wrote a bash + vim workflow that pulls youtube video transcripts into buffers and searches them with fzf
i work at a devops consultancy and we have about 180 youtube videos. recorded architecture reviews, internal tech talks, client postmortems, vendor integration demos, conference presentations people bookmarked. they're all shared in a markdown file in our wiki which is basically a list of youtube links with dates. useless for finding anything unless you remember exactly when the video was recorded.
i wanted a way to search these videos by what was actually said in them without leaving the terminal. so i built a workflow around vim, bash, and fzf.
the first piece is a bash script that takes a youtube url and pulls the full transcript using transcript api. it saves the transcript as a plain text file named after the video title with the date prepended. one file per video. all the transcripts live in a directory called ~/transcripts.
the second piece is an fzf wrapper script. it runs fzf with --preview against the transcript directory. as you type, fzf fuzzy matches across all transcript files and the preview window shows the matching file with the match highlighted. select a result and it opens the transcript in vim with the cursor on the first match. i bound this to a key in my shell so i can hit ctrl-t and start searching immediately.
the vim side is where it gets useful. each transcript file has a yaml front matter block at the top with the video title, date, speaker, tags, and the youtube url. i wrote a small vim function that reads the url from the front matter and opens it in the browser. so the workflow is: fzf to find the video, vim to read the transcript, one keypress to open the actual youtube video if i need to watch it.
the ingestion script is about 30 lines of bash. curl to call the api, jq to parse the json and extract the transcript text, a few lines to generate the yaml front matter, and tee to write the file. i have a text file with all 180 urls and a for loop that processes them. the whole batch ran in about 4 minutes.
the fzf wrapper is maybe 15 lines. the vim function is 8 lines. the keybinding is one line in my vimrc.
about 180 videos indexed as plain text files. the consultants on my team use it before client calls to search for whether we've discussed a similar architecture before. i use it almost daily to find specific things from recorded tech talks. the nice thing about plain text files is that grep works, ripgrep works, fzf works, vim's built in search works. no special tooling needed beyond what's already on my machine.
the whole thing took an afternoon and i haven't changed anything since.
