🎯 Intro — Why Install More Skills?
Hermes ships with 90+ built-in skills and can even write new ones from workflows it sees in your chat. But those auto-created skills only come from what Hermes learns from you. The skills in this video are built by people who already solved problems Hermes can't handle on its own — and packaged those solutions into installable skills. 0:00
📝 1. Planning with Files
Problem: long conversations make the agent forget details you mentioned earlier — faster on Opus models. Too many messages in the context window → the model loses track of smaller details. 0:52
Solution: move the plan into three files in your project folder: task_plan.md (splits the main task into smaller sub-tasks), findings.md (every problem + how it was solved, reusable learnings), and progress.md (how far the agent has gotten).
🤝 2. Delegate (Turns Hermes into an Orchestrator)
Problem: Hermes already has a Claude Code skill and a Codex skill, but it won't use them unless you ask — and you have to say which tasks go to which agent in every prompt. 3:10
Solution: the Delegate Setup skill (actually a set of skills) acts as an orchestrator across every coding tool installed on your machine. It figures out which tool suits which task, hands the work over, and gives you one collective report.
| Detail | What it means |
|---|---|
| Main skill | delegate setup coordinates the others |
| Dedicated skills | Claude Code, Codex, Cursor (and more — 18 to choose from) |
| Run mode | Delegated agents run without asking permission per command (no approval stalls) |
| Flow | Identifies installed tools → maps which area each agent handles (with model name + effort level) → you approve → runs → collective report |
🔇 3. RTK (Terminal Output Filter)
Problem: every terminal command returns output the agent has to read — but 90% of it is noise that bloats the context. Too much context → harder to focus → worse answers. 4:47
Solution: RTK cuts command output and keeps only what the agent needs, before it reaches the context window. It has a database of the commands agents run most often, so it knows how to filter each one. Example: when running tests, the agent only needs the failed results — RTK keeps only failures by default.
🛡️ 4. Mantis (Google's Security Review Skills)
What it is: a set of security review skills by Google, each focusing on a different area of security review — for mobile or web apps. 7:14
Mantis Plan (plans the security review process before you begin) and Mantis Reflect (goes back over findings, makes sure failures are documented properly). Install all of them — they're separate pieces of one review process.🌐 5. Agent Reach (Access Restricted Sites)
Problem: agents get denied access to valuable research sites — like Reddit — because those sites don't want models scraping their content. 8:44
Solution: Agent Reach gives Hermes access to 15+ platforms — Reddit, GitHub, Instagram, Facebook, Twitter — using free methods (costs nothing). It was the #1 trending project on GitHub recently.
🧲 6. Skill Retriever (Save 9K Tokens/Turn)
Problem: only a skill's name and description go into the context window — but they're sent with every message. Past 100 skills, that's a huge chunk of context gone before you've said anything. 9:46
| Mechanism | Detail |
|---|---|
| Pre-model step | Runs before Hermes calls the model |
| Chunking | Splits your message into model-understandable chunks; does the same to every skill |
| Matching | Matches your message against the skill list; sends only the highest-scoring descriptions |
| Form | A plugin; enable "Skill Retriever" from the plugins area |
🔗 github.com/moonlight-lupin/agent-skills (plugins/skill-retrieval)
🧭 My Evaluation (vetted against real Hermes)
The video presents all six as equally worth installing. They aren't. I checked each claim against what Hermes actually ships today, plus the real GitHub repos. Verdict per skill:
| Skill | Verdict | Repo health (checked) |
|---|---|---|
| 1. Planning with Files | ✅ Worth it — the hook is the point | 26.6k⭐, MIT, active |
| 2. Delegate (orchestrator) | ⚠️ Mostly redundant — Hermes has native delegation | 1.7k⭐, MIT |
| 3. RTK | ✅ Worth it — but it's a tool, not a skill | 79k⭐, Apache-2.0 |
| 4. Mantis | ✅ Worth it — real Google project | 943⭐, Apache-2.0 |
| 5. Agent Reach | 🟠 Effective but skates ToS — know what you're doing | 78k⭐, MIT |
| 6. Skill Retriever | ❌ Skip — solves a problem Hermes largely handles | 53⭐, MIT |
✅ 1. Planning with Files — the hook claim holds up
pre_tool_call, plus plugin hooks through the same dispatcher). So a hook that force-injects task_plan.md / findings.md / progress.md is genuinely possible, and "the model can't ignore the plan" is a real, verifiable mechanism — not marketing.AGENTS.md, .hermes.md, global SOUL.md) that are auto-injected every conversation, and persistent memory (MEMORY.md / USER.md). The "agent forgets details" problem the video describes is also partly addressed natively. Planning with Files earns its place specifically where you need the plan to be immutable and forced, not advisory — a real difference for long, drift-prone tasks.⚠️ 2. Delegate — the orchestrator Hermes already is
delegate_task tool ("spawn isolated child agents for parallel workstreams") and built-in skills for Claude Code, Codex, and OpenCode — the exact "send a task to another coding agent" pattern this skill set recreates. The video says "Hermes won't use those skills unless you ask" — which is true of any skill, not a gap this project uniquely fills.✅ 3. RTK — legit, and the video is honest about what it is
rtk-ai/rtk, 79k⭐, Apache-2.0) is a Rust CLI proxy that filters terminal output before it hits the context — the "only failed tests, not passing ones" example is exactly right, and "60–90% token reduction on common dev commands" is the repo's own stated figure.✅ 4. Mantis — the real deal from Google
google/mantis is a genuine Google project — "a modular, stack-agnostic toolkit of security review skills for AI coding agents to autonomously find and fix vulnerabilities." Apache-2.0. The video's framing (install all of them, run via Hermes cron against a hosted app, get Slack alerts) is the single best workflow in the whole list — it uses Hermes's actual strengths (always-on, cron, messaging) for a task that genuinely needs repeating.🟠 5. Agent Reach — effective, but be clear-eyed about what it does
❌ 6. Skill Retriever — a 53-star fix for a mostly-solved problem
skill_view, and only names+descriptions ride in the index. The "11,000 → 2,300 tokens" figure is their own number from their own setup, and it presumes a very large skill library (100+ skills) that most users don't have.💡 Key Takeaways
- Half this list protects the context window, not adds features. The best skills fix a real, recurring problem.
- Hooks are Hermes' superpower. Planning with Files uses a hook to force plan files into context — "most agents don't support hooks. Hermes does."
- Delegate turns Hermes into a director, not a coder. It maps tasks to the right tool (Claude Code/Codex/Cursor) and reports collectively.
- 90% of terminal output is noise. RTK filters it before the context window — agents only need failures, not passing results.
- Cron jobs make security reviews hands-off. Mantis + Hermes' always-on scheduling = scheduled reviews + Slack alerts.
- Skill names/descriptions leak context. Past 100 skills, the index alone eats ~11K tokens. Skill Retriever cuts it to ~2.3K.
- Everything is agent-agnostic. The ideas transfer to Claude Code, ChatGPT, or any AI coding agent.