Originally published at https://tekmag.thsite.top/top-10-github-repos-this-week-adhd-ai-ponytail-and-more/
The Next New Thing is a developer show that ranks the week's most-starred GitHub repositories, and its September 11, 2026 episode centers on one idea: making AI coding assistants behave like senior engineers. The ten repos it highlights split into three groups. Output formatters reshape how an agent writes and talks. Harnesses keep the code it generates lean. And one tool trims the context window so the agent spends less on noise. Five of the ten projects verify on public GitHub. The other five could not be confirmed, so this article covers the five that do.
Key Takeaways
- The lead project: ECC, the most-starred repo of the week at about 257,000 stars, is a multi-harness system that adds skills, memory, and security controls to Claude Code, Codex, Opencode, and Cursor.
- The theme: most of the list are "skills," a lightweight format that changes how an existing agent behaves instead of replacing the model underneath it.
- The headline number: the Ponytail repo publishes a benchmark showing roughly 54 percent less generated code, 20 percent lower cost, and 27 percent faster runs when its "strategic laziness" rules are applied.
- The gap: five of the ten projects named in the episode could not be located on public GitHub when we checked, so they are reported here as episode claims, not verified repos.
Why AI Coding Assistants Took Over This Week's List
Almost every project in this week's top ten is a skill or harness for an existing coding assistant rather than a new model. They change how Claude Code, Codex, or Cursor writes code, formats its replies, and manages its context, without touching the model underneath. According to the episode, the hosts grouped the repos into skills, harnesses, and a single context-window optimizer, and that grouping is why the list leans on the layer between the model and the developer.
The pattern is familiar to regulars of this site. We flagged a similar spread of open source projects, from a large Claude plugin collection to a free Codex build, in an earlier Top 10 GitHub repos roundup, and this week's list pulls the same direction: the interesting work is in how an agent is steered, not in a new model release.
The full list comes from the Next New Thing episode published September 11, 2026, where Andrew Warner and Adam walk through the ten projects. Below, the five that verify on GitHub, then the five that do not.
ECC Is the Most-Starred Project This Week
ECC is an open-source agent harness that layers skills, instincts, memory, and security controls onto coding assistants, and it carries the highest star count on the list at roughly 257,300 stars with 38,500 forks. It is a performance system for the agent rather than a new model, and it supports Claude Code, Codex, Opencode, and Cursor from a single setup.
According to the repo's own material, ECC won an Anthropic hackathon and ships self-hosted documentation under an MIT license. In practical terms it is the standardization layer: if you run the same coding agent across several projects, ECC is the project you reach for to keep that behavior consistent across every session. You can pull it from the ECC repository on GitHub.
Ponytail Encodes Strategic Laziness
Ponytail is a skill that makes an agent prefer the smallest change that passes the tests, and its authors frame the goal as the best code being the code you never wrote. In the benchmark the repo publishes, that philosophy cut generated code by about 54 percent, lowered cost by about 20 percent, and made runs about 27 percent faster. It ships as a set of slash commands, including /ponytail, /ponytail-review, and /ponytail-audit, and it runs on Claude Code, Codex, and Grok Build.
At around 137,000 stars it is the second-most-adopted project on the list, and the README is available in English, Spanish, and Korean. The tradeoff is worth stating plainly: a lazy agent writes less, which is faster and cheaper, but it will also skip the edge cases you would have wanted it to handle. You can review the rules before installing them from the Ponytail repository.
I-have-ADHD Reshapes an Agent's Output
The I-have-ADHD project is a skill that forces an agent's replies into short, numbered, action-first steps with no preamble. It exists for developers who find a long AI response harder to follow than a compact one, and it encodes ten rules for keeping output scannable. According to the repo, it works with Claude Code and ships in seven languages, including English, Chinese, Japanese, Thai, and Vietnamese.
At about 44,000 stars it is the smaller of the two output-shaping tools on the list, and its value is narrow: it changes how the agent talks to you, not what it produces. If a verbose assistant is your main friction point, that is exactly the layer to fix. The project lives at the I-have-ADHD repository on GitHub.
Humanizer Strips the AI Tell From Text
Humanizer is an agent skill that removes the patterns that make writing read as machine generated, smoothing out the phrasing that most often gives AI output away. It works as a detector-and-rewriter pair and follows the Agent Skills standard, so any agent that supports that format can call it directly. At roughly 47,000 stars it is one of the more-adopted output-cleaning tools on the list.
The fit for a developer workflow is specific: when an agent drafts a README, a commit message, or a support reply, Humanizer is the step that keeps the result from sounding like it came from a chatbot. That is also what this very site uses to audit its own drafts, so there is a small irony in the topic. The tool is available at the Humanizer repository.
