🛠️ One Command Syncs Skills Across Every AI CLI

Here is something Congress learned this week: once an AI writes something for you, denying it gets complicated.

Rep. Anna Paulina Luna's office says Claude was used for "spellcheck" on a defense funding amendment. The summary circulating had the tells, the sentence structure, the smoothed edges. Her office says no legislation was ever written by AI. Maybe. But the gap between "spellcheck" and "drafted with AI assistance" is exactly the gap that erodes public trust in institutions that are still figuring out their AI policies while using AI every day. That story matters to us not because of the politics, but because it shows where the accountability gap lives when agents start touching consequential output.

Meanwhile the repos kept coming. Nine pass the bar today, led by a cross-tool skill-sync that the Claude Code community is actually using. I will hand you those first.

[Repo] runkids/skillshare, 2,296 stars. Syncs skills across Claude Code, Codex, OpenClaw, and other AI CLI tools with one command. If you run more than one AI coding tool (and most operators do), you have been managing skills separately per tool. This kills that tax and makes team skill-sharing instant.

[Repo] interviewstreet/hiring-agent, 2,063 stars. AI agent that evaluates and scores resumes. A real use case operators can drop into a hiring workflow today, not a demo, it runs the full evaluate-and-score loop autonomously.

[Repo] 0Chencc/clawgod, 1,446 stars. A runtime patch applied on top of official Claude Code, not a third-party client. Works across Claude Code versions as they ship. The category of "extend Claude Code without forking it" is getting real, and this is the most-starred entry in it.

[Skill] BehiSecc/VibeSec-Skill, 972 stars. A Claude Code skill that writes secure code and catches common vulnerabilities before they ship. If you vibe-code fast and audit slow, this is the guardrail that runs before you do.

[Repo] koala73/worldmonitor, 59,549 stars. Real-time global intelligence dashboard: AI news aggregation, geopolitical monitoring, and infrastructure tracking in one situational-awareness interface. Niche for most operators but extremely high signal for anyone running geopolitical or macro-sensitive pipelines.

[Repo] stretchcloud/claude-code-unified-agents, 735 stars. Unified agent architecture for Claude Code. The snippet is thin but the star count is real and it is trending in the Claude Code community today.

[Repo] agno-agi/agno, 40,835 stars. Build, run, and manage agent platforms. One of the most-starred agent-framework repos actively maintained. If you need a platform layer above raw SDK calls, this is the one to evaluate.

[Repo] traceloop/openllmetry, 7,231 stars. OpenTelemetry-based observability for LLM and GenAI apps. You cannot optimize what you cannot trace. This connects your agent stack to standard observability tooling without a proprietary lock-in.

[Repo] ConardLi/easy-dataset, 14,518 stars. Creates datasets for LLM fine-tuning, RAG, and eval. If you are building any RAG pipeline or planning a fine-tune, generating a clean eval dataset is the step most operators skip and later regret.

[MCP] getgantry/gantry, Native macOS app for managing and monitoring Docker containers, locally and over SSH, with a built-in MCP server. Free and open source. The MCP server is the part that matters: it makes Docker management an action your agent can take, not something you switch context to handle manually.

[MCP] fennaraOfficial/fennara-godot-ai, AI chat and agent tooling for Godot with MCP support, local runtime, and native editor integration. If you are building anything in Godot or exploring game-engine-backed simulation environments for agents, this is the MCP integration that exists today.

Want to run a setup like this yourself? aigent-OS ($49) →

What shipped and what is shifting:

- GLM 5.2 Fast via Wafer is live on Vercel AI Gateway. Their own benchmarking shows 2x higher throughput than competing providers across small-context, large-context, and tool-call scenarios. If you route inference through AI Gateway and latency is a real constraint, this is worth a test today. Vercel Changelog

- The White House shortened the deadline to drop quantum-vulnerable cryptography. Any agent stack that handles auth tokens, API keys, or encrypted storage over a long time horizon now has a compressed compliance window. This is not theoretical. Ars Technica

- OpenAI and Broadcom unveiled Jalapeño, a custom LLM inference chip. Operator consequence: custom silicon is collapsing inference costs at the top end, which means the open-weights cost floor will follow in 12-18 months. The model you pay per token for today will get cheaper, faster than the headline suggests. OpenAI Blog

- Simon Willison flagged a new tell in hiring: LLM-cowritten applications now link to LLM-generated portfolio sites, which link to LLM-generated GitHub projects, with plausible-but-hollow READMEs. The full-stack fake is the new spam. The hiring-agent repo in Drops today is a real answer to this. Simon Willison

- Databricks' technical founders made the case for open frontier ecosystems, arguing every company needs to build its own "Agent Cloud" rather than rent one from a hyperscaler. The operators who understand this thesis will have a moat in two years. The ones who rent forever will not. Latent Space

This week's technique: context and cost control. How to keep Claude Code sharp across long sessions without bleeding tokens.

Most operators ignore /clear until the session is already broken. By then you have paid for compaction and the context window is already drifting. The right move is proactive: clear before the task changes, not after the output degrades.

Three things to know cold:

1. /clear resets the conversation context but keeps your CLAUDE.md rules in place. Use it between distinct work chunks, not just when something goes wrong. 2. Compaction kicks in automatically when the context window fills. It summarizes the session and drops detail, sometimes the wrong detail. If you are mid-task on something precision-critical, finish the task before the window fills, or break the task into shorter chunks that each fit cleanly. 3. Long sessions get expensive and sloppy together. The fix is not a bigger context limit: it is narrower task scopes, a clean /clear between them, and a CLAUDE.md that carries forward the state you would otherwise lose. Anything you want Claude to remember across clears lives in the file, not in the conversation.

You will know it is working when your token costs per task drop and you stop seeing Claude hedge answers with "based on our earlier discussion" mid-task.

No single doc covers all of this, but the memory architecture is at Claude Code docs: memory.

Move: wire traceloop/openllmetry into your agent stack in an afternoon.

Most operators run agents blind. They know the output; they have no idea what the model called, how long each tool took, or where the token budget went. OpenLLMetry fixes that with OpenTelemetry-native tracing, the same standard your existing infra probably already speaks.

Here is the move:

1. pip install opentelemetry-sdk traceloop-sdk and add Traceloop.init(app_name="your-agent") at your entry point. Done in five minutes. 2. Every LLM call, tool invocation, and span now flows into whatever OpenTelemetry backend you already run (Jaeger, Grafana Tempo, Honeycomb, Datadog, your choice, no new SaaS required). 3. Set a span attribute on every agent task with a job ID. Now you can filter by task, not just by session, and see exactly which tool calls ate your budget.

You will know it worked when you can pull up a trace for a failing agent run and point to the exact call that went wrong, without reading the full conversation log.

Repo: traceloop/openllmetry.

This week's featured pick: aigent-OS.

Who builds with this? Operators who already ship with Claude Code and are tired of rebuilding the same memory architecture, skill stack, and session discipline from scratch every project. aigent-OS is the self-improving Operator OS for Claude Code: a structured layer that accumulates what you learn and makes it available the next time you open a session, for $49. The store link is in the button below.

Nine repos, two MCPs, five signals, and one congressional office learning the hard way that "spellcheck" is a narrower claim than it sounds when the output is that smooth.

The skills-sync repo is the one I would clone first. If you run more than one AI CLI tool, you already feel the tax of managing them separately. That repo removes it in one command.

See you Friday.