14 Hermes Agent Skills Worth Installing — Tested Against Hundreds

📄 Skills, in one minute 0:22

A Hermes skill is just a markdown file that teaches the agent how to do a task, step by step. Installing one is deliberately low-ceremony: copy the GitHub repo URL and ask your Hermes agent to install it. From then on the skill sits in the agent's memory and fires either automatically — when your prompt matches what it's for — or on demand, when you name it explicitly.

The framing that opens the video is the one worth carrying through the whole list: a good skill stays under roughly 15 kilobytes. Past that it starts bloating agent memory and burning tokens on every call, because skill descriptions live in the prompt. Size is a proxy for whether the author actually did the editing work.

14 SkillClaw — the library that edits itself 0:48

Roughly 2,000 GitHub stars, and the recommendation is to install it before anything else — because it changes what happens to every other skill you install afterwards.

After each session, SkillClaw prompts the agent to run an evolution loop over its own skill library: review which skills were used, deduplicate ones that overlap, rewrite the weak ones, and update the library — without you intervening. It is maintenance that normally never gets done, because nobody schedules time to prune their agent's markdown files.

The argument for putting it first is compounding. Week one, the setup is marginally sharper. By week four you have a library that has been pruned and rewritten several times over, purely as a side effect of normal use. The appeal, as framed in the video, is that it runs passively — no new habit to maintain.

13 Matt Pocock's skills pack 1:40

Fifteen skills from Matt Pocock, who runs a newsletter with over 60,000 readers. Three are singled out.

Grill Me interviews you before the agent writes any code — five targeted questions about what you actually want built. It sounds like friction until you connect it to why agents mis-build things: they were never given enough context, and they don't ask. Grill Me makes the asking mandatory.

Caveat strips token bloat out of long sessions, with a claimed reduction in token usage of up to 75%. On real multi-hour workflows that is a direct cost line, not a micro-optimisation.

Teach Me — the presenter's favourite — has the agent teach you a subject and then structure the practical takeaways as readable HTML. Pocock has reportedly been using it to learn to solve a Rubik's Cube. The skill isn't teaching content; it's a framework that gives the model a teaching structure to follow.

12 Defuddle — stop paying for nav bars 2:58

Every time an agent opens a web page it ingests the whole thing: navigation, footer, cookie banner, sidebar ads. All of it costs tokens; almost none of it carries signal.

Defuddle strips pages down to clean reader-mode markdown before the agent processes them. The stated result is that the agent reads the web three to four times more efficiently — the same pages, minus the furniture.

The conclusion is stated flatly: if your agent does research, competitive analysis, market research, or documentation lookups, this isn't optional.

11 Humanizer — the last check before publishing 3:51

This one already ships with Hermes, and the point made is that hardly anyone uses it. A lot of genuinely useful agent work is writing — emails, proposals, copy, social posts — and it tends to come back sounding unmistakably generated. Readers have been trained to spot it.

The clever part is the source. Wikipedia maintains a page called Signs of AI writing, cataloguing the tells in depth. Someone scraped that page into a skill that re-syncs whenever the page is updated — so the detector's own reference material keeps current instead of freezing at whatever the tells were on the day it was written.

You run the agent's output through it and get back something in a natural voice. For anyone publishing AI-assisted content — which in 2026 is most people doing this work — it's positioned as the last line of defence before something goes public.

10 youtube-full — because cloud IPs get blocked 5:06

Here the list shifts from efficiency tools to capability expanders — skills that change what the agent can do at all.

The default YouTube skill bundled with Hermes breaks on any VPS or cloud environment, because YouTube blocks cloud IP ranges. If your agent runs anywhere other than your own machine, it simply cannot read YouTube. This replacement handles transcript extraction, channel browsing, playlist parsing, and video search, with no Google API key, backed by an API that processes around 15 million transcripts a month.

The use case given is specific and worth stealing: feeding transcripts in as domain training material — for example, using YouTube scripts to teach an agent to write better YouTube scripts.

9 Composio — integrations without the OAuth tax 5:51

Over 20,000 GitHub stars. Anyone who has tried to wire Gmail into an agent knows the shape of the problem: the Google console is intimidating, the OAuth flow is easy to get lost in, and that's one tool. Multiply by every SaaS product you actually use.

Composio collapses that into a single integration surface — Gmail, Sheets, Slack, Notion, HubSpot, Salesforce, and over a thousand others. No hand-rolled OAuth, no juggling API keys per service.

