A few weeks back I wrote about Hermes Agent and what it can actually do once it learns your patterns. The question I kept getting after that piece was simpler than I expected: “okay but how do I make it smarter, not just chattier?”

The answer isn’t a better prompt. It’s agent skills.

One Topic: Make Your AI Agent Smarter With 8 GitHub Agent Skills

What is an agent skill, in plain words?

Think of your AI agent (Claude Code, Codex, Hermes, OpenClaw, whatever you use) as a smart new hire on day one. It’s capable, but it doesn’t know your tools, your shortcuts, or how your team works. A skill is a small folder of instructions you hand it, like a cheat sheet, that turns that generic new hire into someone who already knows the job. A single install gives agents a complete toolkit they can use on their own, working across Claude Code, Codex, Gemini CLI, and more.

Most agent skills come in one of three shapes: a skill pack your agent loads directly, an MCP server that connects it to outside tools, or a small command-line tool it calls when needed. Nearly all of them are free and open source. What you pay for is your AI subscription (Claude, OpenAI, etc.) and occasionally a paid API for a specific service.

Here are 7 worth knowing about, picked because each one fixes a different real gap: research, memory, file handling, security, and integration.

1. last30days skill – stop manually checking five sites for “what’s happening now”

Every AI model has a knowledge cutoff. Ask it about something from last week and it either guesses or tells you it doesn’t know. This agent skill fixes that gap by going out and checking for you.

You give it a topic, and last30days skill researches that topic across Reddit, X, YouTube, Hacker News, and Polymarket, then synthesizes a grounded summary of what actually happened in the recent past. It even picks up signals from prediction markets, which is a clever touch, since people betting real money tend to be more honest than people posting opinions.

If you’re researching a competitor’s product launch, or trying to understand what your audience is actually talking about before you write content. Instead of opening ten browser tabs and reading for 90 minutes, you type one command and get a sourced brief in a few minutes.

It installs directly into Claude Code, Codex, Hermes, or any of over 50 agent hosts that support the Agent Skills format. No setup is required for Reddit, Hacker News, or Polymarket. A short one-time wizard unlocks X and YouTube. It’s MIT licensed, completely free, and built so your research stays on your own machine rather than getting logged somewhere.

2. pm-skills – give your whole product team the same playbook

If you manage or work alongside product managers, you’ve probably seen the same problem repeat: every PM prompts their AI differently for the same task, like writing a PRD or running a competitor analysis. The output quality becomes a lottery.

This is a marketplace of 100+ agent skills, commands, and plugins covering the full product lifecycle, from discovery and strategy to execution, launch, and growth. Specific skills handle things like persona building, market sizing, roadmap creation, and go-to-market strategy. There’s even a skill that documents a vibe-coded app and audits the gap between what the documentation claims and what the code actually does, which catches a class of bug most tools miss entirely.

Instead of each PM re-explaining their process to the AI every time, the whole team installs the same skill pack. Everyone’s PRDs follow the same eight-section structure. Everyone’s output look the same. If you’re a leader trying to standardise how your team works with AI, pm-skill is the cleanest way to do it without writing your own internal documentation from scratch.

To use install through the Claude Code plugin marketplace with one command, then add only the specific skill packs your team needs (market research, execution, GTM, and so on). It also works with Cursor, Codex, and Gemini CLI. The skill definitions are open source; you only pay for whichever AI model you’re already using.

3. apple/container – give risky agent tasks their own sandbox

This one matters more than people realise once an agent starts running code or scraping websites on your behalf. If something goes wrong, you don’t want it touching your actual Mac.

It’s a tool that creates and runs Linux containers as lightweight virtual machines on a Mac, written in Swift and optimised for Apple silicon. Unlike typical setups where every container shares one big virtual machine, this gives each container its own dedicated, lightweight virtual machine, providing hardware-level isolation and a smaller attack surface.

If you’re running an agent that scrapes the web, executes code it just wrote, or connects to outside tools, you want that activity boxed off from your real files and credentials. This agent skills gives every risky task its own clean, disposable environment. If something breaks or behaves unexpectedly, it stays contained, literally.

It requires a Mac with Apple silicon and a recent macOS version, since it depends on virtualisation features Apple added at the OS level. It’s free, open source, and installs directly from GitHub. No ongoing cost beyond the hardware you probably already own.

4. Agent-Reach – one tool instead of five different scrapers

Before this kind of tool existed, giving your agent access to Twitter or Reddit meant wiring up separate, often rate-limited official APIs for each platform, or paying for access.

Agent-Reach agent skills gives your AI agent the ability to read and search Twitter, Reddit, YouTube, GitHub, and a few other platforms through one CLI, with zero API fees. Under the hood it uses the same free, open tools individual developers already rely on, rather than official paid endpoints.

