Video summary

I Tested 100+ Hermes Agent Automations. These Are The Best

Main summary

Key takeaways

Technology

Summary of technological concepts & “best Hermes skills” (top 14)

The speaker (Dibby) tests 100+ Hermes “skills” (small Markdown-based task modules distributed via GitHub repos) and recommends 14 skills that are genuinely useful instead of “trash” ones. A recurring constraint is token/memory efficiency: a “good Hermes skill stays under 15 kilobytes” to avoid bloating agent memory and wasting tokens.


Core Hermes skill system (explained)

  • Skill format: a Markdown file that guides an agent step-by-step.
  • Install method: copy a GitHub repo and ask the Hermes agent to download/install.
  • Execution: installed skills live in the agent’s memory and can be triggered during prompts or manually by instructing the agent to use a specific skill.

Top 14 recommended skills (with key features)

#14 Skill Claw (passive “self-improvement”)

  • After each session, it runs an evolution loop:
    • reviews which skills were used,
    • deduplicates overlaps,
    • rewrites weak skills,
    • and updates the skill library automatically.
  • Goal: agent gets better over time without manual intervention.
  • Emphasized as passive/background automation.

#13 Matt Pocock’s “skills pack” (15 skills total; 3 standouts)

  1. Grill Me

    • Before coding, the agent interviews the user with 5 targeted questions
    • Intended to reduce misinterpretation/mis-building by collecting requirements up front.
  2. Caveat

    • Reduces token bloat in long sessions
    • Claims up to 75% lower token usage
  3. Teach Me

    • Makes the agent teach topics using an HTML-structured output
    • Example claim: teaching solving a Rubik’s Cube

#12 Defuddle (webpage token-efficiency / reader mode)

  • Web research issue: reading full webpages (nav, footer, cookie banners, side ads) burns tokens.
  • Defuddle converts webpages into clean “reader mode” markdown before the agent reads them.
  • Result: agent reads 3–4× more efficiently.
  • Recommended as critical for research, competitive analysis, market research, documentation lookups.

#11 Humanizer skill (anti–AI writing style)

  • Used for copywriting outputs (emails, proposals, social posts) that often sound like “AI slop.”
  • Claims to use a skill generated from “signs of AI” patterns and auto-updates when that reference page changes.
  • Purpose: rewrite agent output into natural human voice as a “last line of defense.”

#10 YouTube Full (cloud/VPS compatibility for transcripts)

  • Problem: Hermes’ default YouTube skill breaks on VPS/cloud because YouTube blocks cloud IPs.
  • Replacement supports:
    • transcript extraction
    • channel browsing
    • playlist parsing
    • video search
  • Claims: no Google API key required; powered by processing ~15M transcripts/month.
  • Use case: agents can learn domain knowledge from transcripts (example: improving YouTube scriptwriting).

#9 Composio (integrations without OAuth/API pain)

  • Connects Hermes agents to common tools (e.g., Gmail, Sheets, Slack, Notion, HubSpot, Salesforce) via a single skill.
  • Claims: 1000+ SaaS tools, no hand-rolled OAuth, no API key juggling.
  • Motivation: integrations were the bottleneck when adding new tools; Composio removes that bottleneck.

#8 Addy Osmani’s dev lifecycle skills pack (24 skills mapped to “/commands”)

  • Provides production-grade skills corresponding to dev workflow stages:
    • /spec, /build, /test, /review-before-merge, /ship
  • Standout: doubt-driven development
    • Mid-task, the agent pauses to extract assumptions, challenge them, reconcile gaps, then continues.
    • For irreversible/production-impact steps, it questions itself first.

#7 Resemble AI Detect (deepfake/AI-content detection)

  • Adds detection for user-submitted content in ingestion pipelines:
    • AI-generated audio/images/video/text
  • Claims:
    • identifies which tool generated content (e.g., 11 Labs, ChatGPT, Claude)
    • applies watermarking/AI labeling so the agent treats it appropriately.

