Skills-hub Blog
Deep dives, deep work.
The newest AI coding features, the skills people actually install, and the patterns that survive a production codebase. Updated daily.
The State of AI Coding in 2026: Tools, Models, Patterns, and What Comes Next
A full snapshot of AI coding in 2026, every major tool (Claude Code, Cursor, Antigravity 2.0, Copilot, Windsurf, Codex, Cline, Kiro, Gemini CLI, Aider), the models behind them, the patterns that won (subagents, scheduled tasks, MCP, skills), and where the field is heading.
All posts
October 2026
OpenAI Codex CLI
OpenAI Codex CLI Updates: Instant Interrupts, Guardian Reviews, and Plugin Caching
The latest OpenAI Codex CLI releases introduce instant interrupt steering, opt-in Guardian reviews for agent handoffs, and performance improvements through plugin manifest caching. Here is what matters for developers writing and using AI coding skills.
4 minClaude Code
Name a Skill verify and Claude Code Will Run It Before Committing
Claude Code v2.1.286 (September 30, 2026) made one skill name load-bearing: if your project or user skills include one called verify, Claude is now told to run it right before committing. Two other folder names silently refuse to load, and the new /doctor prompt-audit reads your skills for instructions written for older models.
7 min
July 2026
GitHub Copilot
GitHub Copilot Browser Tools Are GA: AI Agents Now Verify Their Own Front-End Work
As of July 1, 2026, GitHub Copilot's browser tools hit general availability in VS Code — letting AI coding agents drive a real browser to verify their own front-end changes. Here's what changed, how it works, and how to apply the same pattern in Claude Code and Cursor.
7 minGitHub Copilot
Kimi K2.7 Code in GitHub Copilot: The First Open-Weight Model in a Mainstream AI Coding Assistant
Moonshot AI's Kimi K2.7 Code landed in GitHub Copilot on July 1 — the first open-weight model in the five-lab model picker. MIT-licensed weights on Hugging Face, hosted on Azure, billed at GPT-5.4 mini tier pricing. Here's what it means for your workflow.
7 minClaude Code
Claude Code's July Pivot: Background Agents Deliver PRs, Manual Mode Is the New Default
Six releases in eight days (v2.1.197–204) shipped Claude Code's biggest behavioral shift of 2026: background agents now commit, push, and open draft PRs autonomously when worktree tasks finish; 'default' permission mode was renamed Manual and enforced across every surface; and dynamic workflow sizing lets you tune agent-pool scale from a single config line.
7 minModel Releases
GPT-5.6 Sol, Terra, and Luna: The Developer's Tier-Selection Guide
OpenAI's GPT-5.6 ships as three distinct models — Sol (frontier agentic), Terra (balanced daily driver), and Luna (fastest/cheapest). Here's exactly when to use each tier in your AI coding stack, with pricing math and real workflow patterns.
7 minMCP
X Ships an Official MCP Server: What XMCP Means for Your AI Coding Workflow
On June 30, X launched XMCP — a hosted MCP server at api.x.com/mcp with 200+ tools auto-generated from the X API OpenAPI spec. Here is how to wire it into Claude Code, Cursor, or any MCP-compatible agent, what you can query, and why every major platform will follow.
6 minClaude Code
claude mcp login & /rewind: The Two Claude Code June 2026 Features Worth Your Attention
Claude Code v2.1.185–v2.1.193 (June 22–26, 2026) shipped two underrated features: OAuth-native MCP server authentication with `claude mcp login`, and /rewind for recovering context after /clear. Here's what they do and how to use them today.
8 minMCP
MCP 2026-07-28: Six Breaking Changes and the New MCP Apps Primitive
The largest MCP revision since launch ships July 28. Sessions are gone, six things will break in your server, and a brand-new primitive — MCP Apps — lets tools ship interactive HTML UIs to any host. Migration checklist and MCP Apps primer inside.
8 minClaude Code
Claude Sonnet 5: The Agentic Upgrade That Changes Your Daily Workflow
Claude Sonnet 5 shipped June 30 with a 13-point Terminal-bench jump, configurable reasoning effort, and introductory pricing that undercuts Opus for most automation. Here's what changes for Claude Code users today.
