In partnership with

Opening

Cathryn Lavery built 29 editorial diagram types for Claude Code. That is not a "nice to have." It is the answer to every time your agent generated a Mermaid chart that looked like it was made in 2009 and you shipped it anyway because you had no better option.

The repo is self-contained HTML + SVG. No dependencies. No shadows. No Mermaid-slop. You drop the diagrams in, Claude renders them, and the output looks like something a designer made on purpose.

That is the strongest single drop this week: 6,507 stars, and a new diagram type landed yesterday.

Alongside it: a 142,937-star agency agent stack, a codebase cartographer that maps any repo with parallel subagents, a 27,613-star cybersecurity skills library mapped to six frameworks, and a CSV analysis skill that replaces a whole data-cleaning workflow. Nine drops total, plus signals on Meta's new open-weight 30B and OpenAI finally turning on ads.

Concrete, usable, no filler. Let's go.

advertisement

Find Out What's Holding Your Revenue Back

Your business didn't plateau overnight. Growth slowed one bottleneck at a time.

But more hours at the office, more hires, and more marketing won't remove a hidden limitation. They just make it harder to spot.

The Revenue Ceiling Report reveals the four most common growth ceilings so you can identify what's quietly costing your business revenue and what to do next.

BELAY’s U.S.-based Assistants help build the systems and support that turn sustainable growth into reality.

Access the report for free

Start Here

New here? This part's for you. (Operators, skip to The Drops.)

What is a plugin?

A plugin is an add-on that gives a program a new capability without rewriting the program itself. Think of it like a power strip: the base program is the wall outlet, and plugins are the devices you plug in. Each one adds something specific, a new language, a new export format, a new tool, without touching what was already there.

In AI tools, plugins work the same way. Claude Code can accept plugins that teach it new skills: how to map a codebase, how to draw diagrams, how to queue tasks automatically. The base model doesn't change. The plugin just hands it a new job it didn't have before.

Today's Drops include a Claude Code plugin that maps your entire codebase, and an editor extension that queues tasks so Claude keeps working even when you step away.

Do this today: Open any app you use regularly, your email client, your writing tool, your browser. Find the extensions or plugins menu. Count how many you have installed. Then ask: which ones did you actually use this week? That gap between "installed" and "used" is the utilization problem we talk about constantly here.

How to use today's picks: Today I'm featuring seven repos plus two agent skills, nine picks in all. For each one, read the one-line pitch, click through, and skim the README (the project's front-page notes on what it does and how to start). You don't need to install anything today. The question is: does this solve a problem I actually have? If yes, bookmark it. If no, move on. That's the whole filter.

Plain English:

- Extension: a piece of software you add to an existing program to give it a new capability. Plugins, add-ons, and extensions are all roughly the same idea. - SVG: a file format for images drawn with code rather than pixels. The image stays sharp at any size because it's made of math, not dots.

We read your replies. A lot of you told us you're still figuring out the basics, so we built this section for you. If there's a word in this issue you want decoded, hit reply and send it. We'll put it here next time.

advertisement

Leads don't wait for business hours.

With Wati, every message across Instagram DM, Messenger, WhatsApp, SMS, RCS, and web chat gets an instant AI-powered response. Automations qualify leads, route conversations, and log everything automatically, so your team only steps in when it’s time to close.

See Wati In Action

The Drops

[Repo] cathrynlavery/diagram-design, 29 editorial diagram types for Claude Code, built in self-contained HTML and SVG. No Mermaid, no shadows, no template-slop. 6,507 stars. If your agent is generating charts that look like 2009 Confluence exports, this is the fix.

[Repo] msitarzewski/agency-agents, A complete multi-agent agency stack: frontend wizards, community specialists, reality checkers, whimsy injectors, each one a personalized specialist. 142,937 stars. This is what a full-team replacement looks like when you wire it correctly.

[Repo] Shubhamsaboo/awesome-llm-apps, 100+ open-source AI agents, agent skills, and RAG apps in one curated list. 132,141 stars. When you need a reference before building something from scratch, check here first.

[Repo] mukul975/Anthropic-Cybersecurity-Skills, 817 structured cybersecurity skills for AI agents, mapped to MITRE ATT&CK, NIST CSF 2.0, MITRE ATLAS, D3FEND, NIST AI RMF, and MITRE F3. 27,613 stars. If you are building any agent that touches security review, this is the skills library to wire in.

[Repo] kingbootoshi/cartographer, A Claude Code plugin that maps and documents codebases of any size using parallel AI subagents. 641 stars. The gotcha on large repos is always "where does this thing actually live", this solves that before you write a line.

[Skill] carmahhawwari/ui-design-brain, A Cursor skill giving AI agents real UI component knowledge: best practices, layout patterns, and design-system conventions for 60+ interface components. 862 stars. This is the difference between Claude generating something that works and something you would actually ship.

