The Next New Thing’s Andrew Warner and Adam Brakhane run through this week’s GitHub roundup, and it lands heavily on skills: small packages that tell a coding agent how to think before it writes anything. Andrew opens by calling it a skills-heavy week, and Adam spends the episode arguing that most of these skills package something you could just say out loud. Then OpenAI deprecates its skills repo mid-roundup, which kind of makes Adam’s point for him.

The top 10

  1. I Have ADHD — forces the model to answer with the facts instead of a wall of text. Andrew’s framing: if the AI treats you like you have ADHD, you stop losing the thread and blaming yourself. The week’s number one repo, and it has been popular before.
  2. ECC — a skill pack that installs an engineering process (plan, test, build, review) so you set it up once instead of re-explaining it in every prompt. Adam’s caveat: it matters most for people without a product or engineering background, because the model will happily build first, test later, and rebuild forever.
  3. Ponytail — encodes the veteran-with-a-ponytail instinct: find the quick path that works. The repo’s date-picker example says it — the amateur installs flatpickr, writes a wrapper component, adds a stylesheet and starts a timezone discussion; the ponytail dev uses the one in the browser.
  4. HumanLayer Skills — five small add-ons, including one that rewrites your project instructions so each rule only fires when it is relevant. That addresses a specific annoyance: Claude Code treats your instructions file as maybe-relevant and quietly ignores parts of it.
  5. OpenAI Plugins — 62 ready-made plugins, described on the show as an app store for Codex. Adam’s point lands here: connecting Codex to Notion is easy, knowing how to use Notion well is not, and the plugin ships the method alongside the connection.
  6. Archify — back from last week’s roundup: turns code or a process into an interactive diagram. Andrew’s case for visuals is that garbage is hard to hide in a chart, and the diagram gives the agent a map of where to add a step instead of re-reading the code every time.
  7. Context Mode — a memory layer that keeps junk out of the agent’s context. It advertises 98% less clutter (its own image says 99%, which the hosts noticed). Adam’s caveat is the useful part: compress a detail the model needed and it re-requests the full context, which on a 300,000-token conversation is an expensive round trip.
  8. Matt Pocock Skills — one of the most-starred repos on GitHub, and the “grill me” skill is a handful of lines: grill the user relentlessly about a plan before building it. Both hosts say the value is the concept, not the file.
  9. Humanizer — rewrites AI-written text to read as human, with a tweet showing a detector scoring the output 100% human. Adam’s prediction: short-lived, because the frontier models keep getting better at writing.
  10. OpenAI Skills — listed at number ten and, in Andrew’s words, no longer available: deprecated in favor of the plugins repository.

The skills argument in the middle

  • Adam’s position: skills often package what he would just say. He dictates with Fluid Voice, so installing someone else’s workflow rarely replaces real work, and half of what he downloads he later deletes.
  • Andrew’s position: even a skill he never installs shows him the creator’s mental model. Reading Matt Pocock’s package taught him the idea of having the AI grill him.
  • Andrew’s question to viewers: should the roundup skip skills entirely and rank repos that are not skills? He asked for reasons, not just votes.
  • The signal they read from OpenAI’s move: a bare skill is not enough. What is interesting is a plugin that bundles an integration, runnable scripts, and a set of skills designed to work together.

Viewer submissions

  • Reckoner — collapses 15 provider billing tabs into one dashboard, shows remaining credit per provider, and warns before a job dies.
  • Tim Harris Skills — 23 agent skills written for people who never read the code, which Adam notes is most people.
  • Clodex — keeps the Claude Code terminal and swaps a GPT model in underneath. Adam’s counter-advice: if you like the GPT models, try Codex itself, since the harness is the part people actually praise.
  • AI Marketing Panel — simulated customer personas that pressure-test marketing copy before a customer sees it. More detail in the persona, better the panel.

The through-line for the week: the interesting work has moved from writing code to deciding how much code should exist, when to stop and plan, and how to keep the agent’s context clean. Almost every repo on the list is a version of that instruction file.