7 minClaude Code
Claude Enterprise-Managed Auth: Zero-Touch MCP Connectors via Okta
Anthropic shipped enterprise-managed authorization for MCP connectors on June 18, 2026. Admins provision Okta once; employees get Asana, Atlassian, Figma, Linear, and four more connectors on first login across Claude Chat, Claude Code, and Cowork — no per-user OAuth.
7 minDevin Desktop
Cascade is Dead: Migrate to Devin Local Today Before Your CI Breaks
Cascade reaches end-of-life on July 1, 2026. In-editor workflows migrate automatically — but CI scripts and automations that call Cascade by name will break silently. Here's the exact grep, the exact fix, and what Devin Local adds over what you're replacing.
8 min
June 2026
Model Release
Gemini 2.5 Pro Deep Think: What the 2M-Token Context Means for AI Coding
Google released Gemini 2.5 Pro with Deep Think on June 22, 2026 — 87.6% on LiveCodeBench, 2M token context, and visible reasoning chains. Here's what actually changes for developers who build with AI.
8 minMulti-Agent
ACP Spec 1.0: Build Your AI Agent Once, Run It in Every IDE
The Agent Client Protocol just hit its first formal spec release under open governance. ACP is JSON-RPC 2.0 over stdin/stdout — the LSP for AI agents — backed by Zed, JetBrains, GitHub, Google, and Cognition. Here is what changed, which tools are compatible, and how to wire Claude Code, Codex, and Devin Local into a single unified command center.
7 minCompliance
Fable 5 Was Pulled in 72 Hours: What the Export-Control Suspension Means for Your AI Stack
On June 12, US export-control regulations forced Anthropic to suspend Claude Fable 5 globally — just 72 hours after launch. Here is the migration playbook, the four codebase dependencies you need to audit, and the model-agnostic architecture that makes the next suspension a non-event.
8 minAI Agents
Arcade's $60M Bet on Agent Authorization: Why Production AI Agents Need a Permission Layer
Enterprise AI agents keep stalling after pilot. Arcade just raised $60M to fix the real bottleneck: not the model, but authorization. What the $72M company builds, why a former Okta leader is behind it, and what every developer needs in a production agent stack.
7 minClaude Code
Claude Code Artifacts: Turn Any Coding Session Into a Live, Shareable Page
On June 18, Anthropic launched Artifacts in Claude Code — session work published as live, auto-updating web pages shareable with authenticated teammates. PR walkthroughs, incident timelines, system explainers, dashboards, and release checklists, all from your terminal.
7 minKiro
Kiro IDE 1.0: Agent Focus Mode, Custom Agents, and the Permission System That Audits Every Tool Call
Kiro shipped 1.0 today with six major features. Agent Focus Mode brings a chat-first layout for parallel sessions; custom agents are Markdown files with four tool-access tags; Natural Language Hooks translate plain English into structured JSON; and the new permission system evaluates every file read, shell command, and MCP call against your declared rules.
8 minClaude Code
Claude Design + /design-sync: The Two-Way Bridge Between Design and Code
Anthropic's June 17 Claude Design overhaul lands design system imports, WYSIWYG canvas editing, and a bidirectional /design-sync command. Claude Code and Claude Design now share components without screenshots or rebuilds — one million users in the first week.
7 minCursor
Cursor Teams Standard vs Premium Seats: The June 2026 Pricing Breakdown
Cursor split Teams into Standard ($32/mo annual) and Premium ($96/mo annual) seats this month, with two separate usage pools and new admin spend controls. Here's how to right-size your team before your July billing cycle.
7 minOpenAI Codex
Codex Record & Replay: Show Your Workflow Once, Get a Reusable Skill
Codex Record & Replay shipped June 18, 2026 — the first AI coding tool feature that turns a demonstrated workflow into a portable, editable SKILL.md. Here's how recording works, what the generated skill looks like, how to replay it with new inputs, and how to export it to Claude Code, Cursor, or any tool on the registry.
8 minDesign-to-Code
Figma Dev Mode MCP: Let Your AI Coding Agent Read the Design Directly
Figma's official Dev Mode MCP server gives Claude Code, Cursor, and Windsurf structured access to components, variables, and Code Connect mappings — replacing screenshot guesswork with machine-readable design context. Here's the complete setup and workflow guide.
