📖 Read the full version with charts and embedded sources on AgentConn →
Six of the fifteen most-starred repositories on GitHub this month are agent skills frameworks. Not models. Not inference engines. Not fine-tuning toolkits. Skills — markdown files with instructions that tell AI agents how to do specific things.
last30days-skill hit 39,200 stars and the #1 trending spot by turning Claude Code into a multi-platform research engine. Superpowers reached 121,000 stars in five months by enforcing design-plan-implement discipline. agent-skills by Addy Osmani crossed 52,200 stars with production-grade engineering workflows. pm-skills carved out 100+ product management commands for a single professional domain.
As Greg Isenberg noted: "I sat down with matt van horn and watched him turn claude code into a real-time research engine with his /last30days skill."
This isn't a coincidence. It's a structural shift in where value lives in the agent stack — and it has direct implications for which tools you should be evaluating, building, and betting on.
What last30days-skill Actually Does
Most AI agents have a knowledge problem: their training data stops months ago, and web search returns SEO-optimized content rather than what practitioners actually discuss. last30days-skill solves this with an engagement-ranked alternative search engine that runs inside your coding agent.
Type /last30days followed by any topic, and the skill orchestrates parallel searches across Reddit, X/Twitter, YouTube, Hacker News, Polymarket, GitHub, TikTok, Instagram, Bluesky, and the broader web. As the explainx.ai deep dive explains, results are ranked by what real people engaged with — upvotes, likes, prediction market odds backed by real money — rather than editorial curation or SEO positioning.
The v3 pipeline works in five stages:
- Entity resolution: Before searching, the skill resolves your topic to relevant people, subreddits, X handles, GitHub profiles, and communities
- Parallel multi-source querying: All platforms searched simultaneously with engagement-based scoring
- Cross-source clustering: Duplicate stories from different platforms are consolidated
- Synthesis: An AI agent judge produces a grounded summary with citations to specific engagement data
- HTML brief generation: Optional shareable dark-mode brief for distribution
The AI Builder Club put the value proposition succinctly: "Google ranks editors. last30days ranks people."
Key stats: MIT licensed, compatible with 50+ agent harnesses (Claude Code, Cursor, Codex, Copilot, Gemini CLI, and more), 1,012 passing tests. Zero configuration for Reddit, HN, Polymarket, and GitHub; optional API keys unlock X, YouTube transcripts, and premium sources.
What makes it a "meta-research skill" rather than just another search tool is the scoring layer. Traditional web search and even Perplexity optimize for coverage and citation density. last30days optimizes for what practitioners actually engaged with — the Reddit thread with 400 upvotes where someone explains why a tool broke in production, the X post where the maintainer admits a known bug, the Polymarket contract where real money says the probability is 73%.
As the explainx.ai article notes, the skill "works best as a signal-discovery layer, identifying what deserves verification rather than serving as a truth authority." It doesn't replace deep research. It tells you where the signal is so you don't waste the first hour figuring out which platform the conversation happened on.
The Four Archetypes: Infrastructure, Curation, Domain, and Meta-Research
last30days doesn't exist in isolation. The skills explosion on GitHub trending reveals four distinct archetypes, each capturing value at a different layer:
Superpowers — Infrastructure Skills (121K Stars)
Superpowers, created by Jesse Vincent (Perl project lead, Keyboardio founder), took a different bet: instead of giving agents capabilities, enforce methodology. The framework mandates design → plan → implement workflows through composable markdown skills that trigger automatically based on context.
The growth — 0 to 121K stars in five months — signals something beyond novelty. As byteiota reported, the trajectory reflects "an industry shift from complex monolithic frameworks toward modular, skills-based architectures that actually ship production-quality code."
Superpowers answers a specific fear: that agents left unsupervised will vibe-code a pile of slop. Its anti-rationalization tables (common excuses to skip steps with documented rebuttals) and mandatory verification gates exist because 10x PRs with 1x reviewers is a real crisis, not a hypothetical one.
agent-skills — Curation Skills (52.2K Stars)
agent-skills by Google's Addy Osmani takes the "Software Engineering at Google" playbook and packages it into 24 structured workflows across six development phases: Define, Plan, Build, Verify, Review, and Ship.
The key distinction from Superpowers: where Superpowers enforces a methodology, agent-skills curates best practices. Every skill requires measurable evidence before completion. Seven slash commands (/spec, /plan, /build, /test, /review, /code-simplify, /ship) map directly to development phases, and four specialist agents (code reviewer, test engineer, security auditor, performance auditor) handle domain-specific verification.
This is curation as engineering infrastructure — turning "how senior engineers actually work" into repeatable agent behavior.
pm-skills — Domain-Vertical Skills (100+ Commands)
pm-skills demonstrates that skills aren't just for engineers. The marketplace provides 100+ agentic skills for product management, organized into plugins covering discovery, strategy, execution, and growth.
Each skill encodes a proven PM framework as domain knowledge. When you discuss product discovery, Claude automatically loads Teresa Torres's Opportunity Solution Tree methodology. Commands chain skills together: /discover runs brainstorm-ideas → identify-assumptions → prioritize-assumptions → brainstorm-experiments in sequence.
pm-skills proves the generalization: any domain with repeatable workflows can be captured as agent skills. Legal research, sales playbooks, security audits — the pattern is the same. The difference is which expert's methodology gets encoded.
last30days — Meta-Research Skills (39.2K Stars)
And then there's the meta category: skills that give agents access to information they couldn't otherwise reach. last30days bridges the gap between what models know (training data) and what practitioners discuss (live platforms). This category will expand as agents need real-time market data, live monitoring, and human-generated signals that no training run can capture.
