Open Source Radar — September 6: write less, install more
Two of the day's three biggest star movers are not frameworks or models — they are opinions about how your agent should behave, shipped as installable rules. A third is one engineer's entire working method, published as a skill library. The through-line: the argument about agents has moved from capability to restraint.
Ponytail (JavaScript, ~128,500 stars, +2,845 today)
The biggest mover on the daily board, and the funniest: a skill that makes your agent think like the laziest senior engineer in the building. Before writing anything, it walks a ladder — does this need to exist, is it already in the codebase, does the standard library do it, is there a native platform feature, is it already a dependency, can it be one line — and only then writes the minimum that works. The maintainer's example is the date picker: a bare agent installs a library, writes a wrapper, and opens a timezone discussion, while ponytail reaches for the native HTML input and ships 23 lines instead of 404.
What makes it more than a joke is that the numbers were rebuilt after a fair public critique. The original single-shot benchmark claimed 80 to 94 percent less code, which a critic correctly pointed out partly measured a chatty baseline model's prose. The corrected version runs headless Claude Code sessions against a real FastAPI and React repo, twelve feature tickets, same agent with and without the skill, and counts added lines in the actual diff: 54 percent less code on average, 22 percent fewer tokens, 20 percent cheaper, 27 percent faster, and no safety regression — while a plain "prefer one-liners" prompt dropped a safety guard. It also ships review and audit commands that hand back a delete-list.
ECC (JavaScript, ~250,400 stars, +1,314 today)
You may know this one under its old name — everything-claude-code now redirects here, and the repo has been rebuilt into what the author calls an agent harness performance system: 68 agents, 286 skills, and 94 command shims covering planning, test-driven development, security review, build repair, and architecture, plus hooks, memory, and a scanner called AgentShield that checks prompts, hooks, MCP config, permissions, and secrets. The organizing idea is stated as a rule: optimize the context window, persist everything else. Skills load only when the task needs them, agents isolate planning from implementation from review, rules are always loaded so you install them selectively, and "instincts" are patterns learned from past sessions with confidence scores that get recalled when relevant.
The interesting part for anyone running more than one agent is the Memory Vault, a plain local Markdown store under the project and in your home directory that Claude, Codex, Hermes, OpenClaw, and Kimi can all read, so context survives a switch of harness. It is MIT-licensed and free, with a paid hosted GitHub App for private repos funding the work — and the README carries an explicit warning to install only from the official repo, npm packages, or the project site, because third-party mirrors have appeared.
mattpocock/skills (Shell, ~253,500 stars, +2,692 today)
The second-biggest mover, and the one with the clearest editorial position. Matt Pocock's pitch is aimed squarely at the spec-driven workflows — GSD, BMAD, Spec-Kit — that try to own your whole process: they take away your control and make bugs in the process hard to resolve. His alternative is small, adaptable, composable skills drawn from decades of engineering practice, organized around four failure modes he says he hits constantly with coding agents.
The first is misalignment, fixed by a "grilling" session where the agent interrogates you about what you're actually building before it touches code. The second is verbosity, fixed by building a shared domain language — a document that lets the agent say "materialization cascade" instead of a sentence, which pays off in naming, navigation, and token spend. The third is broken code, fixed by feedback loops: types, browser access, and a red-green-refactor discipline. The fourth is the ball of mud, addressed by a survey skill that scans a codebase for places to deepen module design and hands you candidates rather than attempting a rescue. Install it as a managed plugin that updates itself, or copy the files in and make them yours — the author's stated preference is that you fork it.
humanlayer/skills (TypeScript, ~2,900 stars, +442 today)
The smallest thing here and the least self-promotional: five focused Claude Code skills from the team behind HumanLayer, each one narrow enough to read in a minute. The most useful is probably improve-claude-md, which rewrites your project instruction file using conditional blocks so the rules actually hold instead of being skimmed past. narrow-react-prop-types tightens component prop types to match the code paths that really run, rather than the union of everything Storybook, tests, and mocks ever pass in.
The other three are more ambitious. build-iterated-agentic-loop sets up a repo-local skill plus a coding-agent workflow in GitHub Actions, with a prompt, a memory file, and reference templates, so an agent can iterate on a problem on a schedule. design-control-loop interviews you about your codebase as a control system — sensor, controller, actuator, disturbances — then builds it as components you can run locally. And show-me explains whatever you're looking at with diagrams and small HTML artifacts. This is what the skills category looks like when it stops being a prompt library: narrow, opinionated, and built to be pulled apart.
Worth watching this week: the highest-leverage thing you can install right now is a constraint, not a capability. Three of today's four picks exist to make your agent write less.
If you had to pick one rule to force on every agent you run — delete before you add, ask before you build, or remember across sessions — which would it be? Tell us in the comments.
Sources: Ponytail (GitHub) · Ponytail agentic benchmark · ECC (GitHub) · mattpocock/skills (GitHub) · humanlayer/skills (GitHub)
