Eight Skills Every Agentic Engineer Needs: David Ondrej's 3.4K-Star Toolkit

A thirty-minute tour of an open-source skills repository that went viral — 42 skills across five categories, 3,418 stars on GitHub, endorsed by the Hermes Agent co-founder and the CTO of OpenSea. David Ondrej walks through his eight personal favourites: global agent guardrails (pre-tool-call hooks that block dangerous commands before they execute), git worktrees (isolated branches for parallel agents), VPS server management, the Go loop skill, a decisions skill, anti-sleep, a custom research agent skill built on DeepAPI, and the research agent web search skill. Every skill is shown in the repository and most come with hooks — shell scripts that enforce the rules programmatically rather than hoping the model follows them from the system prompt.

Don't use AI Agents without using these 8 skills
📺 David Ondrej ⏱️ 30:51 📅 2 August 2026 👁️ 29,788
Agent skills Guardrails Git worktrees Go loop DeepAPI

📌 The repo 0:00

davidondrej/skills3,418 stars on GitHub, MIT, Python, created 24 June 2026, pushed today. 42+ skills across five categories: agent-orchestration, ops-and-setup, research-and-web, skill-authoring, and thinking-and-docs. Most skills ship with hooks — shell scripts enforced programmatically via pre-tool-call hooks, not system-prompt suggestion.
"The CTO of OpenSea Chris Maddern said that this is one of the most useful open repos of agent skills out there. And the co-founder of Hermes Agent, Morgan Linton, called it possibly the single best repo of agent skills."

🛡️ 1–2. Guardrails & git worktrees 2:00

#SkillCategoryWhat it does
1Global agent guardrailsops-and-setupPre-tool-call hook — blocks dangerous commands (recursive delete, wipe git history, fork bombs, piping internet into shell, breaking reflog) before they execute. Ships with three hooks. "This is not just putting it in the system prompt — that is not strong enough."
2Git worktreeagent-orchestrationCreates isolated copies of the same repo on separate branches — enables parallel agents without file conflicts. "Realistically, as soon as you have more than five agents in the same repository, you need this."
The guardrails skill is the first one he shows and the first one he says to install. The rationale: YOLO mode is the only way to run multi-agent systems at scale, and YOLO mode requires programmatic protection. The hooks are shell scripts that run before every tool call and scan for dangerous patterns — not a system prompt asking the model to behave. "If any agent ever runs recursive delete of your entire machine, you can pretty much say goodbye."

☁️ 3–4. VPS & Go loop 7:15

#SkillCategoryWhat it does
3VPS server managementops-and-setupProvisions and configures a Hostinger VPS — firewall, SSH keys, Git, Docker, security hardening. "If you still don't have your own VPS server in 2026, let me show you just how easy it is."
4Go loopagent-orchestrationTurns any short, vague instruction into a specific, executable goal loop — the same mechanism Codex added earlier this year. "Give this to any agent and it will turn any short and vague instruction into a specific Go loop."
Sponsored. The VPS skill is demonstrated on a Hostinger VPS — the video is sponsored by Hostinger, and the link is in the description. The skill itself is real and in the repo; the commercial relationship is disclosed.

🧠 5–6. Decisions & anti-sleep 18:07

#SkillCategoryWhat it does
5Decisionsthinking-and-docsForces the model to surface the key decisions it made during a task — what it was thinking, what choices it weighed, and why it picked one. Inspired by a Victor Tallinn tweet. "The model will tell you: these are the four main decisions I made."
6Anti-sleepops-and-setupPrevents the machine from sleeping — essential for agents running overnight. Works alongside Ghostly and Herder. "You just want to make sure that overnight they keep running."
The decisions skill addresses a structural problem, not a convenience problem. Models make choices silently — which library to use, which architecture to follow, which trade-off to accept. Without a decisions skill, none of those choices are visible to the human reviewer. The skill extracts them. That is auditability, not convenience.

🔍 7–8. Research agent & web search 26:30

#SkillCategoryWhat it does
7Research agentresearch-and-webA custom skill he built on top of DeepAPI — orchestrates multi-step web research. "I created this one myself."
8Web search (DeepAPI)research-and-webHis most-used skill — "hundreds of times per day by my agents." Powers the web search and research category. Most skills in this category use DeepAPI (deepapi.co) as the backend.
The eighth skill — web search — is his most-used by an order of magnitude. "'Every single day' is an understatement. This skill gets used hundreds of times per day by my agents." He built it on DeepAPI, which he works with closely. The entire research-and-web category in his repo runs on it.

🔮 What this means

The video is a guided tour of a personal skills repo. The signal it sends about the category is larger than any individual skill.

Skills are the lever, and the repo is the proof. David Ondrej's repo went viral because it demonstrated something people suspected but had not seen organised: a personal collection of 42 skills, grouped by workflow stage, shipping with programmatic hooks, endorsed by the people who build the harnesses — and the message was not "look at my tools" but "you need to master this."

He is explicit about the thesis: "The ability to create and design effective, powerful agent skills is going to be one of the most valuable abilities in 2026, 2027 and the years going forward. Whoever can control the agents and whoever can best steer them — those are the people who are going to win."

That is the same thesis this site has been documenting all day: from Matt Pocock's formal language for writing agent instructions, to the Telnyx marketplace shipping 238 skills, to the Brainqub3 founder discovering a weekend's skills outperformed a six-month platform. The pattern is consistent, and David Ondrej's repo is the personal-scale proof that it works.

🔍 Claims checked

ClaimResult
David Ondrej skills repo existsdavidondrej/skills — 3,418 ⭐, MIT, pushed today. 42+ skills in 5 categories.
Global agent guardrails uses hooks, not system promptsConfirmed — three hooks at repo root, pre-tool-call shell scripts
Git worktree skill enables parallel agentsConsistent — git worktrees are the standard solution for parallel branch work
VPS skill provisions a Hostinger serverSponsored — Hostinger link in description. Skill exists in repo.
Research skills powered by DeepAPIdeepapi.co referenced — not independently verified
Endorsed by Hermes Agent co-founder and OpenSea CTOHis claim — X/Twitter posts referenced; not independently read
Method note. 30-minute video — 609 segments transcribed. Verified 5 August 2026.

💡 Key takeaways

  1. Skills write for agents, not humans. "Everything we build, it's not for humans anymore. It's for the agents."
  2. Hooks enforce rules programmatically — system prompts do not. The guardrails skill blocks dangerous commands before execution, not by asking nicely.
  3. Git worktrees are the parallel-agent primitive. Five or more agents in one repo need isolated branches.
  4. Decisions make model reasoning visible. Without a decisions skill, the model's choices are invisible to the reviewer.
  5. Web search is the most-used skill by an order of magnitude. Hundreds of times per day.
  6. Skills are the lever. "Whoever can control the agents and whoever can best steer them — those are the people who are going to win."

🔗 Resources

🕐 Timestamps

0:003.4K-star repo — 42 skills, 8 favourites
2:001. Global agent guardrails
5:172. Git worktree
7:153. VPS server management
11:534. Go loop
18:075. Decisions
23:186. Anti-sleep
26:307–8. Research agent & web search
☰ View all