Links, tools, and ideas I'm collecting along the way.
75 links · Updated March 2026
No links found
Browser Use CLI 2.0 — 2x the speed, half the cost. Connects to running Chrome via direct CDP for efficient browser automation.
Direct CDP integration is the right call for speed. This could be a solid tool for agent-driven browser tasks without the overhead of full Playwright.
Claude Code Channels lets you control your Claude Code session through MCPs, starting with Telegram and Discord — message Claude Code directly from your phone.
This bridges the gap between desktop-bound coding sessions and mobile accessibility. Being able to check on or steer a Claude Code session from your phone changes the workflow entirely.
Claude's Cowork feature now supports Projects — keep tasks and context in one place, with files and instructions stored locally. Import existing projects or start fresh.
Local-first project context is the right move. This makes Cowork feel less like a chat window and more like a workspace. The import flow lowers the barrier significantly.
A categorized collection of design inspiration sites — web design, landing pages, SaaS websites, navbars, CTA sections, and animation references.
A solid reference list to bookmark. Having inspiration sorted by component type (navbar, CTA, animation) is more useful than generic 'dribbble browsing' when you need targeted direction.
Weekly roundup of the fastest growing GitHub projects — agency-agents (a complete AI agency) gained 23.2K stars, and obra/superpowers (agentic skills framework) hit 100K stars.
The fact that the most-starred projects are agent orchestration frameworks says a lot about where the industry is headed. Worth watching both repos.
A look at physical games designed to pull people away from screens — a reminder that the best interfaces sometimes have no pixels at all.
Nice counterpoint to the constant stream of digital tools. Sometimes the most creative inspiration comes from stepping away from the screen entirely.
A tool that connects macOS native apps to AI through the Model Context Protocol — bridging desktop applications with AI capabilities.
MCP as the universal glue between native apps and AI models is a pattern that keeps gaining traction. This kind of bridge tool makes MCP practical for non-developer workflows.
A link share from Ole Lehmann — content creator known for AI workflow and content strategy insights.
The most starred AI repo of the month — an open source AI company you install in one command. 147 agents, 12 divisions, 50K GitHub stars in 14 days.
The framing as an org chart rather than a framework is doing heavy lifting here. People understand companies; they don't understand agent graphs. Smart positioning.
A Claude Code skill that replicates Emil Kowalski's design engineering style, producing high-quality UI in one shot.
A proof point that design taste can be encoded as a skill file — the output looks like Emil did it himself. This is the 'taste as code' idea made real.
A Babel plugin that turns Tailwind-like classes into Motion animations. Zero imports, zero runtime overhead.
Animation has always been the gap in utility-first CSS — this closes it by making motion as composable as spacing or color.
A walkthrough of combining Paperclip for orchestration, gstack for engineering skills, and autoresearch for automated R&D into one AI-powered development stack.
A concrete recipe for assembling an AI-first dev team — Paperclip as the manager, gstack as the hands, autoresearch as the lab. The real value is seeing the pieces fit together.
A SKILL.md file that teaches AI editors layout, typography, spacing, and motion — with adjustable settings for variance, motion, and density.
A technique for going from 'AI generates generic UI' to 'AI generates opinionated, high-quality UI' — taste is teachable via structured context files.
You can embed !`command` syntax in SKILL.md files to dynamically inject shell command output when a Claude Code skill triggers.
A technique for making Claude Code skills context-aware at runtime — instead of static instructions, your skill can pull in live data (git status, env vars, API responses) right when it fires.
Boris Cherny's workflow of running 10-15 parallel Claude Code sessions daily and updating CLAUDE.md after every correction to prevent repeated mistakes.
The key insight: after every correction, update CLAUDE.md so the mistake never happens again. Turns your config file into a living rulebook that compounds.
A purpose-built Claude Code skill for email marketing workflows.
Skills expanding beyond code into marketing territory — email marketing is a structured enough domain that a skill file can meaningfully guide the output.
15 specialized skills and 6 power tools that transform Claude Code into a virtual engineering team — from office hours to QA to shipping.
The full-stack AI development workflow made concrete. Not a single skill but an entire engineering process — plan, build, review, test, ship — all as composable slash commands.
A design system built on top of shadcn/ui where every component has been tweaked to reflect a specific aesthetic sensibility.
The real value here is the idea that design systems should have a point of view — shadcn gives you the primitives, but taste comes from opinionated defaults.
