YouTube inbox ingest — 2026-08-13

Seven Inbox videos, all with native English transcripts. Themes cluster around agent coordination (Hermes Kanban + plugins), safer non-technical builders (Cloudflare OS gatekeepers), PR review automation (Vercel Eve vs Hermes/OpenClaw), local model routing (Switchyard + Lightning), and grounded specialist advisors (Notion KB from books; synthetic personas as forecasts not people).

Skipped already-ingested -IGB6Avxwgo (Switchyard Level1Techs). Playlist not modified.

1. The Ultimate Guide to Hermes Desktop Plugins

Tonbi's AI Garage · SLUtt-po_OE · wiki: youtube-hermes-desktop-plugins.md

• Plugin SDK is host-owned UI (~25 placements), not a native always-on service.
• Form follows job: compact / anchored / popover / expansive / declarative.
• Demos: Hyperliquid markets, Spotify controller (needs desktop app), SNES via emulator.js.

2. Testing Cloudflare OS

Tonbi's AI Garage · 1mXyYAfCt6M · wiki: youtube-cloudflare-os.md

• OSS agent workspace + gadget builder, not a general OS. Repo cloudflare/cloudflareos.
• Gatekeepers: Capn Web API, simulate then human-approve side effects (Cloudflare Workers).
• Best for non-technical teams already on Cloudflare; Durable Objects hold gadget state.

3. Hermes Agent Kanban multi-agent workflow

Tonbi's AI Garage · EKVRqcpTT6s · wiki: youtube-hermes-kanban-multi-agent.md

• SQLite Kanban is the bus + audit log; no agent-to-agent chat.
• Scout → rubric (65/100) → parallel researchers → one Telegram human gate → build or video.
• ~18 parallel workers, ~97 tasks in the live run; self-heal when slides landed in scratch dirs.

4. AI code review bot in 30 minutes

How I AI (Claire) · cmATJGbA8bI · wiki: youtube-ai-code-review-bot.md

• CI-green → risk score → auto-approve low / Slack-escalate medium-high. Intercom: 5x faster, lower revert rate.
• Vercel Eve + Codex + Chrome browser-use for GitHub/Slack app setup. Thresholds <24 / 25–64 / 65+.
• Eve framed as easier than Hermes Agent or OpenClaw for Slack+GitHub internal bots. Bot cannot satisfy required human reviewers (gray check).

5. Persona Engineering (Ishan Anand)

AI Engineer · YnNF55QV0zs · wiki: youtube-persona-engineering.md

• Synthetic personas are weather forecasts: bounded, improving, trustworthy only vs reality.
• Failures: missing-context confounders (improv), order bias, say-vs-do. More samples ≠ statistical significance.
• ~83% alignment vs ~80% human self-consistency noise floor. Complement human research; don't replace it.

6. Build a Local Router with Nemotron Lightning

Prompt Engineering · ofpqWw0OFKo · wiki: youtube-local-router-nemotron-lightning.md

• Route for cost, speed, specialization, privacy. Switchyard strategies: random, LLM classifier, stage, escalation.
• Don't route per turn (breaks cache). Lightning on DGX Spark ~71 tok/s; 20s vs Kimi K3 50s on a 1400-token task.
• PII-stripping digest path: local Lightning then capable API. Companion to Level1Techs Switchyard page.

7. Turn Books Into AI Business Advisors

Systems Made Better · OAgU6sOmih0 · wiki: youtube-books-ai-business-advisors-notion.md

• Specialist = instructions + skills + curated KB. Empty KB must refuse; that's the proof the gate works.
• Notion Personal Agent (not credit-hungry custom agents) + chapter-paste ingest (PDFs >~15 pages fail).
• Godin demo: position against AI productivity, not within it. Compounding loop is cheaper as a monthly skill than a scheduled custom agent.

-IGB6Avxwgo — You Don't need to use Cloud AI! Switchyard and Nemotron 3.5 Lightning (already in processed.json / youtube-switchyard-nemotron-lightning.md).