The author of the Dubibubi channel sifted through hundreds of Hermes Agent skills and picked 14 he would keep, plus three honorable mentions. Below — each skill in format: name, description, link, install code.
Top 3: install first
Agent-Reach
The agent can reason, plan, and build. But as soon as you ask it to find something on the internet, it goes blind: Twitter requires a paid API, Reddit returns 403, YouTube blocks cloud IPs. Agent-Reach solves this in one install: Twitter, Reddit, YouTube, and GitHub become available without API fees.
- Why: full network access
- Cost: 0 API keys
- Feature: fallback paths around blocks
The main reason: every other skill does something the agent already does. This one gives the agent something it never had — seeing the whole internet, not just what's easily scrapable. Plus: when a platform changes rules and blocks one integration, Agent-Reach already has a fallback path. The agent won't even notice.
Link: https://github.com/Panniantong/Agent-Reach
Install code:
help me install the skill Agent-Reach from https://github.com/Panniantong/Agent-Reach
make-interfaces-feel-better
A rare find: the developer turned an article about details that make interfaces feel better into a skill. Install once, and the agent applies all of it automatically every time it builds UI.
- What's inside: line breaks, radii, animations
- Application: automatic on every UI
Instead of remembering these details every time, the agent just does them: preventing headlines from orphaning a single word; matching nested button radii to their cards; making icons respond to hover with opacity, scale, and blur; keeping stats numbers steady with tabular figures; preventing animations from getting stuck on rapid clicks.
Link: https://github.com/jakubkrehel/make-interfaces-feel-better
Install code:
help me install the skill make-interfaces-feel-better from https://github.com/jakubkrehel/make-interfaces-feel-better
oh-my-hermes
Turns a single agent into a committee that checks itself. AI easily becomes an "agreement machine": whatever you say, it agrees. This skill expands a session into a multi-agent workflow: the task is decomposed, specialized agents or external CLI workers (Codex, Gemini, Cursor) are assigned, they run in parallel or stages, and the result is verified rather than left halfway.
- What it does: decomposes task into subtasks
- Verification: result is verified
- Savings: up to −50% tokens
Inspired by oh-my-claude (36k stars). A single agent session expands into a full multi-agent workflow. Practical result: bigger tasks get done faster, less manual prompting, proper reviews and QA without your involvement.
Link: https://github.com/witt3rd/oh-my-hermes
Install code:
help me install the skill oh-my-hermes from https://github.com/witt3rd/oh-my-hermes
Efficiency: skills
SkillClaw
After every session the agent runs an evolution cycle: reviews used skills, removes overlaps, rewrites weak ones, updates its library — without your involvement. Install this first, before anything else.
- What it does: self-updates skills
- When: after every session
Skills age, duplicate, lose relevance. SkillClaw solves this forever: the agent cleans its own library. Install first so other skills immediately enter a clean registry.
Link: https://github.com/AMAP-ML/SkillClaw
Install code:
help me install the skill SkillClaw from https://github.com/AMAP-ML/SkillClaw
Matt Pocock Skills
A pack of 15 skills. The author highlights three: Grill Me — five questions before writing code; Caveat — cuts token bloat in long sessions, up to −75%; Teach Me — teaches any subject and formats takeaways in HTML.
- Grill Me: 5 questions before code
- Caveat: −75% tokens in long sessions
- Teach Me: teaches via AI → HTML
Grill Me interviews you first — five targeted questions about what you actually need to build. Annoying until you realize that's exactly why agents build the wrong thing: missing context. Caveat cuts token bloat. Teach Me is the favorite: teaches any subject and outputs a ready HTML cheatsheet.
Link: https://github.com/mattpocock/skills
Install code:
help me install the skill Matt Pocock Skills from https://github.com/mattpocock/skills
Defuddle
When an agent visits a page it reads menus, footers, cookie banners, and ads — burning tokens for nothing. Defuddle strips the page to clean markdown before the agent processes it. Web reading becomes 3–4× more efficient.
- What it gives: clean markdown instead of garbage
- Effect: web reading 3–4× more efficient
Most pages are 90% noise. Defuddle works in read-only mode and feeds the agent only content. Token savings and clean context — no menus, footers, or popups.
Link: https://github.com/kepano/defuddle
Install code:
help me install the skill Defuddle from https://github.com/kepano/defuddle
Humanizer
Already built into Hermes but underused. Rewrites AI text into a natural human voice, pulling edits from Wikipedia. The author calls it the last line of defense before publishing.
- What it does: rewrites text in human voice
- Edit source: Wikipedia
AI style is instantly recognizable: bureaucratese, passive voice, uniform sentence length. Humanizer breaks these patterns and pulls facts from Wikipedia, making the text alive.
Link: https://github.com/blader/humanizer
Install code:
help me install the skill Humanizer from https://github.com/blader/humanizer
Capabilities: skills
youtube-full
Replaces the default YouTube skill that breaks on VPS and cloud machines. Works without a Google API key: transcripts, channels, playlists, search.
- Transcripts: yes
- Channels / playlists: yes
- Search: without Google API key
If the agent can't read YouTube, half the internet doesn't exist for it. Works via an API processing 15M transcripts/month. The author is serious: YouTube transcripts are a great way to inject niche expertise into agents.
