⚔️ Your Model Access Just Became a Political Variable

There is a 1,600-skill library for Claude Code on GitHub right now, and I would bet most of you have not touched it yet.

antigravity-awesome-skills crossed 42,000 stars. It installs in one command. Inside: specialized skills for Claude Code, Cursor, Codex CLI, Gemini CLI, and more, organized into bundles you can activate by category. Marketing ops, code review, architecture analysis, writing workflows. Not prompts stapled to a readme. Actual installable skill definitions that drop into your agent's context layer and change what it does by default.

The drop pattern this week is skills. Three skill-adjacent repos in today's list. The best operators I know are not prompting harder, they are building reusable capabilities that carry across every session. That is the whole thesis. One install, every session gets sharper.

The governance story is worth noting too: OpenAI and Anthropic are now gating frontier models behind White House approval. Anthropic separately cut a deal giving California state government access at half price. Both moves in the same week. The model access game is getting political in ways that matter for your stack planning.

Nine repos today, two of them skills libraries you can actually install before lunch. Let's go.

[Repo] agency-agents, A complete multi-role agent crew: frontend wizards, Reddit community ninjas, reality checkers, whimsy injectors, and more. Each agent is a specialized expert with its own persona and toolset. Clone it when you need a crew, not a chatbot. 118,726 stars and trending today.

[Skill] antigravity-awesome-skills, 1,600+ installable agentic skills for Claude Code, Cursor, Codex CLI, and Gemini CLI. Comes with a CLI installer, bundles, and workflow packs. This is the skill library you build your stack on top of. 42,021 stars.

[Repo] awesome-llm-apps, 100+ agent and RAG apps you can actually clone and run, not just read about. The breadth here is the value: deep research agents, coding assistants, document analyzers, multimodal pipelines. 116,056 stars.

[Repo] logto, Authentication and authorization infrastructure for SaaS and AI apps, built on OIDC and OAuth 2.1 with multi-tenancy, SSO, and RBAC baked in. The piece most agent builders bolt on last and regret. 12,562 stars.

[Skill] ai-marketing-skills, Open-source AI marketing skills covering growth experiments, sales pipeline, content ops, outbound, SEO, and finance automation. Drop these into your agent and stop writing the same growth prompts from scratch. 2,746 stars.

[Repo] ragflow, RAG engine that fuses retrieval with agent capabilities to build a proper context layer for LLMs, not just a vector lookup. The production-grade RAG path most hobby setups skip. 83,869 stars.

[Repo] LlamaFactory, Unified fine-tuning interface for 100+ LLMs and VLMs. When you are done prompting and ready to train, this is the tool. Covers LoRA, QLoRA, and full fine-tune paths in one place. 72,765 stars.

[Repo] Prompt-Engineering-Guide, The canonical guide to prompt engineering: papers, lessons, notebooks, and worked examples covering RAG, agents, and context engineering. The reference you send when someone asks where to start. 76,072 stars.

[Skill] Vibe-Skills, All-in-one AI skills package for general-purpose agents: expert-level capabilities and context management in a single install. Useful when you want a capable default baseline without building every skill yourself. 2,350 stars.

[MCP] metaharness, Scaffold your own branded agent harness with a CLI, MCP server, memory layer, learning loop, and witness-signed releases. Works with Claude Code, Codex, and others. The move here is building a reusable harness once instead of hand-wiring every new agent project from scratch.

[MCP] CodeDebrief, Local-first static analysis that turns source code into deterministic, source-grounded workflow maps for coding agents via MCP. The non-obvious use: feed it to Claude before a refactor and get a map of what breaks before you touch a line.

OpenAI and Anthropic are gating frontier models behind White House approval. Both labs restricted new model access at the Trump administration's request during an ongoing cybersecurity review. For operators: if your stack depends on frontier access, model availability is now a political variable, not just a technical one. SecurityWeek

Anthropic cut a half-price deal with California's state government. As the federal government tightens access, Anthropic is deepening its relationship with California's public sector, giving state agencies Claude access at 50% of standard pricing. The operator read: government vertical is moving fast on Claude adoption. TechCrunch

China's Z.ai released GLM-5.2, an open-weight model with repository-scale coding and vulnerability discovery. No vendor guardrails. Rivals top US models on code benchmarks, at a fraction of the cost. The threat is not the model, it is what it means that capability at this level is now ungated. Forbes

Tidal stops paying royalties on AI-generated music, effective immediately, with AI labeling from July 15. Not a ban, but a hard demonetization line. Every platform is going to draw this line somewhere. If you are building in any content-adjacent vertical, watch where the line lands in your sector. The Verge

Nvidia's AI chip sales in China have stalled as Huawei takes the lead. Huawei is filling the gap left by export controls. For operators building on GPU infrastructure, the supply-chain picture underneath your cloud provider is shifting, and cost floors will follow. AP News

This week's technique: custom slash commands. Codify any workflow you repeat into a single /command call.

