Best Hermes Skills in 2026: The 17 I Would Install First

Who wrote this, and how

I'm Muhammad Shadab Shams, an AI Automation Consultant. I ship products by directing agents rather than typing every line myself, and I currently run a dozen production agents across two live properties. Skills are the layer that made that possible, so this is not a curiosity list.

How the numbers were checked. Every star count, repository owner and description in this post was pulled live from the GitHub API on August 27, 2026. That matters because skill repos are moving fast enough that a figure from three weeks ago is already wrong. Two examples: mattpocock/skills is at 238,275 stars, not the ~224k circulating in most write-ups, and blader/humanizer has 38,179 stars, a number almost nobody quotes at all.

What I did not verify. Three of the seventeen (Composio skills, Browser Harness, Loopy / Loop Library) did not resolve to a confirmable canonical repository during checking. They stay on the list because the jobs they do are real and I use those categories, but they carry an explicit unverified flag instead of a made-up number. A missing figure beats an invented one.

Last updated: August 27, 2026.

At a Glance

What an Agent Skill Actually Is

An agent skill is a written procedure your agent loads on demand when a task matches it. Think of it as a recipe card rather than a plugin: it does not add a capability to the model, it tells the model how you want a job done, every time, without you re-explaining.

The difference shows up in three places:

1. Without skills, you re-explain the same workflow every session, output quality tracks whatever the model feels like that day, and multi-step work drifts because nothing anchors the process.
2. With skills, the steps are written down once, the agent pulls them when the task matches, and the same procedure travels across projects and machines.
3. Hermes ships with a good set already. The seventeen below are the external packs people bolt on when they want a setup that behaves like infrastructure rather than a chat window.

The bottom line: A skill converts a workflow you keep in your head into one the agent can execute without you.

The Whole Roster, Ranked by Stars

Ordered by verified GitHub stars, highest first.

⚠️ On star counts: Stars measure attention, not quality. A 65-star skill from Resemble AI is a production tool from a real company. A 200,000-star repo can still be wrong for your workflow. Read the star column as "how many people found this," never as "how good this is."

The honest version of the advice. If you install all seventeen before you have run a single job, you will spend your first week debugging your setup rather than shipping.

• Agent-Reach: So the agent can see the public internet at all.
• addyosmani/agent-skills: So engineering work has gates instead of vibes.
• Humanizer: So nothing you publish reads like a press release.
• Defuddle: So web pages arrive as clean text instead of nav bars and cookie banners.
• Minions: So you can see what five parallel agents are actually doing.

Those five cover research in, work through, and output out. Everything else is a specialist you add the first time you genuinely need it.

Skip on day one: The cybersecurity library (huge, and pointless until you have an authorized target), OpenMontage (only if you actually produce video), Resemble Detect (only if you handle user-generated media), and SkillClaw (it improves skills from real usage, so it needs usage to exist first).

1. Agent-Reach — 75,763 stars

Panniantong/Agent-Reach

Agent-Reach is a CLI that gives your agent read access to the public internet across X, Reddit, YouTube, GitHub, Bilibili and XiaoHongShu, with no paid API keys. That last part is the reason it has 75,000 stars.

The problem it fixes is specific and familiar. Your agent writes code fine, then you ask it to check reactions on X or summarize a YouTube tutorial and it falls over: blocked scrapers, paywalled APIs, login walls, unparseable HTML.

After installing, the instructions become plain language: search X for what people are saying about a tool, pull a YouTube transcript and summarize it, or check whether anyone on Reddit has hit the same bug.

• Use it for: Content research, competitor checks, social listening, debugging against community reports.

2. Defuddle — 9,173 stars

kepano/defuddle

Defuddle takes a messy web page and returns the main content as clean HTML or Markdown. URL in, readable article out. It comes out of the Obsidian Web Clipper world, which is why it is unusually good at deciding what on a page is actually the content.

Agents choke on raw HTML. Navigation, ad slots, cookie banners and six levels of nested divs all burn tokens and none of it is the article. Defuddle strips it before the agent ever sees it.

