Yapture
[ OVERVIEW ]
Built a local tool that turns videos into transcripts, then files them into a queryable playbook. Those transcripts feed an AI agent that files each lesson into a playbook I can query when a project starts or stalls.
[ PROJECT SNAPSHOT ]
[ LOGO ]
[ TYPOGRAPHY ]
[ COLOUR PALETTE ]
Yap Cream #FFF6E9 · Ink #17150F · Yap Lime #C9F24D · Tangerine #FF6B35 · Moss #2E7D4F · Chili #D93B28. Flat colour only, no blues or gradients.
[ BRAND IN USE ]
Speech-bubble 'Y' mark, Yap Cream containers with hard ink borders, 3px offset ink button shadows, Tangerine in-progress accents, and flat status badges.
[ THE CHALLENGE ]
"EVERYONE SAVES VIDEOS THEY MEAN TO COME BACK TO. ALMOST NO ONE GOES BACK."
[ THE CONTEXT ]
Saved, never revisited
People save reels and videos full of good advice, then rarely go back to them. Video is slow to search and slow to skim, so going back means rewatching eight minutes to find the thirty seconds that mattered.
The fix is behind a paywall
The tools I tried demand a sign-up, cap you on credits, or run a 30 to 60 second ad before handing anything over. Enough friction to kill a daily habit.
A transcript is not a playbook
Even with the text in hand, raw transcripts pile up the same way the videos did unless something files them where they can be found again.
[ WHAT WE WERE UP AGAINST ]
"Every free transcription tool put a login, a credit limit, or an ad between me and the transcript, which slowed down a habit I wanted to do daily."
[ WHAT I DID ]
MOVE 01 / SCOPE
Built the tool the market would not
Rather than work around the limits, I specified what a transcription tool should do with none of them, then built it. Any YouTube or social link, no sign-up, no credit cap, no ad, no limit on length, running entirely on my own machine.
MOVE 02 / SPEED
Two paths to the same transcript
Where a platform already publishes captions, the tool pulls them natively and returns a transcript in seconds. Where it does not, Whisper transcribes the audio locally as a fallback. Most videos take the fast path, which is where the speed gain comes from.
MOVE 03 / STRUCTURE
Transcription is the input, not the product
A transcript on its own is just more text to lose. Each one is handed to an agent that files the lesson into a fixed category, so the playbook builds itself as the transcripts come in. Sales, go-to-market, marketing, design and branding stay separated, and the whole thing can be queried by an LLM or read straight through.
THE breakdown.
[ THE BREAKDOWN ]YAPTURE LOCAL LAUNCH & LIVE WORKFLOW
PASTE A LINK. GET EVERY WORD.
BULK TRANSCRIPTION QUEUE IN PROGRESS
ALL TRANSCRIPTIONS COMPLETED
LOCAL PROCESSING & FEATURE OVERVIEW
LIVE STREAMING TRANSCRIPTION VIEW
01. CORE TRANSCRIBER
Built a local web page that takes a video link and returns a clean transcript. It pulls the audio with yt-dlp, transcribes with Whisper on Apple Silicon, copies the result to the clipboard, and saves a Markdown file tagged with the creator, title and link.
02. ALWAYS ON
Set it to launch on startup, so there is no terminal and no npm run dev. It runs in the background and is ready the moment it is needed, which is the difference between a tool used daily and one that gets forgotten.
03. SPEED
The first YouTube test took about 5 minutes, and most of that was download time rather than transcription. I switched to pulling the platform's own captions first, downloading audio only, and keeping the AI model loaded in memory. YouTube came down to 5.6 seconds, an improvement of about 50x, and Instagram reels to about 17 seconds.
04. BULK AND USABILITY
Added bulk mode: open a batch of tabs, copy every link, paste them in at once. It filters out non-video links, skips anything already transcribed, and queues the rest. A bookmarklet handles one-click transcribing from the page you are already on.
05. PLAYBOOK STRUCTURE
A transcript is only useful if it can be found again. I locked the main categories early so new lessons cannot drift into the wrong section, added an Unsorted file for anything that does not fit, and set a rule to merge duplicates rather than file them twice. The result is one organized playbook rather than a folder of loose text files.
06. AUTOMATED HANDOFF
Finished transcripts land in a folder the playbook agent reads from. One command sends it through the whole batch: it files each lesson into the right section and moves the source file to a processed folder so nothing gets handled twice. From there the playbook is queryable: ask the agent how to approach a launch, a pricing decision or a cold outreach sequence and it's built to answer from the lessons already filed.
[ VALIDATION ]
The first transcript matched its source word for word. Speed went from about 5 minutes to 5.6 seconds for YouTube and about 17 seconds for reels.
[ REFLECTION ]
Sometimes building it beats searching for it. An afternoon produced a better tool than the ones I had been fighting with. The lesson that mattered most was structure: being off by one degree at the start compounds the further you go, so I locked the categories before the playbook grew any bigger. The second was to verify against the source rather than trust a summary of it.
[ NEXT STEPS ]
Put the handoff on a schedule so transcripts reach the playbook unattended. Reels under 15 seconds. A monthly duplicate sweep. The same flow applies to anything people save and never revisit, recipes included.