Most operators know skills exist. Fewer know you can write your own slash commands that go beyond skills: a command can trigger a multi-step workflow, set context, load a file, or run a pre-defined agent task, all from one /name.

Here is how to build one:

1. Create a markdown file at .claude/commands/your-command-name.md inside your project (or ~/.claude/commands/ for user-wide commands). 2. Write the command body in plain markdown: the instructions Claude should follow when you invoke it. Reference $ARGUMENTS to pass a value at call time (e.g. /review src/api/routes.js). 3. In session, type /your-command-name (or /your-command-name ). Claude runs the full workflow defined in the file.

You will know it worked when you can replace a five-message setup routine with one /command and Claude executes the full workflow without any re-briefing.

The gotcha: command files are just markdown, so vague instructions produce vague execution. Treat the command file like a CLAUDE.md for that specific task: precise, ordered, with explicit output format.

Claude Code docs: custom slash commands

The move: install a skill bundle and test it against a live task before committing it to your stack.

Most people install a skills library and forget to validate it. Here is the five-minute version:

1. Install antigravity-awesome-skills using the CLI installer (npx antigravity install per the repo readme). 2. Pick one bundle that matches a workflow you already do, marketing ops, code review, or writing. Activate that bundle only. 3. Run your most-repeated task with the skill active. Compare the output quality and context usage against your baseline.

You will know it worked when Claude's first response on that task is closer to your finished output than your usual draft, without extra prompting.

The reason to test narrow first: 1,600 skills in context at once is a cost and coherence problem. Operators who load everything and notice "it got weird" are the ones who skip the bundle selector. Activate what you need, measure it, then expand.

We build The AIgent's engine in the open. Each week, an honest look at what we are making, what broke, and where it is headed. This week begins a story told in parts. It is called FlowStack, the machine that dreams in pictures. Part one.

It started with a frustration anyone who has run a content channel knows in their bones. The work is endless, and most of it is not the creative part. Write the script. Find the visuals. Generate the voice. Stitch it together. Fix the framing. Write a title that does not get buried. Upload it. Schedule it. Do it again tomorrow, and the day after, forever.

FlowStack began as a refusal to accept that. The idea was almost arrogant in its ambition: what if a machine could do all of it? Not one slice, but the whole pipeline. From a blank idea to a finished, SEO-tuned video sitting live on YouTube, Instagram, and Facebook. The vision was never a tool. It was a factory that makes art.

And not for one channel. For a whole constellation of them, each with its own soul. One for the meditators. One for the music drifters. One for the people who just want to be told a story. Each with its own look, its own pace, its own reason to exist. The bet was that a single machine could hold a dozen distinct creative identities at once, and let every one of them stay completely itself.

What made it strange and wonderful is that the channels grew personalities. Not as a metaphor: they became characters. There was Veile, the nocturnal one, all deep indigo and smoky violet and burnished gold, a figure half-swallowed by shadow on rain-slicked streets. Veile was the most beautiful and the most difficult, and we will get to why. Around her grew a whole cast, each with a fingerprint: a palette, a mood, a way of moving the camera, a way of speaking.

The breakthrough that made a cast possible had a name: the Archetype. Instead of hardcoding how each channel works into the guts of the app, everything got driven by a configuration that could be reasoned about, merged, and overridden: the visual style, the platform settings, the pacing, the voice. It meant a new channel could be born without writing a single new line of channel-specific logic. That was the quiet triumph nobody outside the codebase ever saw, and the reason any of this scaled past one channel. Personality, it turns out, is a feature, not decoration.

None of it came easy, and the series ahead is the honest version. The bug where videos kept coming out shorter than their own audio, hunted for days and fixed by two characters of config. The meditation that suddenly played too fast because a branding intro lied about its frame rate. Veile, whose gothic soul the image models flatly refused to draw, until we taught the machine to whisper what it would not let us shout. Days of investigation for fixes you could write on a napkin. That is software.

This is part one. The struggles, the self-healing fixes, and the lessons they beat into us are coming, week by week. We are not going to tidy it up. You get the wins and the tangles both, because the tangles are where the real work actually lives.

Building something like this yourself, a machine meant to run while you sleep? Hit reply and tell us where it is breaking for you. We read every one.

Nine repos, two skill libraries you can install before your next session, and a White House footnote on your model access. The skills layer is where the compounding happens now. Build it once, carry it everywhere.

See you Wednesday.