8 minCursor
Cursor 3 Cloud Agents: /in-cloud, Bugbot 3×, and the PR Babysitting Pipeline
Cursor 3.7 (June 17, 2026) ships isolated cloud subagents via /in-cloud, reusable environment snapshots that spin up in under 10 minutes, and /babysit for overnight PR preparation. Bugbot hit 3× faster reviews the week before. Here's how to wire it all together.
8 minOpen Source AI
Cohere North Mini Code: The Open-Source 30B Agent That Fits on One H100
Cohere's North Mini Code landed June 9 as the first Apache-licensed agentic coding model that genuinely outperforms much larger proprietary alternatives — 33.4 on the AI Coding Index, 256K context, single-H100 FP8 deployment. Here's the full developer playbook.
8 minOpen Source
OpenCode: The Model-Agnostic Terminal Agent That Just Hit 160K Stars
OpenCode connects to 75+ AI providers from a single CLI, reads code through Language Server Protocol, and costs $2–8 per day instead of $20–40/month. Here's why 7.5M developers made the switch — and how to set it up.
8 minClaude Code
Claude Agent SDK Billing Split: What the June 15 Change Breaks and How to Fix It
On June 15 2026, Anthropic split Claude subscription billing into two buckets: one for interactive use, one for autonomous agents. If you run claude -p, GitHub Actions automations, or any SDK-driven pipeline, your credits now drain from a separate pool — and stop when they run out.
8 minClaude Code
Claude Code Week 24: fallbackModel, Safe Mode, and /cd for Production Resilience
Claude Code's June 2026 Week 24 update shipped four features that matter most to teams running agents at scale: fallbackModel chains, --safe-mode troubleshooting, /cd workspace switching, and nested sub-agents five levels deep.
8 minClaude Code
Fable 5's 1M-Token Context: A Practical Guide to Whole-Codebase Analysis
Fable 5 ships with a 1-million-token context window — enough to load your entire codebase without summarization tricks. Here's what actually works in practice: sizing repos, caching strategy, recall tradeoffs, and a four-phase migration playbook.
8 minKiro
Kiro CLI 2.1: Headless Mode Brings AI Code Review to Every CI/CD Pipeline
Kiro CLI 2.1 ships headless mode and skills as slash commands. Here's how to wire Kiro into GitHub Actions for automated code review, dependency audits, and PR summaries — no browser, no interactivity, no waiting.
8 minClaude Code
Claude Code Ultracode: The xhigh + Orchestration Mode That Rewires Your Sessions
Ultracode is not just a higher effort level — it pairs xhigh reasoning with automatic parallel workflow orchestration so Claude plans and runs multi-agent pipelines on your behalf. Here's the complete guide: what it does, when to use it, and how to stay inside your token budget.
8 minClaude Code
Claude Fable 5 in Claude Code: The Mythos-Class Upgrade for Real Coding Work
Fable 5 just dropped as Claude Code's new default option — 80.3% on SWE-Bench Pro, the first Mythos-class model cleared for general use, and free on subscriptions through June 22. Here's what changes for your day-to-day coding workflows.
8 minClaude Code
Claude Opus 4.8: 88.6% SWE-bench, Cheaper Fast Mode, and What Developers Need to Know
Opus 4.8 lands at 88.6% SWE-bench Verified and 69.2% on the harder Pro set — nearly 5 points ahead of 4.7. Here's what the benchmark gap means in practice, how the 3× cheaper fast mode changes cost math, and when to route tasks to Opus vs Sonnet.
8 minAI Models
GPT-5.5 for Agentic Coding: What the Terminal-Bench 2.0 Leader Means for Your Workflow
GPT-5.5 leads Terminal-Bench 2.0 at 82.7% and cuts hallucinations by 52.5% versus GPT-4.5. Here's what changes for developers running agentic coding pipelines — API migration, SKILL.md compatibility, MCP tool-call format, and when to switch.
8 minClaude Code
Claude Code Hooks: Complete Guide to All 30+ Lifecycle Events
Claude Code hooks fire deterministically at 30+ lifecycle points — before and after every tool call, on session start and end, when files change on disk, during context compaction. Here's the definitive map with real configs for safety enforcement, auto-formatting, and session context loading.