The /plan-eng-review skill lets you request diagrams — user flows, data flows, branching — and the act of creating them surfaces edge cases and bugs.
The insight isn't about AI generating diagrams. It's that forcing a system to explain itself visually is a debugging technique — diagramming is thinking.
Click UI elements, add notes, and generate structured context that AI coding agents can understand — visual annotation for agent-powered fixes.
Bridging the visual gap between what you see and what the agent understands. Click, annotate, share — the context goes directly to Claude Code or Cursor in a structured format.
A skill that automatically improves other skills — run, score, identify failures, adjust prompts, keep what works. Went from 56% to 92% quality pass rate.
Meta-skill engineering: a skill that optimizes other skills through automated testing loops. The jump from 56% to 92% quality is the headline, but the approach — score, adjust, retain — is the real technique.
A tool that runs autonomous improvement loops — 100 experiments while you sleep, compounding results over time.
The idea of an autonomous research loop that compounds its own improvements overnight. Set it running, wake up to results. Karpathy's approach to automated experimentation.
AI-powered marketing platform that creates blog posts, social media content, ad copy, and marketing briefs — all in your brand voice.
A full-stack AI marketing platform — strategy, content creation, scheduling, and analytics in one tool. Positioning as an agency replacement at a fraction of the cost.
Three prompt templates for AI coding: spec simplification, thorough planning with verification steps, and strategic debugging over trial-and-error.
From the Notion co-founder — 'How can we make this SIMPLER and DUMBER while still achieving our goals?' A forcing function against overengineering, applied at the prompt level.
A curated list of 18 design inspiration sites including Awwwards, Behance, Dribbble, SiteInspire, Mobbin, and more.
A solid bookmark-worthy reference for design research. Covers the full spectrum from awards sites to niche collections like Dark Mode Design and Minimal Gallery.
An open-source checklist to help you plan, build, and grow your design system.
A practical planning artifact for anyone starting or scaling a design system. Not another framework — just a structured checklist to make sure you haven't missed anything.
Google Stitch announces their biggest update ever — positioning as a 'vibe design partner' for UI generation.
The official announcement from Google Stitch — they're leaning hard into the 'vibe design' framing. This is Google's play in the AI-assisted design space.
Google Stitch's biggest update — generating app screens, landing pages, and UI flows from a single prompt.
The 'vibe coding' equivalent for design. One prompt to full UI flows. Eight examples of wildly different outputs from the same tool — the range is what makes it interesting.
Open-source framework for building internal coding agents, with case studies from Stripe, Ramp, and Coinbase.
LangChain's answer to the internal coding agent question — open-source framework with real production case studies from Stripe (Minions), Ramp (Inspect), and Coinbase (Cloudbot).
Official insights from the Claude Code team on skill design patterns — what works, what doesn't, and how to structure skills for reliability.
First-party guidance on skills from the people who built the system. The most shared tweet in the batch (42k bookmarks) — clearly fills a gap in understanding how to write good skills.
AI creative agent platform that orchestrates specialized models across video, image, and audio — plan, generate, iterate, and refine with full context.
The force multiplier framing: not replacing creatives, but making them prolific. Orchestrates multiple AI models for coordinated creative output across video, image, and audio.
Maestro.dev — a developer tool platform shared in reply to a coding discussion.
A tool recommendation surfaced through developer conversation — worth checking out for the developer tooling ecosystem.
Open-source marketing skills suite for Claude Code — 33 skills, 197 automated evals, Composio integration for HubSpot/Salesforce/Meta Ads, and 61 CLI tools.
The most comprehensive skill suite I've seen — 33 skills with 197 evals and CI testing. The Composio integration is the standout: one MCP server connecting to 500+ marketing tools.
Open-source meeting assistant that runs 100% locally — LLM suggestions, embeddings, and transcription all on your Mac via Ollama, nothing hits the network.
The local-first approach applied to meeting notes. No data leaves your machine — transcription, knowledge base, and AI suggestions all run on Ollama. Privacy by architecture.
A design-forward shadcn component kit for interfaces that stand out — from Amsterdam.
A shadcn variant that prioritizes visual polish over utility defaults. For when you want the shadcn DX but need the output to actually look distinctive.
A project management platform for managing multiple concurrent projects with customizable AI agents, real-time activity tracking, and scratchpad tools.
Project management reimagined for the AI era — parallel projects with customizable agents that have names, emoji, and profile pictures. The 'parallel timeline' framing is interesting.