This is genuinely useful for competitive research and staying current on a niche topic, without burning through expensive API quotas. It recently crossed 22,000 stars on GitHub, with over 1,500 new stars in a single week, which tells you builders are actually using this daily, not just bookmarking it.

Install with a simple pip command, then run the built-in doctor command to check which platforms are ready to go. Reddit and a couple of others need a one-time cookie setup tied to a browser session (the docs recommend using a secondary account rather than your main one, for safety). Everything else works out of the box. The only possible cost is an optional $1-a-month server proxy if you’re running it from a server instead of your own laptop.

5. firecrawl/cli and the Firecrawl Skill – let your agent read an entire website, not just one page

These two ship together, so it’s worth covering them as a pair. Most AI agents are decent at reading one page you give them. They’re much worse at exploring a whole site on their own.

The Firecrawl CLI and Firecrawl Skill teach agents how to search, scrape, interact with, and crawl websites for live web work, and separately how to wire Firecrawl into actual application code when you’re building something more permanent. Once installed, it gives the agent six core tools: scraping single pages, crawling entire sites, web search, mapping all the URLs on a domain, structured data extraction, and fully autonomous research that finds its own sources.

Say you want your agent to read your entire documentation site before answering questions about it, or pull pricing and feature data from five competitor websites. Doing this manually means dozens of copy-paste cycles. With this skill, one instruction does the whole job. The clever part is that results get written to files on your computer instead of getting dumped straight into your AI’s working memory, which keeps large crawls affordable instead of burning through your token budget.

One install command sets it up across Claude Code, Codex, OpenCode, and other detected agents on your machine at once. There’s a generous free tier to start, with paid plans once you’re crawling at real scale.

6. Qdrant MCP servers – real memory instead of a messy text file

Most personal agents fake memory badly. They either forget everything between sessions, or someone is manually maintaining a growing text file that the AI re-reads every time, which gets slow and messy fast.

This agent skill is an official MCP (Model Context Protocol) server for keeping and retrieving memories in the Qdrant vector search engine, acting as a semantic memory layer. Instead of matching exact keywords like a search bar, it understands meaning, so your agent can ask “what did we decide about the budget” and get the right answer even if those exact words were never typed.

This is the real upgrade once you’re tired of re-explaining context every session, spending token unnecessary, or once your agent is handling enough documents and conversations that a flat file genuinely can’t keep up. It’s the difference between an assistant that remembers facts and one that remembers your situation.

You run the Qdrant database itself (locally through Docker, or hosted on Qdrant Cloud), then point your MCP-aware agent at it. Qdrant the database is open source and free to self-host; cloud hosting is a paid option if you’d rather not manage a server. Embeddings, the way text gets converted into something searchable by meaning, can run through OpenAI or free local models.

7. Headroom – the one that actually lowers your AI bill

If you’re paying for Claude Code or Codex by usage and feeling that bill creep up, this is the agent skills to install first.

Headroom compresses tool outputs, logs, files, and retrieval chunks before they ever reach your AI, claiming 60 to 95 percent fewer tokens for the same quality of answers. A real example from its own documentation: a 10,144-token log file got compressed down to 1,260 tokens, and the AI still caught the same error it would have found in the full version.

Beyond the obvious cost savings, it adds a feature that mines your agent’s failed attempts and quietly writes corrections back into your skill instruction files, so the same mistake doesn’t repeat. There’s also shared memory across different agents, so Claude, Codex, and Gemini can draw from the same compressed context instead of starting fresh each time.

Like other, install with a single pip or npm command. It works as a library you call directly in code, a proxy that needs zero code changes, or a simple wrapper command around your existing agent. It’s open source, so the only thing you’re really paying for is the AI usage you’re now using less of.

How to actually start

Don’t install all eight agent skills at once. Pick the one that maps to your biggest actual pain this week.

If your problem is stale information, start with last30days-skill or Agent-Reach. If it’s your AI not knowing your whole website, documentation or need better internet access, start with the Firecrawl skill. If you’re worried about cost, headroom pays for itself fastest. If you’re leading a team that needs consistency, pm-skills is the one to standardize on.

Run that single one for a month on a real recurring task. Measure what it actually saved you. Then decide if you need a second one.

---

Interested in travel or photography, read last week’s LensLetter newsletter about Why Photographers Experience Travel Differently

Read last week’s JustDraft about key takeways from London Tech Week

---

Two Quotes to Inspire

The best leaders don’t chase every new capability. They ask which one removes the most friction from their team’s worst week, and start there.

The real advantage is not having more AI tools. It is giving one agent the right memory, rules, and access.

---

One Passage Summary From My Bookshelf

All the mental energy you use to elaborate your misery would be far better used trying to find the one seemingly impossible way out of your current mess. Spend zero time on what you could have done, and devote all of your time on what you might do. Because in the end, nobody cares; just run your company.

From The Hard Thing About Hard Things by Ben Horowitz