Theo Browne released the agent skills he uses every day

A coding agent that writes code is now the easy part. The next question is who checks the pull request before it merges.

Theo Browne, the YouTuber behind t3.gg, put his personal set of agent skills on GitHub two days ago. The pack is called tstack and already has 412 stars. It works with Claude Code and OpenAI Codex. The repository is MIT licensed.

What is in the pack

The repository holds nine skills. full-send runs an agent through the whole pull request flow. It proves the change, asks for a review, files the pull request, and merges when the review passes. full-stop does the same work but pauses before the merge, so a person can look at it. council asks three different models to review the same diff. It reports where they agree and where they do not. sol-loop runs a review with GPT Sol over and over until the model stops asking for changes. file-pr writes the pull request body around the problem the change solves rather than a list of files it touched. babysit-pr watches a pull request through CI and acts on real failures. takeover picks up somebody else's pull request and keeps the first author on the commit. modernize rebases an old pull request and asks whether it is still worth landing. leaving-pr-comments cleans up the voice an agent uses when it writes a comment.

Who it is for

An engineer who runs a coding agent and has to review every diff by hand. The skills take the small decisions off the engineer. Is the branch up to date? Did CI pass? Does the pull request body say why the change is needed? Does the agent's comment read like a person wrote it? The engineer still looks at the code. The pull request lands only after a second model has read the same diff.

How to use it

The README says to clone the repository and copy the skill files. For Claude Code, the folder is ~/.claude/skills/. For Codex, it is ~/.codex/skills/. Each skill is a Markdown file the agent reads at the start of the session. The pack needs an authenticated GitHub CLI and Python 3. The council and sol-loop skills need an account at more than one model provider.

Browne frames the pack as a starting point, not a product. Every skill is small enough to read in a few minutes. That is the whole point. An agent that follows a rule nobody can read is harder to trust. A short Markdown file next to the code is easier.

The pack is a snapshot of one person's workflow. The author gives no promise to keep it current. For a team that already runs Claude Code or Codex every day, the value is simpler. It is a chance to read how one careful user set up the loop. The parts that fit, stay.

Source

Primary source: t3dotgg/tstack on GitHub.