The framing is that integrations used to be the bottleneck: every new tool meant new connection work before the agent could touch it. Removing that bottleneck is what makes "operate my business stack" a realistic instruction rather than a project.

8 addyosmani/agent-skills — the full dev lifecycle 6:34

The most-starred entry on the list at around 65,000 stars, from Addy Osmani, senior engineering manager on Google Chrome. It's 24 production-grade skills organised behind eight slash commands mapped to the development lifecycle: /spec before writing code, plan to decompose, build to implement, test to verify, review before merge, ship to deploy.

The standout is doubt-driven development. At each major decision mid-task, the agent stops and argues with itself: it extracts the assumptions it is making, challenges each one, reconciles the gaps, and only then continues.

7 Resemble AI Detect — screening what the agent ingests 7:31

This sounds niche until the threat model lands. Agents ingest user-submitted and scraped content — text, audio, images, video — and in 2026 a meaningful share of that is itself AI-generated. Some of it is built specifically to manipulate what an agent concludes or outputs.

The skill adds deepfake and synthetic-content detection to any ingestion pipeline: it flags AI-generated audio, images, video, and text, traces which tool likely produced it — ElevenLabs, ChatGPT, Claude — and marks it as synthetic for the agent's benefit.

If your agent scrapes broadly or leans on web search for factual detail, this is the filter that stops synthetic material entering the reasoning chain unlabelled.

6 Mission Control — a dashboard for your fleet 8:30

3,700 stars, and the start of what the video calls the infrastructure layer: tooling for running several agents at once.

The problem is visibility. Once you're running one agent for research, one for coding, and one for client communication, you have no consolidated view — no cost tracking, no status, no way to dispatch work across the fleet without going terminal by terminal.

Mission Control provides the fleet dashboard: task dispatch, agent health monitoring, real-time cost tracking, and live status across the stack. It's described as the line between an amateur multi-agent setup and one you can actually operate.

5 OpenMontage — agentic video production 9:04

12,000 stars, and the most unexpected entry on the list: an open-source agentic video production system — 12 pipelines, 52 tools, and over 500 agent skills in a single install.

You give it a reference video — a short, a reel, anything. The agent reads the transcript and analyses pacing, scene structure, key frames, and tone. Then, before generating a single frame, it returns two or three differentiated concepts, the full tool path it intends to use, a cost estimate, and a preview. You approve a direction before spend happens.

The worked example: a 60-second Pixar-style animated short — six generated motion clips, narration, music, and word-level captions — estimated at about $1.33 in total.

4 Anthropic Cybersecurity Skills 11:09

Over 700 structured skills mapped to the MITRE ATT&CK framework — the standard public knowledge base of adversary tactics and techniques drawn from real-world observation. Threat modelling, vulnerability assessment, secure code review, and incident response playbooks, all queryable by the agent.

The pitch is economic rather than technical. A solo builder or small team cannot justify a dedicated security engineer before meaningful scale — but is still shipping real products to real users holding real data. This pack covers the basics in the interval, which the video calls one of the better trades available in software.

3 oh-my-hermes — turn one agent into a committee 12:07

The problem it targets is agreeableness. Models are agreeable by default: say the output is broken and it concurs; say it's perfect and it concurs with that too. Self-verification rarely happens unless something forces it.

Inspired by the oh-my-claude skill (around 36,000 stars), oh-my-hermes turns a single session into a coordinated multi-agent workflow. It decomposes the task into subtasks, assigns specialist agents or external CLI workers — Codex, Gemini, Cursor — runs them in parallel or in staged pipelines, then verifies the result instead of stopping at a plausible-looking half-answer.

Claimed benefits: faster execution on large tasks, less manual prompting, better QA through specialist roles, live visibility via session logs, and lower token cost through smart routing — up to 50% savings — with execution persisting until verification passes.

2 make-interfaces-feel-better 13:18

There's a widely-shared article called Details that make interfaces feel better — the kind of post developers bookmark and never apply, because by the time you're building you're not thinking about it. Someone converted the entire thing into a skill, so the agent applies it automatically on every UI it builds.

What that covers: text wrapping so headlines don't orphan a single word on the last line; concentric border radius so a button inside a card has corners that actually match; contextual icon animations using opacity, scale, and blur so icons respond to interaction; tabular numbers so statistics don't visually jump as values update; and interruptible animations that don't freeze when you click too fast.