• Use it for: Research pipelines, clipping into notes, feeding clean context into any agent. Pair it with Agent-Reach and your research input problem is basically solved.

3. youtube-skills — 572 stars

ZeroPointRepo/youtube-skills

This package pulls YouTube transcripts, searches videos, browses channels and extracts playlists without fighting yt-dlp blocks on cloud IPs. "Summarize this YouTube video" is one of the single most common agent failures, and this is purpose-built for it.

1hermes skills install skills-sh/ZeroPointRepo/youtube-skills/skills/youtube-full

• Use it for: Learning from tutorials, competitor video breakdowns, turning a conference talk into notes. However, the underlying transcript API runs on a credit model, so confirm the current free tier before you build anything that depends on it at volume.

4. Matt Pocock skills — 238,275 stars

mattpocock/skills

The most-starred skill pack on this list by a wide margin, and the philosophy is the reason: small, composable, editable skills rather than one large framework.

The pitch is that big process frameworks take control away from you. When your agent has a giant opaque process and something goes wrong, you cannot debug it. When it has no process at all, work drifts. Matt's set sits deliberately in the middle: small enough that you can open one and hack it.

1npx skills@latest add mattpocock/skills

Then run the setup skill once per repository.

• Use it for: TypeScript and app engineering, ticket triage, planning, day-to-day shipping.

5. addyosmani/agent-skills — 90,115 stars

addyosmani/agent-skills

Production engineering skills built around a lifecycle: define, plan, build, verify, review, ship. What it fixes is vibe coding with no gates. These skills force habits senior engineers already have: write the spec, cut the work small, test it, review it.

1npx skills add addyosmani/agent-skills

• Use it for: Real products rather than throwaway scripts, especially frontend work and anything spanning more than a day.

6. make-interfaces-feel-better — 3,024 stars

jakubkrehel/make-interfaces-feel-better

A skill dedicated entirely to UI polish: animation, type, icons, hover states, hit areas, shadows, optical alignment. The details that separate an interface that works from one that feels intentional.

AI-built UIs have a signature. They function, and they feel slightly off. Spacing is almost right. Hover states are dead. Buttons are a few pixels too small on mobile. Nothing is broken and everything feels cheap. Run it as a second pass after your agent ships a page, and ask it to review for feel rather than features.

1npx skills add jakubkrehel/make-interfaces-feel-better

• Use it for: Landing pages, dashboards, internal tools, any frontend an agent just built.

7. codebase-memory-mcp — 40,809 stars

DeusData/codebase-memory-mcp

An MCP server that indexes your codebase into a persistent knowledge graph, covering 158 languages with sub-millisecond queries and a claimed 99% reduction in tokens. It ships as a single static binary with no dependencies.

The problem is context. Every time an agent needs to understand your repo, it reads files, and reading files is the most expensive thing it does. A knowledge graph answers structural questions (what calls this, where is this defined, what breaks if I change it) without loading the files at all.

• Use it for: Any repository large enough that your agent keeps rediscovering the same architecture. The token savings compound daily.

8. oh-my-hermes — 291 stars

witt3rd/oh-my-hermes

Multi-agent orchestration built natively for Hermes primitives, inspired by oh-my-claudecode. Small repo, disproportionate value if you work on things bigger than one session. What it fixes is one agent freestyling a large task: skipping requirements, producing a weak plan, then declaring victory with no proof.

Suggested flow: research → interview → ralplan → ralph

• Use it for: Unfamiliar domains, multi-day features, anything where plan quality matters more than speed.

9. Minions — 624 stars

agent37-platform/minions

A Kanban-style mission control for Hermes: create tasks, watch them run, review output, and stop long jobs from vanishing into terminal scrollback.

Hermes is strong. Managing five parallel sessions across five raw terminals is not. You lose track of what finished, what is stuck, and what is silently waiting on you. Minions provides an in-progress / review / done board, live streaming of tool calls, human sign-off before anything is marked done, per-task model settings, scheduled jobs, and a local-first SQLite database.

