anthropics/skills sits at 168,968 stars. google/skills just showed up behind it at 18,056. The agent-skills standard is being contested in public, on GitHub, and whoever wins it owns the distribution layer under every agent you build for the next two years.
That fight is happening at the platform level, but the underlying idea applies to your own work too: reusable instructions compound. The better your prompt library, the less time you spend starting from scratch.
HubSpot put together a practical collection of prompts designed to help you automate repetitive work, move through everyday tasks faster, and get more useful output from ChatGPT.
The Drops
Repoanthropics/skills
168,968 stars · trending · anthropics/skills
The reference implementation for Agent Skills, the format most Claude tooling already speaks.
Here is the real problem it solves. Every agent framework invents its own plugin shape, so a capability you build for one tool dies at the next one's border. Skills give the Claude ecosystem one shared format, and the star count says the network effect already picked a side.
Quick start: browse the skill folders and copy one whose structure matches a capability you already give your agent by hand.
Catch: it is Anthropic's format, and Google just contested it in public. Watch which one your tools support natively before betting the stack.
Repogoogle/skills
18,056 stars · google/skills
Google's own Agent Skills library, the direct answer to Anthropic's move.
Use it for: watching which conventions Google keeps versus breaks before you commit to one format.
Catch: brand new; momentum and staying power unproven.
RepoDonchitos/Claude-Code-Game-Studios
23,855 stars · Donchitos/Claude-Code-Game-Studios
Turns Claude Code into a full game studio: 49 agents, 73 workflow skills, an org chart of coordinated staff.
Use it for: seeing what agents-as-staff looks like at real scale, then stealing the coordination patterns.
Catch: a reference to study, not a dependency to install.
Repooomol-lab/open-connector
4,659 stars · oomol-lab/open-connector
An open-source auth gateway wiring 1,000+ SaaS providers into agents via SDK, CLI, MCP, and HTTP.
Use it for: making "connect my agent to everything" an afternoon job instead of a month of OAuth.
Catch: plumbing, not magic; the security model is still yours to own.
MCPOpenLabs-so/openanalytics
Privacy-first web analytics with revenue attribution and an MCP server built in. No cookies, no cross-site profiles.
Use it for: letting your agent query real conversion data without shipping user data to a third party.
Catch: young project; check the MCP surface against your stack before wiring it in.
Toolhongnoul/hwatu
A fast, interruptible verification browser for AI coding agents: 35ms checks, pixel diffs, live human hand-off.
Use it for: closing the verify-before-trust loop when your agent ships UI changes.
Catch: built for coding-agent loops; overkill if you only check UI by hand occasionally.
AffiliateMindStudio
A standard getting contested in public is exactly when you do not want your agents welded to one vendor's format. MindStudio lets you build and deploy agents across 200+ models, no API keys, free tier to start. Worth a look before the dust settles.
We may earn a commission.
Start Here
Verify your agent's browser output before you trust it, not after.
Most agent UI work still ends with you eyeballing a screenshot and hoping. Wire a verification pass into the loop instead:
1. After your agent makes a UI change, run a headless check that diffs pixels against the last known-good state.
2. Interrupt on any diff above your threshold and hand control back to you, not to the agent's own judgment.
3. Let the agent proceed autonomously only on changes that verify clean.
TRY THIS: point hwatu from today's Drops at your agent's next UI change and let it interrupt you once. You will know it worked when broken layouts stop surviving three commits because nobody looked until Friday.
Free PDF: 45 Crazy Side Hustles Worth Stealing
Discover weird businesses, overlooked opportunities, and practical ways people are making money right now. Free PDF sent straight to your inbox. Bonus: Stock market + passive income research and guides.
Frontier Signals
Writer built a model to cut your token bill, not to top a leaderboard. A post-training variant of Z.ai's open GLM-5.2, aimed squarely at deployment cost. If your agent bill is the line item you dread, this one is for that complaint. (TechCrunch)
DeepSeek V4 Pro 0813 landed on OpenRouter with no announcement. Quietly routable today; worth an eval swap if you are already OpenRouter-native. (Simon Willison)
The White House authorized private security firms to hack overseas cybercriminals. A first, and it changes the threat model for anyone running agents with real infrastructure access. (Ars Technica)
|
Recommended reading
If you like The AIgent, a small group of operator-tier publications worth your inbox: see the shortlist. |
Community Pulse
This week on the Builders Board: post what you shipped, get unstuck, and tell us what to build next. The room is builders only and the replies are fast.
The AIgent is free for everyone, every day, and reader support is what keeps it that way. If it earns its place in your inbox, you can back it at theaigent.xyz/support: Supporter is pay-what-you-want, Sustainer is $5 a month. Nothing is paywalled and nothing will be.
Before You Go
Two companies shipped competing skills repos in the same week without a press release between them. That is a standard getting fought over in commits instead of courtrooms, and it is worth watching closely.
Reply and tell us which drop you opened first. We read every one.
See you Monday.