8 minOpen Source
Hermes Desktop: The MIT-Licensed Agent That Remembers Everything Between Sessions
Nous Research shipped a native cross-platform desktop app for Hermes Agent on June 5 — 100+ PRs in a week. Here's what persistent goals, durable kanban, and checkpoint rollbacks mean for developers who run long-horizon work across multiple sessions.
8 minGitHub Copilot
GitHub Copilot AI Credits: First-Week Reality Check (June 2026)
On June 1, GitHub Copilot dropped flat-rate billing and switched to token-based AI Credits. One week in: some developers report bills jumping from $29 to $750/month. Here's the math, who gets hit hardest, and how to stay in budget.
8 minMCP
MCP 2026: The Stateless Protocol Upgrade Every Server Author Needs to Know
The next Model Context Protocol release (finalising July 28, 2026) is the biggest spec change since launch: stateless core, JSON Schema 2020-12, hardened OAuth, cloud-native routing headers, and a formal deprecation policy. Here's what changes and how to migrate.
7 minCursor
Cursor 3.6 & 3.7: Auto-Review Mode, Canvas Design, and Smarter Agent Control
Cursor 3.6 ships Auto-review Mode — a three-stage classifier (allowlist → sandbox → LLM judge) that decides which tool calls run, run sandboxed, or need approval. Cursor 3.7 adds Design Mode for visual canvas iteration and interactive context usage reports. Here's how both reshape your daily workflow.
7 minDevin Desktop
Windsurf Is Now Devin Desktop: ACP, Devin Local, and What the Rebrand Actually Changes
On June 2, 2026, Cognition retired the Windsurf brand and shipped Devin Desktop — a Rust-rewritten local agent (30% more token efficient), native Agent Client Protocol (ACP) support, and a unified IDE that runs Claude Code, Codex, and any ACP-compatible agent in the same Kanban command center.
8 minCodex
OpenAI Codex and GPT-5.5 on Amazon Bedrock: The Developer Playbook
OpenAI's GPT-5.5, GPT-5.4, and Codex coding agent went GA on Amazon Bedrock on June 1, 2026. Here's what changed, how to wire it up, and what per-token pricing means for teams who were paying per seat.
7 minKiro
Kiro Requirements Analysis: How AWS Uses an SMT Solver to Kill AI Slop Before It Starts
Kiro's May 2026 Requirements Analysis engine applies a three-stage neurosymbolic pipeline — LLM rewrite, formal logic translation, SMT solver proof — to catch logical contradictions in your spec before a single line of code is generated.
7 min
May 2026
AI Coding Tools
Kiro 0.12: Parallel Tasks, Quick Plan Mode, and Spec-to-Code in a Single Pass
Kiro 0.12 (May 2026) ships parallel task execution, Quick Plan Mode, and pre-implementation requirements analysis. Here's what changed, when to use each mode, and how these features stack up against the tool's original approval-gate workflow.
8 minClaude Code
Claude Code Dynamic Workflows: Orchestrate Hundreds of Background Agents
Three releases across 48 hours (v2.1.152–154, May 27–29) introduced dynamic workflows that orchestrate tens to hundreds of background agents, a MessageDisplay hook that reshapes the terminal UI, and hot-swappable skills. Here's the complete breakdown.
8 minIndustry
Gemini CLI Is Now Enterprise-Only: What Every Developer Needs to Know
Google accepted 6,000 community pull requests, accumulated 100K GitHub stars, then announced on May 19 that Gemini CLI's free tier ends June 18. Here's what happened, what Antigravity CLI actually offers, and which alternatives are worth switching to now.
7 minAI Coding Tools
Amazon Kiro: The Spec-Driven IDE That Takes Prototypes to Production
AWS launched Kiro to close the gap between vibe coding and shipped software. It converts natural language into EARS-notation specs, maintains code from those specs, and fires agent hooks on save. Here's the full workflow, MCP integration, and how it compares to Cursor, Windsurf, and Claude Code.
8 minGoogle Antigravity
Google Antigravity 2.0 SDK: Build Custom Agents on Your Own Infrastructure
Google launched Antigravity 2.0 at I/O on May 19, 2026 with a public SDK, an agy CLI, and a Managed Agents API. Here's everything a developer needs to ship a custom agent against Gemini 3.5 Flash today.