#6 Mission Control (multi-agent infrastructure + visibility)

  • For running multiple Hermes agents (research, coding, client comms).
  • Provides a fleet dashboard:
    • task dispatch
    • agent health monitoring
    • real-time cost tracking
    • live status across agents
  • Positioned as the difference between “amateur agent setup” and “multi-agent machine.”

#5 Open Montage (agentic video production system)

  • Claimed as the world’s first open-source agentic video production system.
  • Claims:
    • 12 pipelines, 52 tools, 500+ agent skills in one install
  • Workflow:
    • user provides a YouTube video/reel reference
    • agent analyzes transcript + pacing/scene/keyframes/tone
    • outputs 2–3 concepts, tool path plan, cost estimate, and preview before generating frames
  • Example cost/scene planning: a 60-second Pixar-style animated short requires multiple generated assets and “will probably cost” ~$1.33.
  • Emphasized for content products and generating video assets without manual editing.

#4 Anthropic cybersecurity skills (security knowledge base mapped to MITRE ATT&CK)

  • Built by security researcher McColl 975.
  • Claims: 700+ structured skills mapped to MITRE ATT&CK.
  • Queryable content includes:
    • threat modeling
    • vulnerability assessment
    • secure code review
    • incident response playbooks
  • Use case: replaces the need for a dedicated security engineer at small-team stage.

#3 O my Hermes (multi-agent “committee” + verification)

  • Turns one agent into a coordinated multi-agent committee that:
    • decomposes tasks into subtasks,
    • assigns specialist agents or external tools (e.g., Codex, Gemini, Cursor),
    • runs in parallel/staged pipelines,
    • and verifies results instead of accepting the first answer.
  • Claimed benefits:
    • faster large tasks
    • better review/QA via specialists
    • session logs for visibility
    • token savings up to ~50%
    • persistent execution until verification passes.

#2 “Details that make interfaces feel better” (UI polish automation)

  • Converts an interface design article into an automated skill.
  • When the agent builds a UI, it applies:
    • text wrapping (avoid orphaned headline words)
    • consistent border radius for card/button combos
    • icon animations (opacity/scale/blur)
    • tabular numbers to prevent stat jitter
    • interruptible animations that don’t freeze on rapid clicks
  • Goal: fix “UI feels off” issues automatically.

#1 Agent Reach (internet access / resilient platform access)

  • Main problem: agents can reason/build, but often become “blind” when asked to browse the internet due to platform restrictions (API limits, 403s, cloud blocking, etc.).
  • Agent Reach claims to provide access to:
    • Twitter, Reddit, YouTube, GitHub and more
    • zero API fees (as claimed)
  • Key differentiator: backup paths when platforms change/block integrations, so the agent “doesn’t notice” failures.
  • Positioned as enabling the whole internet, not just easily scrapable parts.

Honorable mentions (brief)

  • Browser Harness: connects to a real Chrome browser so the agent can click/scroll/fill forms; writes helper code when it encounters unfamiliar pages (“self-healing”).
  • Code Base Memory MCP: indexes code into a persistent knowledge graph; claims huge scaling tests (e.g., Linux kernel scan) and token reductions; multi-language support.
  • Loop library: provides feedback-cycle loops for performance optimization (measure → improve → keep if better → repeat), rather than one-shot “do it once” instructions.

Main speakers/sources (as stated)

  • Speaker: Dibby (builder of apps using Hermes agent; claims $47,000 generated in <50 days with these tools)
  • Referenced creators/companies:
    • Matt Pocock (newsletter; creator of a skills pack)
    • Addy Osmani (built the dev lifecycle skills pack)
    • Resemble AI Detect (deepfake detection tool/pack)
    • McColl 975 (Anthropic cybersecurity skills)
    • Anthropic cybersecurity skills pack
  • Inspiration mentioned: “O my Claude” skill (for committee/verification design)

Original video