NLP · YouTube API · Next.js · 2026 · Solo — trend research, NLP implementation, interaction design, TDD, verification, deployment
Clapmark
Turn any YouTube transcript into timestamped chapters and apply them directly to your video — no LLM, no upload, pure algorithm.
The problem
YouTube chapter timestamps improve watch time and search ranking, but writing them manually takes 15–30 minutes per video. No lightweight tool reads your actual transcript content and applies chapters via the YouTube API in one step. Most creators skip chapters entirely or copy-paste timestamps from memory.
Architecture
Key decisions
TextTiling over LLM segmentation
TextTiling (Hearst 1997) is deterministic, cheap, and produces topically coherent segments without any external API call or model dependency. An LLM would require a paid API key and introduce latency and non-determinism for a task that is well-solved algorithmically.
TF-IDF chapter naming with stop-word filtering
TF-IDF scores segment-specific vocabulary against the full transcript, surfacing distinguishing terms rather than common words. Stop-word filtering prevents generic function words from dominating titles. No prompt engineering required.
Description splice with regex boundary detection
Before writing, the product scans the existing description for a line matching the YouTube chapter timestamp pattern and replaces only that block. If no chapter block exists, it appends after a blank line. This preserves all other description content, hashtags, and links.
OAuth write guard and diff preview
The apply action requires a server-side OAuth access token. The UI shows a live preview of the updated description before committing the write. The product will not write to videos it cannot read first.