Turning Hermes Agent Skills to Dashboards (Why This is BETTER)

Skill → Dashboard: 40 minutes of skill work turned into 4 days of dashboard work — and it's still worth it.

## The Pipeline The job is pulling YouTube videos for boxmining.ai: fetch videos, filter shorts, grab transcripts from YouTube transcripts.com, pull comments, and store everything in a database. Hermes ran this as a single long skill through Python scripts the AI generated on the fly.

## Why Hermes Broke Hermes would "sometimes lie to you that it ran something." Real example: it told the creator there was an empty transcript on a video that had none, and once logged 20 views when the real number was 4,000. After 1–2 weeks the fetcher started silently missing new videos, and the only fix was to manually nudge it with prompts like "have a transcript. Make sure you get the transcript next time." Designing the skill was the easy part — sorting out errors is now "the biggest challenge to AI."

## Why a Dashboard Beats It The dashboard is written in TypeScript after AI-assisted generation, not in an LLM loop. It is "probabilistic vs deterministic" — the script runs the same way every time, shows you the step it's on, surfaces YouTube API quota usage, and lists exactly how many videos, comments, and transcripts were fetched. No tokens spent per run. Cron jobs can trigger it.

## The Real Cost The creator originally estimated a 1-hour conversion. Authentication (Supabase), role-level security, login, UX flows, and double-click protection stretched the build to 4 days across Claude Code 4.7 and 4.8. Edge cases that "an idiot pressing the button twice" exposes — stale "still running" flags, YouTube API keys silently failing, RAM blowups from concurrent runs — all had to be hand-coded because the app is "a stupid computer" without an LLM in the loop to self-correct.

## The Rule Use Hermes for one-off or exploratory work. Build a dashboard for anything daily, repeated, and quota-bound — once it works, it is "very reliable and dependable."