Link: https://github.com/ZeroPointRepo/youtube-skills
Install code:
help me install the skill youtube-full from https://github.com/ZeroPointRepo/youtube-skills
Composio
Connects the agent to Gmail, Sheets, Slack, Notion, HubSpot, Salesforce — over a thousand SaaS tools. No custom OAuth or key juggling.
- Integrations: 1000+ SaaS
- Coverage: Gmail, Slack, Notion, CRM, etc.
Previously every SaaS needed its own token, OAuth, and wrapper. Composio gives a unified layer: agent says "send email" and Composio handles the Gmail API. For production agents this is a must-have.
Link: https://github.com/ComposioHQ/skills
Install code:
help me install the skill Composio from https://github.com/ComposioHQ/skills
addyosmani/agent-skills
The most starred pack on the list — 65k stars. 24 production skills around eight slash commands: spec → plan → build → test → review → ship. The key feature is doubt-driven development: the agent pauses and argues with itself before big decisions.
- Skills: 24
- Commands: 8 slash commands
- Philosophy: doubt-driven development
Normal agent: "done, here's the result." This one: "I don't like this approach, here's an alternative, let's discuss." That's how seniors work — they doubt before committing. The pack makes the agent behave like a senior.
Link: https://github.com/addyosmani/agent-skills
Install code:
help me install the skill addyosmani/agent-skills from https://github.com/addyosmani/agent-skills
Resemble AI Detect
Detects deepfakes in any format: audio, images, video, text. Identifies the tool used to generate content and flags it before the agent processes it.
- Detects: deepfakes in any format
- Identifies: generation tool
Generative content is flooding the internet. Agents consuming external content need an input filter. Resemble AI Detect puts that filter in place.
Link: https://github.com/resemble-ai/detect-skill
Install code:
help me install the skill Resemble AI Detect from https://github.com/resemble-ai/detect-skill
Infrastructure: skills
Mission Control / Minions
A dashboard for a fleet of agents: task dispatching, health monitoring, real-time token spend. The difference between a homemade setup and a well-oiled multi-agent machine.
- Dashboard: entire fleet
- Monitoring: health + tokens
When you have more than three agents, chat management turns into chaos. Minions gives a single control panel: who's doing what, token burn, where things failed. Without this, scaling a fleet is impossible.
Link: https://github.com/agent37-platform/minions
Install code:
help me install the skill Mission Control / Minions from https://github.com/agent37-platform/minions
OpenMontage
The world's first open-source agentic video production system: 12 pipelines, 52 tools, 500+ skills. Drop a video link and the agent analyzes pacing, structure, scenes, and tone, then produces 2–3 concepts with a cost estimate. Example: a 60-second Pixar-style animation costs around $1.33.
- Pipelines: 12
- Tools: 52
- Skills: 500+
Video production is the most expensive content part. OpenMontage makes it agentic: from reference to finished clip with budget estimate. For marketing, education, demos — this is a paradigm shift.
Link: https://github.com/calesthio/OpenMontage
Install code:
help me install the skill OpenMontage from https://github.com/calesthio/OpenMontage
Anthropic Cybersecurity Skills
700+ skills mapped onto the MITRE ATT&CK matrix: threat modeling, secure code review, incident response playbooks. A temporary replacement for an expensive security engineer for small teams.
- Skills: 700+
- Basis: MITRE ATT&CK
Security engineers are expensive and busy. This pack gives the agent red and blue team knowledge: vuln hunting, attack modeling, code review checklists. For startups without dedicated security — a lifesaver.
Link: https://github.com/mukul975/Anthropic-Cybersecurity-Skills
Install code:
help me install the skill Anthropic Cybersecurity Skills from https://github.com/mukul975/Anthropic-Cybersecurity-Skills
Honorable mentions
Browser Harness
Connects the agent to your real Chrome. It clicks, scrolls, fills forms. If it hits a page it doesn't know how to handle, it writes the missing helper and continues.
- Author: browser-use team
- Mode: real Chrome
Headless browsers break on complex SPAs, captchas, dynamic content. Browser Harness gives the agent your real browser with your cookies, sessions, extensions. The agent sees exactly what you see.
Link: https://github.com/browser-use/browser-harness
Install code:
help me install the skill Browser Harness from https://github.com/browser-use/browser-harness
codebase-memory-mcp
Indexes your entire codebase into a persistent knowledge graph. 28 million lines of Linux kernel indexed in 3 minutes. After indexing, the agent spends 120× fewer tokens studying the codebase.
- Speed: 28M lines in 3 min
- Token savings: 120× less
Without indexing, the agent reads files one by one — tokens burn, context is lost. codebase-memory-mcp builds the graph once, and then the agent "knows" the whole codebase. For large projects — the only way to give the agent full understanding.
Link: https://github.com/DeusData/codebase-memory-mcp
Install code:
help me install the skill codebase-memory-mcp from https://github.com/DeusData/codebase-memory-mcp
Loopy
Adds a feedback loop: measure, keep if it worked, repeat until goal. Instead of "make the site faster" you get targeted improvements verified by measurement.
- Purpose: feedback loop
- Principle: measure → keep → repeat
Agents often do things "head-on" and consider the task done. Loopy forces them to verify results with metrics and iterate until improvement. This turns a one-shot run into an optimization process.
Link: https://github.com/Forward-Future/loopy
Install code:
help me install the skill Loopy from https://github.com/Forward-Future/loopy
Source
Original video: 14 Hermes Agent Skills You NEED To Install Right Now by Dubibubi. Translated and adapted for V&L notes.
