claude plugin eval: test skills and plugins like you test code
The claude plugin eval command runs skills on real cases against a no-plugin baseline: six grader types, tips for stable cases and costs to know.
AI Agents & Skills
Skills, MCP, how agents work and keeping costs down. · 20 articles
All
Claude Code
Claude API
LLM & NLP
AI Agents & Skills
Tips
AI Automation
Tools
System Design
Reading
ReAct and Toolformer: the origins of every agent's think–act loop
The papers behind tool-using agents: ReAct interleaves reasoning and acting, Toolformer learns when to call APIs; plus inherited risks.
Lessons from Anthropic's multi-agent research system
The orchestrator–worker architecture behind Research, performance and token cost figures, prompting principles, evaluation and when not to use multiple agents.
Harnesses for long-running agents: keeping memory across context windows
How Anthropic uses an initializer agent, coding agent, JSON feature list, progress file, init.sh and Git to run multi-day agent projects.
MCP 2026-07-28 goes stateless: what you need to know
The major changes in the MCP 2026-07-28 spec: no handshake or sessions, server/discover, subscriptions/listen, MRTR, deprecations and how to upgrade.
Build an MCP server in Python in 30 lines
Write an MCP server with the official Python SDK, test it with MCP Inspector, connect Claude Desktop and Claude Code, and fix common errors.
Securing remote MCP servers: OAuth 2.1, CIMD and fatal mistakes
Putting an MCP server on the internet safely: Origin checks, the OAuth 2.1 flow with PKCE and resource, client registration via CIMD, no token passthrough.
Writing Agent Skills the Anthropic way: right description, under 500 lines, evals first
SKILL.md principles: be concise, pick degrees of freedom, name and describe for discovery, progressive disclosure, checklists, evaluation-first.
Context engineering: fewer tokens, stronger signal for your agents
Anthropic's engineering post summarized: context rot, right-altitude system prompts, just-in-time retrieval, compaction, notes and sub-agents.
Designing tools for AI agents: wrap the workflow, not the API
Anthropic's process and five principles for effective tools/MCP: consolidate by task, namespace, meaningful results, token efficiency, clear specs.
The 20 most-used Claude skill and Claude Code tip repos (Oct 2026)
20 top-starred GitHub repos for Claude: CLAUDE.md tips, workflow skills, token saving, UI design, writing, and skill security scanning.
OpenClaw: a self-hosted AI assistant you message on Telegram/WhatsApp that does real work on your machine
One of the fastest-growing open-source personal-agent projects (formerly Clawdbot, Moltbot). Powerful, but it needs a safe setup.
Workflow or agent? 5 patterns from Anthropic for picking the right architecture
Prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer — and when you actually need an autonomous agent.
What are Agent Skills? Understand SKILL.md before installing dozens of them
A skill is a folder with a SKILL.md that teaches an agent one specific job. It’s an open standard used by Claude Code, Codex, Gemini CLI, Cursor and more.
What is MCP? The “USB-C port” connecting AI to databases, GitHub, Slack…
The Model Context Protocol is a common standard for AI apps to reach external tools and data. Build a server once; every MCP-capable client can use it.
Superpowers: a skill set that makes Claude Code plan and test before coding
Instead of “build me an app”, Superpowers walks the agent through brainstorm → plan → batched execution → TDD → verification before saying done.
AutoHarness: let Claude Code turn each session’s lessons into skills
An open-source plugin that captures reusable approaches from your sessions as skills — then updates, merges and prunes them over time.
The Ralph Wiggum loop: run an agent over and over until the tests go green
A “dumb but effective” technique: keep the agent running on the same prompt and stop only when a check that can’t lie says it passed.
Hallmark: a design skill so AI-built sites stop looking AI-built
Purple gradients, identical rounded cards, generic copy — Hallmark uses page structures, themes and a 65-point “slop test” to block those tells.
“AI can clone any app without source code”? A reality check on REA
REA helps AI agents investigate how binaries, APKs and web apps behave. It’s real and useful — but the clickbait overstates it and skips the legal and safety risks.
No matching articles
Try another keyword or clear the topic filter.
Never miss a new article
Follow binhAI via RSS (Feedly, Inoreader, Slack…) to get new articles as soon as they’re published.