out of reach for this route.
Should the transcript go into the deliverable?
Usually as an appendix or a footnote, not in the body. The body needs the argument; the timestamps are there so a reviewer can check it.
What if the recording is three hours long?
That is the case this is built for. You search the text for the four sentences that matter instead of watching for the one you need.
More on Doubao Work

Every Felo skill for Doubao Work

Five skills written for Doubao Work. They follow the same pattern on this platform, so you can combine them in one session without re-explaining your setup.

Felo SearchLive web research with sources you can citeFelo SlidesGenerate a deck from a single promptFelo YouTube SubtitlingTranscripts from talks, demos, and reviewsFelo X SearchTrack posts and replies on XFelo SuperAgentLong-running workflows that keep their memory
Beyond ready-made skills

One API key. A full agent toolkit.

Start with a skill, then compose the exact models, live data, knowledge, and output tools your Doubao Work workflow needs — the same key every Felo skill already uses.

Models

Use cost-efficient LLM access when your agent needs custom reasoning or generation.

  • LLM API
  • Chat completions
  • Model access
Explore LLM API →

Live information

Give agents current web and social signals instead of relying on stale context.

  • Web Search
  • X Search
  • Web Fetch
Explore live data tools →

Knowledge and output

Ground work in reusable knowledge, then turn it into shareable deliverables.

  • LiveDoc
  • PPT generation
  • Mindmaps
Explore knowledge tools →

Creative tools

Extend the workflow from research to visual and office-ready results.

  • Image generation
  • Page tools
  • Office outputs
Explore platform tools →
Explore API Platform →Read the API docs →

Keep reading

Felo YouTube Subtitling on its own pageThe skill, apart from any single platformThe Doubao Work guideDocuments, schedules, and computer useFelo Search for Doubao WorkWhen the source is a page rather than a recording