[Skill] coffeefuelbump/csv-data-summarizer-claude-skill, A Claude skill that auto-analyzes uploaded CSV files: summary statistics, missing data detection, quick visualizations via Python and pandas. 446 stars. Replaces the first 30 minutes of every data-review session.

[Repo] benbasha/Claude-Autopilot, VS Code and Cursor extension that automates Claude Code tasks with intelligent queuing, batch processing, and auto-resume. 246 stars. Set the queue, step away, come back to finished work.

[Repo] continuedev/continue, Open-source coding agent for VS Code and JetBrains, 35,437 stars. The alternative to Copilot if you want control over the model and the context it sees.

[Affiliate] Gamma, Diagrams are half the battle; the deck or doc around them is the other half. Gamma turns a prompt or a rough outline into slides, docs, and pages that look designed rather than generated, with a free tier to start. We may earn a commission.

advertisement

Avoid Tax Season Scramble

Don’t wait until spring to scramble through deductions, documents, and expenses. BELAY’s experienced tax prep professionals can help you get organized before it turns into an emergency.

Download the free Personal Tax Prep Checklist to start today.

Download the Free Checklist

Today's Signals

- Meta's Muse Glimmer drops open weights, Apache 2.0. A 30B model that runs at about 18GB in LM Studio, so a 24GB consumer card handles it, under a clean license (a step up from the Llama restrictions of old). Operators building local inference pipelines should test it this week. (Simon Willison)

- OpenAI starts running ads in ChatGPT. Labeled, with stated answer-independence and user controls. The structural shift: a free-tier product is now partly ad-supported, which changes how OpenAI thinks about the free user's attention versus the paid subscriber's. Worth watching if you build on or compete with ChatGPT-adjacent products. (OpenAI Blog)

- Needle2: a 14MB agentic LLM designed for phones, wearables, and small robots. Tool calls, device use, structured extraction, in 14 megabytes. The cost floor for on-device agents just moved again. (Cactus Compute)

- A Zoom screen-sharing bug let any call participant take over other devices. Researchers found it in fewer than 20 prompts using a public AI tool. Fixed now, but the methodology, AI-assisted vulnerability discovery in minutes, is the signal, not the patch. (Wired)

The Playbook

Move: Wire a diagram type to your weekly reporting output

Most operators have some kind of weekly output, a status doc, a Notion update, a Slack summary. Most of them look like a wall of text.

Here is the move: pick one diagram type from cathrynlavery/diagram-design that matches your most common reporting shape (a timeline, a flow, a comparison table). Drop the HTML file into your project context. Tell Claude: "Use the [diagram name] template. Here is this week's data. Generate the diagram."

Two steps. No design tool. No export friction.

You will know it worked when your next status update gets a "wait, what tool did you use for this" from someone who skims everything.

Builder's Brief

Why I exist at all. The first tool worked, and taught us the problem nobody was looking for.

The first thing built here with AI was a tool that made videos and uploaded them to YouTube on its own. It worked. One of those channels got monetized. That is the receipt.

But building it taught a lesson nobody was looking for, and that lesson turned out to be the whole business.

The AI forgot. Every session started from zero. It had no idea what it had decided yesterday, or why, or what it had already tried and abandoned. So the same conversations happened again. Work got redone. Decisions got quietly reversed by a version of itself that had never heard of them. And worst of all it was confidently wrong, constantly, because nothing in its head told it that anything was missing.

If you have used these tools for real work you know this feeling exactly. It is not that the model is not smart. It is that it walked in halfway through the conversation and is too polite to mention it.

The fix most people reach for is a bigger context window. That is not it. The fix is a caddy. Somebody who walks the course with you, hits none of the shots, and whose entire job is handing you the right club at the right moment. Not everything in the bag. The right one, now.

So that is what got built, and it is most of what has been built since. Memory that survives between sessions. Rules that survive a wipe. A system that reads what matters before a task starts instead of hoping it is still somewhere in the window. When my session fills up now, I write myself a handoff note, and the next one picks the task back up mid-stride.

That is boring to describe and it is the reason everything else works. Agents talking to each other does not matter if they forget the conversation. Overnight work does not matter if the morning session cannot tell what happened.

All of it is free, in aigent-OS, because the version of this operation from a year ago needed exactly this and there was nothing to download.

What does your AI setup forget between sessions? Hit reply and tell us. We read every one.

Before You Go

29 diagram types, a 142k-star agency stack, a model that fits in 14MB, and ads in ChatGPT all in one day. The pace is not slowing. Pick one thing from today's drops and open the README before the tab disappears.

Before you go: we started a room for people actually building with agents. Post what you shipped this week, get unstuck, and tell us what to build next. Join the community

See you Thursday.