Source: raw/10_Hermes_Agent_Skills_You_NEED_To_Install_Today.md and raw/10_NEW_Hermes_Agent_Skills_You_Need_To_Install_This_Week.md (auto-caption transcripts)
Creator: Sharbel (YouTube @sharbelxyz) · URLs: https://www.youtube.com/watch?v=5WQk9WiWcYQ (13:19, uploaded 2026-08-16) and https://www.youtube.com/watch?v=bfPTHS9r7gs (12:14, uploaded 2026-10-25) · No sponsors named. Both promote his own Mission Control dashboard repo; the August list also mentions his “Nova” YouTube skill. Repo links are in the video descriptions, not the transcripts.
Two ten-item lists of community skills, plugins and MCP servers that Sharbel runs on Hermes. Most also work in Claude Code or other agents. The repos were not opened for this article, so names come from the captions and the effect claims are his or the repos’. Confidence is low for that reason.
Key Takeaways
- His August advice: start with the three that cut cost, then add abilities. “Skills are free to install. What costs you is burying the good ones under 40 that you never use.”
- Two measured numbers. Codebase Memory claims 99% fewer exploration tokens across 158 languages; he measured 95% in his own test. Caveman claims 65% fewer output tokens; he shows a side-by-side but gives no number. A JetBrains test relayed elsewhere in this wiki put Caveman’s real saving near 8.5% (see Token Optimization).
- Read each skill before installing. Several are platforms or MCP servers rather than
SKILL.mdfiles, and the security pack’s “Anthropic” name is not official (his warning).
August 2026 list (cost first, then abilities)
| Item | Kind | What it does, per the video |
|---|---|---|
| Defuddle | Skill | Strips navigation, footers, cookie banners and ads so a web page arrives as clean Markdown; “not optional” if your agent does research |
| Caveman | Skill | Ultra-compressed output mode; drops preamble, recaps and offers of more help |
| Codebase Memory | MCP server | Indexes a repo into a graph once; the agent queries the map instead of rereading files |
| Humanizer | Skill (29.6 KB) | Rewrites agent text to remove AI tells, built from Wikipedia’s “Signs of AI writing” page and auto-updating with it. Large, so invoke it late, once, ideally in a fresh session |
| Agent Reach | Skill | Reads X, Reddit, GitHub and other platforms without API fees, with backup routes when one path is blocked (Reddit returns 403 to cloud IPs, he says) |
| Marketing skills | Pack of 49 | SEO, paid ads, conversion, copywriting, buyer psychology; “judgment”, not tools. His first install if you sell anything |
| Composio | Platform | One install for 1,000+ app connections (Gmail, Sheets, Slack, Notion, HubSpot, Salesforce) with no API key or OAuth setup by hand |
| Open Montage | Video system | 12 pipelines, 100+ tools, 700+ skill files. Reads a reference video, then proposes 2-3 concepts, the tool path, a cost estimate and a preview before spending anything |
| oh-my-claudecode (captions: “Oh my cloth”) | Orchestration | Splits a task into subtasks for specialist agents in parallel or staged pipelines, then verifies the result; his quality gate for irreversible work |
| Superpowers | Skills framework | ”Over 260,000 stars”; a method for breaking work down and executing it. “Takes an afternoon to properly understand” |
October 2026 list (counted down to his top pick)
| # | Item | What it does, per the video |
|---|---|---|
| 10 | Diagram design | Clean flowcharts, architecture diagrams and explainers in your brand colours |
| 9 | Research skills library (unnamed) | Data analysis and scientific-method skills; he ran a significance test on his video stats and got a chart with a p-value |
| 8 | Security pack (unnamed) | 800+ security skills; probes your app like an attacker and returns a ranked fix list. Community-made despite “Anthropic” in its name |
| 7 | Taste skill | Design rules for spacing, type, layout and motion; reworks generic AI-looking pages |
| 6 | HyperFrames | Writes a video as an HTML file and renders it; his test was a 10-second product intro from one sentence |
| 5 | Agent Reach | Repeated from August |
| 4 | GBrain | A local second brain stored as a connected map; Hermes-native; can ingest an Obsidian vault and answer with sources |
| 3 | Concise-answer skill (unnamed) | Leads with the answer and cuts filler |
| 2 | Planning with Files | Native Hermes plugin; writes the plan, steps, problems and progress to files and puts them back in front of the agent every turn, so a crash or /clear does not lose the plan |
| 1 | Ponytail | Makes the agent think like a lazy senior developer: does this code need to exist, can it reuse something, is it built in? His test: the normal agent wrote three files, Ponytail reused existing code in a few lines |
Try It
- Install Defuddle first and compare the token count on one article with and without it.
- Before installing anything else, run
/context all(Hermes) or/context(Claude Code) and note the baseline, then check again after each install. - Give the source repo link to your agent and ask it to scan the skill for malicious code before it installs.
- Keep Humanizer out of your always-loaded set; call it for one final pass.
Open Questions
- Exact repo names and owners for the unnamed October items (research library, security pack, concise-answer skill) and for “Marketing skills”.
- His Codebase Memory test conditions (repo size, task) are not given.
- Whether “Planning with Files” is first-party Nous or community is not stated.
Related
- Claude Code Token Optimization — independent tests of Caveman, RTK and Ponytail
- Superpowers — the framework he calls an operating system for agents
- oh-my-claudecode — the orchestration layer in the August list
- Open Montage and HyperFrames — the two video tools
- Wiring GBrain into Hermes — setting up the second-brain pick
- Anti-AI-Slop Guide — where Humanizer fits among writing-cleanup tools
- Writing Your Own Hermes Skill (Sharbel) — his case for building your own instead