1npx minionsai

Then open http://localhost:6969.

• Use it for: The moment you run Hermes as a worker rather than a chat toy. This is the skill that makes parallel agents feel manageable, and it is why it made my day-one five despite the modest star count.

10. Loopy / Loop Library — unverified

Repeatable agent loops with checks and stop rules. The category matters more than the specific package: a loop without a stop rule is how you wake up to a burned budget and 400 commits of nonsense.

I could not confirm a canonical repository for this one during fact-checking, so treat the entry as a category recommendation rather than a specific install. If you already run loops, the discipline to copy is simple: define the success check, define the stop condition, and cap the iterations before you start.

11. Composio skills — unverified count

Distributable skills that teach agents to work with Composio: Tool Router, auth, sessions, triggers and production patterns.

Connecting an agent to Gmail, Slack, GitHub or a CRM is messy, and most of the mess is authentication and session handling rather than the API calls. Composio is the integration layer, and these skills teach the agent the safe patterns around it.

1npx skills add composiohq/skills

• Use it for: Any setup that needs to take real actions in real apps rather than edit local files. Confirm against Composio's documentation before relying on it.

12. Browser Harness — unverified

The agent drives your real browser, in your real session, and learns reusable helpers as it goes. The learning part is the interesting bit: rather than re-deriving how to navigate a site every run, it accumulates helpers.

Verify the source before installing anything that gets access to a logged-in browser session, because that is one of the highest-trust permissions you can grant an agent.

13. OpenMontage — 51,586 stars

calesthio/OpenMontage

An open-source agentic video production system: 12 production pipelines, 100+ tools, and over 700 agent skill and production-knowledge files. It is trying to turn a coding agent into something closer to a small video crew.

Video is a pile of separate tools: script, b-roll, edit, captions, export. Most agents can discuss video competently and produce nothing. OpenMontage aims at the actual production workflow.

• Use it for: Short-form content, explainers, product demos. However, this is the heaviest install on the list by a wide margin, so do not add it until you are genuinely producing video with an agent in the loop.

14. Anthropic Cybersecurity Skills — 31,306 stars

mukul975/Anthropic-Cybersecurity-Skills

817 structured cybersecurity skills across 29 security domains, mapped to six frameworks: MITRE ATT&CK, NIST CSF 2.0, MITRE ATLAS, D3FEND, NIST AI RMF and MITRE F3. Apache 2.0.

⚠️ Read the name carefully: This is a community project, not an official Anthropic product. The repository name is doing a lot of work it has not earned, and plenty of people install it assuming a vendor guarantee that does not exist.

It also contains offensive and dual-use techniques intended for authorized testing, defense, research and education. Point it only at systems you own or have written permission to test.

• Use it for: Security reviews, incident response drafts, threat modeling, authorized red-team prep, or learning security workflows with an agent as tutor.

15. Resemble AI Detect — 65 stars

resemble-ai/detect-skill

A media safety skill that detects AI-generated or manipulated audio, images and video using Resemble's detection stack. The smallest repo on the list, and one of the few built by a company whose actual business is the thing the skill does.

• Use it for: Content moderation pipelines, brand safety review, "is this clip real" checks. Treat detection scores as signals rather than absolute proof.

16. Humanizer — 38,179 stars

blader/humanizer

A skill that rewrites AI-sounding text so it reads like a person wrote it, built on Wikipedia's documented "signs of AI writing" patterns. At 38,179 stars it is one of the most popular skills in existence.

The problem is instantly recognizable. Your agent's draft is factually useful and reads like a press release: "pivotal," "landscape," "it's not X, it's Y," three-item lists in perfect parallel, and hype with no specifics behind it.

How it works in practice:

1. Rewrite once without preserving the AI structure.
2. Scan what remains for AI tells.
3. Rewrite again.
4. Keep every fact, invent no names or numbers.