A multi-step system for creating Claude Code skills — interviews, classification, drafting, quality reviews, test generation, and iteration.
A meta-skill: a systematic process for building skills themselves. Interview, classify, draft, review, test, iterate. Treating skill creation as an engineering discipline.
A critique of the 'future of coding is specification' narrative — arguing that detailed specs inevitably devolve into pseudocode, defeating the purpose.
A contrarian take worth sitting with. If your spec has to be precise enough for an agent to execute, you're just writing code with extra steps. The real question is where the useful abstraction boundary lives.
A tool that detects and removes common AI writing patterns from text, making AI-assisted prose read like it was written by a human.
The problem isn't AI-generated text — it's that AI-generated text all sounds the same. This tool attacks the homogeneity, not the generation.
A project management tool designed specifically for agencies and client-focused work.
Agency-specific project management — built for the workflow of managing multiple client projects simultaneously rather than generic task tracking.
119 rules across 11 categories — animations, timings, UX laws, typography, audio — distilled from userinterface.wiki into a single skill file.
Codified design intuition. Instead of hoping your AI knows good UI, you give it a structured rulebook. The rules aren't prescriptive — they're the kind of knowledge that separates craft from output.
Adding context: fork to a skill runs it in an isolated subagent with a fresh context window, keeping the main context clean.
A technique for keeping complex skill executions from polluting your main context — the subagent pattern means skills can do heavy lifting without bloating the conversation.
Google's Stitch SDK lets you program design directly — generating UI screens, landing pages, and app flows from code.
Design-as-code is finally real. When you can program design, the barrier between prototype and production dissolves.
A comprehensive guide to building web apps that work offline and sync when connected.
This is where things are heading. The sync engine is the new database.
Six professional Framer templates made free — Subflow, Syntrix, Cutflow, Viralup, Linear Studio, and Daniel Cross.
Free Framer templates from a marketplace creator. Good starting points for landing pages and portfolio sites without building from scratch.
Beautiful, composable chart components built on Recharts and Radix.
Finally a chart library that doesn't look like it was designed by committee.
A built-in Claude Code command that analyzes your codebase and suggests optimizations — took one minute and saved hours.
A hidden gem in Claude Code's built-in commands. One-minute analysis that surfaces optimization opportunities you'd otherwise spend hours finding manually.
A framework for categorizing AI workflows into four types: Skills, Plugins/Tools, Agents, and documentation (CLAUDE.md).
A useful mental model for organizing your AI automation — analyze your sessions, identify patterns, then sort them into the right abstraction level: skill, plugin, agent, or doc.
Making the case for SQLite as your primary production database.
Pair this with Litestream for replication and you've got something real.
Curated font pairings from Google Fonts with single-command installation — automates font setup in layout and CSS files.
The shadcn pattern applied to typography — curated font pairings you install with one command. Eliminates the manual font config dance in layout and CSS files.
Gives coding agents the context they need to work with your components — Radix, Base UI, updated APIs, component patterns, and registry workflows.
The bridge between 'AI knows how to code' and 'AI knows how to code in your system.' Component-aware agents make fewer mistakes because they understand the abstractions they're building on.
A GitHub repo covering Claude Code commands, sub-agents, skills, MCP servers, memory, hooks, and agent teams — 'the playbook.'
A comprehensive reference repo for Claude Code patterns — commands, sub-agents, skills, MCP, memory, hooks, and teams all in one place. The community's unofficial manual.
Building an AI-powered operational system using Claude Code for scheduling, email triage, and task management — the digital assistant dream realized.
The shift from 'AI writes code' to 'AI runs my operations.' When the agent handles scheduling and email, you're not just saving time — you're changing what a solo operator can manage.
A practical guide to evaluating agent skills with code, prompts, and real results — because most skills are AI-generated and untested.
The unglamorous but critical work: how do you know if a skill actually works? Code, prompts, and real results for evaluating quality. The testing gap in the skills ecosystem.
Google Drive, Gmail, Calendar, and every Workspace API accessible from the command line — with 40+ agent skills included.
Google making their entire productivity suite agent-accessible via CLI. The signal isn't the tool — it's that Google is designing for agents as first-class users alongside humans.
Complete AI coding workflow system with 18 hook events, 5 agents, 7 reference guides, and support for Claude Code, Cursor, and 32+ agents.