Matt Van Horn's original launch tweet captured it: "Last 30 days of research. 30 seconds of work."
The community reception underscores how much latent demand existed for this category. The Startup Ideas Podcast declared: "This Claude Code skill just replaced my entire research workflow." It pulls Reddit threads, X posts, and web results — all from the last 30 days — synthesized into one expert-level answer. For operators who previously spent the first 30 minutes of any research task figuring out which platform had the relevant discussion, that compression is the entire value proposition.
The same pattern shows up in how practitioners actually compose these tools. Allie K. Miller built "Claudeopedia" by combining Karpathy's llm-wiki concept with the /last30days skill and a custom /wiki command — a personal knowledge base that stays current by running meta-research on its own entries. When skills combine, they create capabilities that none of the individual components could deliver alone.
The Discovery Problem: 60,000 Skills and Counting
The skills explosion has created its own bottleneck. A Show HN thread about curating 1,000 skills from 60,000+ GitHub repos highlights the scale: there are now more published skills than any practitioner can evaluate. Quality varies wildly. As AgentConn has documented in our analysis of the agent skill supply chain, 36% of agent skills have security flaws — they execute on load, and vetting them is non-trivial.
The HN thread on Agent Skills (544 points, 260 comments) surfaced the tension directly. User smithkl42 admitted: "We've defined enough skills in the last month or two that if we were to put them all in CLAUDE.md, we wouldn't have any context left for coding." The skill-proliferation problem mirrors the npm-dependency problem: abundance creates a discovery, trust, and management overhead that partially offsets the productivity gains.
The skeptics' case: HN user iainmerrick argues "any specific format for skills.md is a red herring" — well-organized documentation in plain English should suffice. User SOLAR_FIELDS reports agents "rarely invoke skills unprompted." The bitter lesson camp argues that throwing more context at better models will eventually make structured skills unnecessary.
The counter-argument: skills work today, with current models, at current context sizes. Waiting for models smart enough to not need structure is a bet against the entire trajectory of software engineering, which has spent decades proving that codified process beats ad-hoc intelligence.
The Operator's Guide: How to Evaluate Skills
If you're choosing which skills to adopt — or building your own — here's the framework that emerges from the data:
Start with what your team already does manually. The best skills encode workflows your team has already validated. If your engineers already follow a design-review-implement cycle, Superpowers codifies that. If your PMs already use opportunity solution trees, pm-skills automates that. If your researchers already triangulate across Reddit, X, and HN, last30days parallelizes that.
Check the trust signals. Star count is necessary but insufficient. Look for:
- Test coverage (last30days: 1,012 tests)
- Release cadence (14 releases = active maintenance)
- Harness compatibility (50+ hosts = not locked to one vendor)
- License clarity (MIT/Apache 2.0 = safe for production)
- Security posture (does the skill execute code? does it access credentials?)
Layer skills by archetype. Infrastructure first (Superpowers or equivalent for methodology enforcement), then curation (agent-skills for engineering workflows), then domain (pm-skills or equivalent for your team's specialty), then meta (last30days for research). Each layer builds on the previous one and compounds the value of the layers below it.
Start narrow. Install one skill, use it for a week, then assess. The biggest risk isn't picking the wrong skill — it's installing twenty and losing context budget to skill definitions that never get triggered.
Why This Matters More Than the Model Race
Here's the structural argument the trending data makes: in a world where models commoditize — where Claude, GPT, Gemini, and open-weight alternatives converge on capability — the differentiator moves to what you tell the model to do, not which model you pick.
Skills are the orchestration layer where that happens. Superpowers' 121K stars say practitioners want enforced methodology. last30days' 39K stars say practitioners want real-time information access. agent-skills' 52K stars say practitioners want codified senior-engineer judgment. The combined signal: value is migrating from the model layer to the skill layer.
This is the same pattern that played out in cloud infrastructure. AWS, Azure, and GCP converged on capability; the differentiation moved to Terraform, Kubernetes, and the orchestration layer above them. Skills are to AI agents what Terraform is to cloud — the layer where operational knowledge gets encoded, versioned, and shared.
For a curated view of the skills that work with your agent host, check the AgentConn Skills Directory — we track compatibility, security posture, and community traction across 50+ frameworks.
What to Watch
The skills ecosystem is still early. Three dynamics will shape where it goes:
Standardization vs. fragmentation. The HN thread flags real problems: incompatible directory conventions (.claude/skills vs .opencode/skills vs .agents/), no versioning standard, no dependency management. These are solvable problems, but they're not solved yet.
Security and trust. Every skill that runs in your agent has access to your codebase, your API keys, and your deployment pipeline. The supply chain risk is real, and the ecosystem needs better vetting infrastructure before enterprise adoption scales.
The validation layer. Skills generate more agent output, which needs more validation — the same judge layer pattern that emerged for agent code review will likely emerge for skill output verification.
The Takeaway
Six of the fifteen most-starred repos on GitHub this month are skills. That's not a trend — it's a market forming. The model race gets the headlines, but the orchestration layer is where practitioners are voting with their stars, their installs, and their workflows.
last30days-skill, Superpowers, agent-skills, pm-skills — each captures a different slice of the agent value stack. Together they make the case that the most important code in your AI workflow might not be the model weights at all. It might be the markdown file that tells the model what to do.
The models are converging. The skills are diverging. The teams that treat skills as engineering assets — versioned, tested, reviewed, shared — will compound that advantage every week the ecosystem matures. Bet accordingly.
Originally published at AgentConn
Top comments (0)