8 minCursor
Cursor Composer 2.5: Build in Parallel, 10x Cheaper, and What Changes for Your Workflow
Cursor's third-generation agentic coding model launched May 18, 2026. Composer 2.5 matches Claude Opus 4.7 on SWE-bench at one-tenth the cost, introduces a 100-sub-agent parallel architecture, and brings targeted RL training that makes it dramatically more reliable on long-running tasks.
8 minWindsurf
Windsurf 2.0: Agent Command Center, Spaces & What the Devin Integration Actually Changes
Windsurf 2.0 (April 15, 2026) ships three features that reframe how developers manage multi-agent work: a Kanban-style Agent Command Center, Spaces for context-bundled task management, and native Devin cloud handoff. Here's what changed and how to reconfigure your workflow.
8 minHow-to
The AI Coding Stack 2026, What to Install on a Fresh Developer Machine
Opinionated, complete setup for a 2026 fresh-machine AI coding stack. One tool per role, editor, terminal agent, autocomplete, MCP servers, skill catalog, daily routine, plus the 10 skills to install on day one.
11 minRoundup
The Top 50 AI Coding Skills for 2026 (Ranked, with Install Commands)
The 50 most-installed AI coding skills on skills-hub.ai right now, ranked and grouped by SDLC phase. One install command per skill. Works with Claude Code, Cursor, Codex CLI, Windsurf, Copilot, Cline, Antigravity, Kiro.
14 minClaude Code
Claude Code Subagents: The Complete 2026 Guide to Agent Teams
Subagents are Claude Code's headline 2026 feature, isolated parallel agents with their own context windows, tool access, and skill loadouts. Here's how to compose them into agent teams that implement, test, review, and ship in parallel.
9 minClaude Code
Claude Code /loop: Cron for AI Agents (March 2026 Update)
/loop is Claude Code's scheduled-task feature, a cron syntax for AI agents that runs skills headlessly, opens PRs, runs tests, and reports back. Nightly security audits, pre-standup pipelines, dependency drift, all without keeping a terminal open.
6 minClaude Code
Talking to Your Terminal: Claude Code /voice Mode (2026)
/voice landed in Claude Code in March 2026, push-to-talk in 20 languages, transcribed to the chat as you go. Here's how it changes the actual rhythm of pair programming with an AI, where it shines, and where it falls down.
5 minGitHub Copilot
GitHub Copilot Flex Billing: The Complete June 2026 Developer Playbook
On June 1, 2026, GitHub Copilot switches from premium request units to AI Credits with flex allotments. Here's exactly what changes, what stays unlimited, how to read your preview bill, and which plan to pick, before the deadline.
8 minStateful Skills: Extending SKILL.md with Memory
Every skill marketplace today ships prompt templates. None of them ship what practitioners actually learn from running the skill 60 times. Stateful skills add three sibling files, MEMORY.md, EXAMPLES.md, CALIBRATION.md, that turn a static artifact into a compounding one. Backward compatible. Open RFC.
April 2026
Roundup
Top 10 AI coding skills every developer needs in 2026
A curated, opinionated list of ten Claude skills that pay back the install cost in the first hour, code review, TDD, security hardening, OWASP, deploys, ADRs, and more. Pulled from real install counts on skills-hub.ai.
9 minRoundup
The 25 best Claude skills (2026), vetted, security-scanned, ready to install
The best Claude skills of 2026, ranked by install count and quality. Code review, testing, deployment, security, and documentation skills for Claude Code, Cursor, and any MCP-compatible tool. Install with one command.
6 minHow-to
Install Claude skills in 30 seconds (npx, no config)
Install any Claude skill in under 30 seconds using a single npx command, no global install, no config file, and the same command works for Claude Code, Cursor, Codex CLI, and any MCP-compatible tool via --target.
5 minComparison
Claude Code skills vs Cursor rules: when to use which (2026)
A side-by-side comparison of Claude Code skills (SKILL.md) and Cursor rules (.cursorrules / .mdc), file format, install paths, scoping, and which one fits each workflow. Both, actually. Here's how to ship one source of truth that targets both.
7 minHow-to
Building your first AI coding skill (SKILL.md tutorial)
A concrete, end-to-end tutorial: write a SKILL.md file, validate it locally with the skills-hub CLI, install it into your project, and publish it to skills-hub.ai with /publish so anyone on Claude Code, Cursor, or Codex can install it.
8 min