Get the transcript of this talk and list every benchmark number it quotes, with the timestamp for each.What comes backA short table of figures, each linked to the minute it was said, so you can jump straight to the claim.
Felo YouTube Subtitling on Pi
Pi Coding Agent
Pi Coding AgentA harness with no browser pane can still watch the talk
Pi keeps its core small enough that video never enters the workflow — and the sentence you need is often forty minutes into a recording. This pulls the captions so the session can search them like any other text file.
Copy it in
cp -r felo-skills/felo-youtube-subtitling ~/.pi/agent/skills/- Bundlefelo-skills/felo-youtube-subtitling
- InputA YouTube URL
- Comes back asTimestamped lines
- ThenOrdinary text search
Pi keeps a deliberately small core and asks you to extend it, so a skill is the intended way to add reach without adding tools to maintain.
- Loads from
- ~/.pi/agent/skills/ · .pi/skills/ · packages
- Install is
- A folder copy, or pi install npm:@scope/pkg
- Worth knowing
- No sandbox — read the SKILL.md before you mount it
Timestamped lines, so a claim can be checked against the minute it was said
What comes back is text the session can search, quote, and disagree with. The timestamps are what make it useful later: a quotation without one cannot be verified by anyone, including you.
04:12Maren
We moved the scheduler off the monolith in the spring, and the first thing we lost was the retry budget.
11:38Maren
The p99 did not move at all. What moved was the time we spent waiting on the queue.
19:05Host
So the win was operational rather than a latency win.
26:44Maren
Exactly. Six weeks of work to make an incident stop happening every quarter.
33:20Maren
If you take one number away from this talk: fourteen pages a quarter became two.
The p99 did not move at all. What moved was the time we spent waiting on the queue.
From a shared recording to notes somebody can use
The order matters: get the text first, decide what matters second, and write it down third.
Runs when
Someone sends a recording and the useful part is somewhere past the first ten minutes.
- 01
Ask for the transcript, not a summary
A summary drops the numbers and the hedges, which are usually the reason you are watching. Get the timestamped lines and decide what matters yourself.
- 02
Search it like a file
Ask for every sentence containing a figure, a version, or a date. Once the captions are text, this is a text task — which is exactly the kind of task a small harness is good at.
- 03
Pull the two lines worth repeating
One that carries the argument and one that names the limit. Keep the timestamps attached; a quote without one cannot be checked.
- 04
Put it beside the decision it informs
A transcript that lives only in this session disappears with it. Write the relevant lines into the repository next to the work they justify.
Leaves behindTimestamped lines you can search, the quotes worth repeating, and a note of which claims the speaker never supported.
Add it on PiTranscripts are a normal bundle, so the install is a copy
Nothing about captions needs a special root. The one Pi-specific habit worth keeping is reading the bundle before it goes global, because a skill that reaches the network deserves a look.
- 01
~/.pi/agent/skills/GlobalEvery project on this machine. Put a bundle here once you have read it and want it everywhere.
- 02
~/.agents/skills/SharedThe folder other harnesses read too, so one clone serves Pi and whatever else you run.
- 03
.pi/skills/ProjectThis repository only, and only after you trust the project. The right place to evaluate a bundle you have not read yet.
- 04
pi install npm:@scope/pkgPackageA package's skills/ directory, kept versioned and updatable. The cleaner route once a bundle earns a permanent place.
Or load it for one session, without copying
pi --skill /path/to/felo-youtube-subtitlingTwo ways to use a talk you have not watched
Both assume the video is a source rather than entertainment, and both end with something checkable.
Does this recording say how they handled a rollback? Quote the answer with its timestamp.What comes backThe relevant lines with timestamps — or a plain statement that the topic never came up, which is also an answer.
Three limits worth knowing before you quote a video
Transcripts are evidence with known failure modes, and they are cheap to check.
Captions are not a transcript of intent
Auto-generated captions mangle names and numbers. If a figure is going into a decision, check it against the recording before you repeat it anywhere that matters.
No captions means no text
A recording without subtitles gives the skill nothing to fetch. That is a source problem, not something a longer prompt will fix.
Context is still yours to supply
The transcript says what was said, not whether the speaker's situation resembles yours. The skill removes the watching, not the judgement.
What Pi users ask about YouTube transcripts
Which videos work?
Public YouTube links that carry captions, including auto-generated ones. Private, unlisted-without-access and login-walled recordings are out of reach.
Can it watch a channel for new uploads?
It fetches a video you point it at. Watching for new uploads is a scheduling job rather than a transcript job, and on Pi that is a routine you would build out of something else.
Where should the transcript live afterwards?
In the repository, next to the decision it informs. Future sessions cannot see this one's memory, and neither can your colleagues.
Does this need a browser or a headless Chrome install?
No. That is the point on a minimal harness: the skill reaches the captions directly, so nothing has to open a page or keep a browser process alive.
Every Felo skill for Pi Coding Agent
Five skills written for Pi Coding Agent. They follow the same pattern on this platform, so you can combine them in one session without re-explaining your setup.
Felo Search
Felo Slides
Felo X Search
Felo SuperAgentOne API key. A full agent toolkit.
Start with a skill, then compose the exact models, live data, knowledge, and output tools your Pi 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
Live information
Give agents current web and social signals instead of relying on stale context.
- Web Search
- X Search
- Web Fetch
Knowledge and output
Ground work in reusable knowledge, then turn it into shareable deliverables.
- LiveDoc
- PPT generation
- Mindmaps
Creative tools
Extend the workflow from research to visual and office-ready results.
- Image generation
- Page tools
- Office outputs