The tip that actually matters: Paste two or three paragraphs of your own writing and ask it to match your voice. Generic de-slopping produces text that sounds like nobody. Voice-matched de-slopping produces text that sounds like you.

17. SkillClaw — 2,517 stars

AMAP-ML/SkillClaw

A system where skills evolve from real use. You work normally. In the background it digests sessions, improves the skills you actually used, removes ones that never fire, and compounds that experience across agents and devices.

• Use it for: Long-running setups where the same categories of work recur. Install it after you have real usage history, since a system that learns from sessions needs sessions to learn from.

Here is the thing missing from every list of this kind: skills consume context.

Not just when they run, but when they are available. Descriptions, triggers and loaded procedures all occupy the same window your actual task needs. Install seventeen packs, several of which contain hundreds of individual skills, and you have spent a meaningful share of your budget before typing a word.

• The cybersecurity library alone is 817 skills.
• OpenMontage carries over 700 skill and production-knowledge files.

Three rules I follow:

1. Install per project, not globally: A video project does not need security playbooks. A security review does not need UI polish.
2. Uninstall what never fires: If a pack has not triggered in a month, it is paying rent with your context.
3. Prefer small composable packs over large frameworks: This is exactly why Matt Pocock's set works so effectively.

Every skill here is free. Context is the actual scarce resource, and it is the one nobody budgets.

What is Missing from This List

Being honest about the gaps is more useful than pretending seventeen is a complete map:

• The two largest skill repositories on GitHub are omitted: obra/superpowers sits at 278,274 stars and anthropics/skills, the official Anthropic library, at 171,918. Their absence is defensible because this list focuses on external add-on packs rather than foundational base frameworks.
• Other notable packs outside the seventeen: Leonxlnx/taste-skill (81,181 stars) to eliminate generic AI design slop; mvanhorn/last30days-skill (59,370 stars) for cross-platform recency research; coreyhaines31/marketingskills (45,797 stars) for CRO and copywriting; and kepano/obsidian-skills (47,352 stars) for Obsidian notes integration.

Honest Limitations

1. Three entries are unverified: Composio skills, Browser Harness and Loopy / Loop Library did not resolve to confirmable canonical repositories. Treat them as category patterns.
2. Star counts decay fast: Figures are a snapshot dated August 27, 2026.
3. Skills do not fix a weak model: A skill is a procedure, not a capability. It makes a capable agent consistent; it does not make a limited one smart.
4. More skills is not better: Past a certain point you are trading context for options you never exercise.

How to Choose in 60 Seconds

1. What does your agent fail at most often? Install the skill for that, alone, and use it for a week. Failure-driven installation beats aspirational installation every time.
2. Is the work repetitive or exploratory? Repetitive means loops, mission control, and memory (Minions, codebase-memory-mcp). Exploratory means research and planning (Agent-Reach, oh-my-hermes, Defuddle).
3. Does anything you produce get published? If yes, Humanizer is not optional. It is the difference between human voice and machine slop.

Glossary of Agent Architecture Terms

• Agent skill: A written procedure an agent loads on demand when a task matches it. Adds consistency, not capability.
• Hermes Agent: The Nous Research agent runtime these skills target, though most work across Claude Code, Codex, Cursor and other runtimes too.
• MCP (Model Context Protocol): The standard connectivity layer for exposing tools and data to agents.
• Knowledge graph: A structured map of entities and relationships used to understand codebases without consuming tokens by reading full files.
• Context window: The working memory an agent has for a task. Loaded skills consume it whether they run or not.
• Skill pack: A repository containing multiple related skills installed together.
• Harness: The layer around the model handling tools, memory, permissions and recovery.
• Agent loop: A repeating cycle where an agent acts, checks the result, and continues. Requires strict stop rules.
• AI tells: Recognizable patterns marking text as machine-written (hype adjectives, uniform parallelism).
• Dual-use: Techniques usable for defense or attack requiring strict authorization.

Frequently Asked Questions

Written by Muhammad Shadab Shams | AI Automation Consultant | aifloxium.online | X @ShadabLoveAi