An ambitious orchestration layer for AI coding — 18 hook events and cross-agent support. The scale is the interesting part: designed to coordinate 32+ agents.
Select any element on your page, tell Claude Code or Codex what to change. Bridges the gap between visual selection and code editing.
The missing link in AI-assisted frontend dev — instead of describing which element you mean, you just point at it. Makes the feedback loop as tight as a design tool.
Anthropic dropped a massive update to the skill-creator skill for Claude Code and OpenClaw.
The official skill-creator from Anthropic getting a major upgrade — the meta-tool for building Claude Code skills. 2.4k likes and 4k bookmarks signal this hit a nerve.
A built-in Claude Code subagent called 'claude-code-guide' that answers questions about CLI, Agent SDK, and Claude API using read-only tools.
A hidden subagent inside Claude Code that can answer questions about itself — uses Glob, Grep, Read, WebFetch, and WebSearch but can't edit. A self-documenting tool.
Impeccable v1.1 — design fluency as agent skills, now supporting Antigravity, VS Code, and universal install.
Design quality as a portable skill that works across AI harnesses. The rename from 'simplify' to 'distill' is telling — it's about refinement, not reduction.
Open-source orchestration layer for autonomous businesses — org charts, goal alignment, task ownership, budgets, and agent templates.
The idea that a company's operating system can be codified into agent workflows. Not just task automation — full organizational structure as infrastructure.
Restructuring a 7.8k-star skill using progressive disclosure — reducing initial load from 1,625 lines to 183 while maintaining functionality.
The key insight: treat instruction files like tables of contents, not comprehensive guides. Separate rules from reference material. Less context upfront = better agent performance.
Instead of asking for 10 ideas, 'snowball' one topic through sub-angles, contrarian perspectives, and practical guides to generate 30 days of content.
A prompting technique for content: start with one seed topic and expand it through multiple lenses rather than asking for a flat list. Depth over breadth in ideation.
A Claude Code skill from Vercel Labs that handles OAuth scopes and webhook verification for building Slack agents.
Vercel abstracting away the worst parts of Slack integration — OAuth and webhooks — into a single skill. 'Anyone can build a Slack agent, even if you've never touched OAuth scopes.'
Knowing the correct terms for UI components dramatically improves AI-generated frontend results — a vocabulary reference for vibe coding.
The bottleneck isn't the AI, it's your vocabulary. Knowing the right term — 'combobox' not 'dropdown with search' — gets you 100x better results. Language as a design tool.
A vision of an agency with 8 departments managed by 38 AI agents and 8 people — agents don't own results, people do.
The key insight isn't the agent count — it's 'agents don't own results, people do.' One accountable human per department. The org chart of the near future.
A guide to transforming Claude Code into a full autonomous coding team with production-ready agents, persistent memory, custom hooks, and slash commands.
Another take on the Claude Code best practices repo — framing it as building an autonomous team rather than using a single tool. The mental model shift matters.
Frameworks for training AI agents in product management — a GitHub repo of PM-specific skill definitions.
Skills expanding into product management territory. Not just code — PM workflows like user story writing, prioritization, and stakeholder communication as structured agent skills.
Exploring a multimodal dashboard approach where details expand inline instead of navigating to new pages — keeping context while diving deeper.
A design pattern worth stealing: instead of breaking flow with page navigation, expand details in-place. Keeps the dashboard context while allowing deep exploration.
Deep dive into Astro's content layer for type-safe content management.
Skill graphs — interconnected markdown files with wikilinks — capture complex domains better than a single SKILL.md file.
The idea that agent knowledge should be a graph, not a document. Single files flatten expertise; linked files preserve the relationships between concepts.
Remotion's agent skills let you create programmatic video animations entirely through Claude Code prompts.
Video as a programmable medium — when you can describe motion in words and get production-quality output, the content creation bottleneck shifts from execution to imagination.
A tip to point your CLAUDE.md to an AGENTS.md file so you maintain a single source of truth for agent instructions.
A simple organizational trick: instead of maintaining two files, have CLAUDE.md reference AGENTS.md. One file to update, one source of truth.
Matt Pocock's CLAUDE.md additions that transform plan mode from unreadably long plans to concise, useful plans with followup questions.
From TypeScript's most visible educator — concrete CLAUDE.md rules that fix plan mode's verbosity problem. Before: walls of text. After: concise plans with follow-up questions.
Reusable CSS layout primitives that actually compose.
The stack and cluster patterns changed how I think about spacing.