🏅 Honourable mentions 14:24

Browser Harness (15,000 stars, from the Browser Use team) connects the agent to your real Chrome browser — clicking, scrolling, filling forms, navigating as a person would. When it hits a page it doesn't know how to handle, it writes the missing helper itself and continues. Self-healing browser automation.

codebase-memory-mcp (11,800 stars) indexes an entire codebase into a persistent knowledge graph. The benchmark cited: the Linux kernel, 28 million lines, scanned in three minutes. Once indexed, exploration reportedly uses around 120× fewer tokens, because the agent queries the graph in sub-milliseconds rather than reading file by file. Support is claimed for up to 158 languages.

Loop Library tackles something subtler. Most prompts tell an agent to do a thing once. A loop tells it what to do with the result: measure it, keep the change if it helped, repeat until you hit the target, stop when it stops improving. "Make this website faster" becomes "find the slowest page, make one focused improvement, measure again, keep it only if it helps, repeat until every page hits target." The skill provides a live catalogue of pre-built loops, and walks you through designing one when nothing fits.

1 Agent-Reach — give the agent the whole internet 16:08

38,000 GitHub stars, and the reasoning for first place is a category distinction rather than a quality one.

The problem: your agent can reason, plan, and build — and then goes effectively blind the moment you ask it to find something online. Twitter is behind a paid API. Reddit returns 403 to cloud IPs. YouTube blocks VPS traffic. Ask for competitor research or what people actually say about a topic, and you get very little back.

Agent-Reach covers Twitter, Reddit, YouTube, and GitHub in one install, with no API fees. The detail that earns it the top spot is resilience: when a platform changes something and breaks one integration — which happens constantly — there's already a backup path mapped, so the agent doesn't notice the outage.

🎯 Key takeaways

1. Size is a quality signal. A good skill stays under roughly 15KB — beyond that it bloats agent memory and taxes every call, since skill metadata lives in the prompt.
2. Install the self-maintaining one first. SkillClaw prunes, deduplicates, and rewrites your library after each session, so it compounds across everything installed later.
3. Most token waste is input you never wanted. Defuddle strips page furniture before ingestion; Caveat strips accumulated session bloat. Both attack cost at the source rather than trimming outputs.
4. Cloud IP blocks are a real architectural constraint. The default YouTube skill fails on any VPS. If your agent doesn't run on your laptop, verify each integration against that environment specifically.
5. Force the agent to interrogate itself. Doubt-driven development and oh-my-hermes both exist because models are agreeable by default and won't self-verify unless something makes them.
6. Ask before you build, not after. Grill Me interviews you first; most mis-built output traces to context the agent was never given and never requested.
7. Preview cost before irreversible generation. OpenMontage returns concepts, tool path, and a price estimate for approval before rendering — a pattern worth copying anywhere spend is involved.
8. Screen what you ingest. A meaningful share of scraped content is now synthetic, some of it engineered to steer agent output. Detection belongs in the ingestion pipeline, not after the fact.
9. Security tooling substitutes for a hire you can't yet justify. 700+ MITRE ATT&CK-mapped skills cover the basics in the gap before a dedicated security engineer makes sense.
10. Multi-agent setups need observability. Once several agents run in parallel, cost tracking and health monitoring stop being nice-to-have and become the thing that makes the fleet operable.
11. Encode taste once, apply it forever. The interface-details skill works because the agent will never think about orphaned words or concentric radii on its own — but it will follow a checklist every single time.
12. Loops beat one-shot prompts. Specifying what to do with the result — measure, keep if better, repeat until target — turns a single action into convergence toward a goal.

🔗 Resources & links

• #14SkillClaw — self-evolving skill library
• #13Matt Pocock Skills — Grill Me, Caveat, Teach Me
• #12Defuddle — reader-mode markdown for agents
• #11Humanizer — strips AI writing tells
• #10youtube-full — YouTube access from cloud environments
• #9Composio — 1,000+ SaaS integrations
• #8addyosmani/agent-skills — dev lifecycle commands
• #7Resemble AI Detect — deepfake / synthetic content detection
• #6Mission Control / Minions — multi-agent fleet dashboard
• #5OpenMontage — agentic video production
• #4Anthropic Cybersecurity Skills — MITRE ATT&CK mapped
• #3oh-my-hermes — multi-agent orchestration + verification
• #2make-interfaces-feel-better — UI detail checklist
• #1Agent-Reach — Twitter, Reddit, YouTube, GitHub access
• HMBrowser Harness — self-healing Chrome automation
• HMcodebase-memory-mcp — codebase knowledge graph
• HMLoop Library — feedback-loop prompt patterns

⏱ Timestamp index