Context Mode Cuts Down on Tool Clutter
Context Mode is an optimizer that keeps an agent's context window small by sandboxes its tool output, reducing context clutter by about 98 percent according to the figures the repo publishes. It also persists session memory and routes calls across 17 platforms through MCP and hooks. The payoff is cost and reliability: a tighter context window means fewer tokens billed and fewer lost threads in long sessions.
At roughly 22,500 stars it is the least-adopted project on the verified list, but it targets a different bottleneck than the output-shapers above. It aligns with the broader direction of recent coding-model work, such as the push toward models that finish tasks with fewer tool calls, because cutting tool clutter and cutting tool calls are two sides of the same cost problem. The repo is at the Context Mode repository.
Five of the Ten Repos Could Not Be Confirmed
The remaining five projects were named in the episode but could not be confirmed as public GitHub repositories when we checked, so we are reporting what the episode said about them rather than verified repo data. They are a Zapier MCP SDK, a set of Human Layer team skills, a Show Me prompt workflow, a project called Artify for code visualization, and a Matt PCO skill collection.
One name in that list, Artify, sits close to Archify, the diagram tool that hit 49,000 stars and appeared alongside a handful of other trending projects in our Archify-led Top 10 GitHub repos roundup in early September. They may be the same project or two different ones, and we could not confirm which. The simplest read is that several of these repos are still private or unpublished, which is common for a project that trends on a video before its repository goes public.
If you work with AI coding assistants every day, the five verified repos above are a practical checklist: a harness to standardize output, a laziness layer to keep code lean, and a context tool to trim the noise. Pick the one that matches your biggest bottleneck and try it on a throwaway project before you wire it into your main setup.
Conclusion
The theme of this week's list is the gap between a capable model and a useful one. The most-starred projects, led by ECC's 257,000 stars, are all working to close that gap in the layer where a model meets a developer.
{
"@context": "https://schema.org",
"@type": "FAQPage",
"mainEntity": [
{
"@type": "Question",
"name": "What is a "skill" for an AI coding agent, and why does it dominate this list?",
"acceptedAnswer": {
"@type": "Answer",
"text": "A skill is a lightweight, usually open-source format that changes how an existing agent behaves: it adds new instructions, output rules, or a way to manage context without replacing the model. Most of this week's top ten are skills or harnesses for tools like Claude Code and Codex, which is why the list is about steering an agent rather than shipping a new model."
}
},
{
"@type": "Question",
"name": "How reliable is Ponytail's claim of 54 percent less code?",
"acceptedAnswer": {
"@type": "Answer",
"text": "It is a number the repo publishes in its own benchmark, not an independent measurement, so treat it as a directional result. The honest read is that a lazy agent writes less, which is faster and cheaper, but it will also skip edge cases you may have wanted handled. Review the rules before installing."
}
},
{
"@type": "Question",
"name": "Why does the article only cover five of the ten repos?",
"acceptedAnswer": {
"@type": "Answer",
"text": "Five projects from the episode could not be located on public GitHub when we checked, so we reported only the five that verify. Naming the unverified five and skipping their unverifiable details keeps the piece honest rather than guessing at star counts and features we cannot confirm."
}
},
{
"@type": "Question",
"name": "Are these projects free to use?",
"acceptedAnswer": {
"@type": "Answer",
"text": "The verified projects are open source. According to the repos, ECC and I-have-ADHD are MIT licensed, and the others follow the Agent Skills standard, which is an open format. The main cost is that each one still runs on a paid model, so a tool like Context Mode mainly cuts what that model bills you for."
}
}
]
}
Frequently asked questions
What is a "skill" for an AI coding agent, and why does it dominate this list?
A skill is a lightweight, usually open-source format that changes how an existing agent behaves: it adds new instructions, output rules, or a way to manage context without replacing the model. Most of this week's top ten are skills or harnesses for tools like Claude Code and Codex, which is why the list is about steering an agent rather than shipping a new model.
How reliable is Ponytail's claim of 54 percent less code?
It is a number the repo publishes in its own benchmark, not an independent measurement, so treat it as a directional result. The honest read is that a lazy agent writes less, which is faster and cheaper, but it will also skip edge cases you may have wanted handled. Review the rules before installing.
Why does the article only cover five of the ten repos?
Five projects from the episode could not be located on public GitHub when we checked, so we reported only the five that verify. Naming the unverified five and skipping their unverifiable details keeps the piece honest rather than guessing at star counts and features we cannot confirm.
Are these projects free to use?
The verified projects are open source. According to the repos, ECC and I-have-ADHD are MIT licensed, and the others follow the Agent Skills standard, which is an open format. The main cost is that each one still runs on a paid model, so a tool like Context Mode mainly cuts what that model bills you for.
Top comments (0)