The fenbs blogAll articles
Page 16 of 48
- Playwright MCP: Letting an AI Agent Drive the Browser for TestsPlaywright MCP is Microsoft’s MCP server that lets a coding agent open pages, click, type and read what happened, using the page’s accessibility tree. How to install it in each client, the options that matter, how it compares with the Playwright CLI and test agents, and how to keep a browser full of your sessions safe. · 7 min read
- Context7 MCP: Current Library Docs for Your Coding AgentContext7 feeds a coding agent up-to-date, version-specific documentation for the libraries it is using, so it stops writing code against APIs that have changed. Who runs it, how to add it to each client, what the API key changes, how it picks a library, what it costs in context, and how far to trust what it returns. · 6 min read
- n8n AI Agents With MCP: Connecting Tools to a Workflown8n’s AI Agent node puts a model inside an ordinary workflow, and its MCP nodes connect that agent to any MCP server, or turn n8n into one. How the nodes fit together, one workflow worked through, where the human approval goes, how credentials are kept, and when n8n is a better fit than code. · 8 min read
- Claude With Microsoft 365: Outlook, Teams and SharePointClaude’s Microsoft 365 connector lets it search your Outlook mail and calendar, Teams chats and meetings, and SharePoint and OneDrive files, and act in them when an administrator allows it. What it reads, what it can change, the Entra consent it needs, its limits, and how to use it safely. · 7 min read
- Remote vs Local MCP Servers: How to Connect and Which Is SaferA local MCP server is a program your client starts on your machine; a remote one is a service at a URL you sign in to. How to connect each kind in the main clients, where your data and credentials go, what each one can harm, and a short guide to choosing. · 9 min read
- Why MCP Servers Eat Your Context Window, and What to DoEvery MCP server you connect adds tool definitions to the model’s context, and every result it returns stays there. Where the tokens go, how to see them in each client, and the fixes that work: tool search, trimming, toolsets, leaner servers and code execution. · 8 min read
- MCP Apps: Interactive UI Inside Claude and ChatGPTMCP Apps is the official MCP extension that lets a server show an interactive page inside the conversation: a chart, a form, a board. What it is, its status, which hosts render it, how the sandbox keeps it contained, and the smallest working shape. · 8 min read
- AI Agent Frameworks Compared: When You Need One and When You Don’tLangGraph, CrewAI, the OpenAI Agents SDK, the Claude Agent SDK, Google ADK and Microsoft Agent Framework all run the same loop for you. What each one is built around, which languages it speaks, how it reaches MCP servers, and how to tell when forty lines of your own code will do. · 8 min read
- Microsoft Agent Framework: What It Is, and Semantic KernelMicrosoft Agent Framework is the successor to both Semantic Kernel and AutoGen. What it is made of, its release status in .NET, Python and Go, a small C# agent and workflow that compile against the current packages, and what changes when you migrate. · 8 min read
- Agent Skills Across Claude, Copilot, Codex and CursorA Claude skill is a folder in the open Agent Skills format, and Codex, GitHub Copilot, Cursor and Gemini CLI read the same format. What differs is where each tool looks and which frontmatter it honors. Here is how to keep one copy of a skill that every assistant on your team can load. · 6 min read
- What Is an Agent Harness? Why the Loop Around the Model MattersThe harness is everything in an agent that is not the model: the loop, the tools, what goes into the context, the permissions, the memory and the sandbox. Why the same model does better in one harness than another, what Claude Code, Codex, VS Code and the SDKs mean by the word, and what it changes when you choose tools. · 7 min read
- MCP Sampling Explained: When a Server Asks the ModelSampling lets an MCP server borrow the client’s model for a completion. How it worked, how it works under the 2026-07-28 specification, which clients support it, and why the specification now tells new servers not to use it. · 7 min read
- Building an MCP Server in C# and .NETA working MCP server in C# with the official SDK: two tools with dependency injection, served over stdio and over HTTP with ASP.NET Core, tested in the MCP Inspector and connected to Claude Code and VS Code. Plus the dotnet new template, and what the SDK does with your errors. · 8 min read
- Writing MCP Tool Descriptions a Model Can UseA model picks and calls your MCP tools from their names, descriptions, schemas and error messages alone. How to write each one so it picks the right tool, fills the arguments correctly and recovers when a call fails, with before-and-after rewrites. · 8 min read
- Spec-Driven Development vs Vibe Coding vs TDDThree ways to build with an AI coding agent, tried on the same small project: where each one breaks, which to reach for when, and how they fit together rather than compete. · 8 min read
- GitHub Spec Kit: Specs, Plans and Tasks for Coding AgentsSpec Kit is GitHub’s open-source toolkit for spec-driven development with an AI coding agent. How to install it, what each command does and which files it writes, a walk-through on one feature, its limits, and where the tasks should live once they exist. · 8 min read
- Spec-Driven Development With Claude CodeClaude Code has no feature called specs, but its own tools cover every step: an interview that ends in a spec file, CLAUDE.md pointing at it, plan mode, a fresh session to build, a subagent to review against the spec, and a board for the tasks. A worked example, and where Spec Kit fits. · 8 min read
- Writing a PRD an AI Coding Agent Can Build FromA product requirements document for an AI coding agent has one job: let it build the right thing without guessing. The seven sections it needs, a template to copy, a filled-in example, how to turn it into tasks, and how a PRD differs from a spec. · 7 min read
- User Story vs Epic vs Task vs Bug: How the Pieces FitA story is a change a user would notice, a task is a piece of the work that delivers it, an epic is a story too big to deliver in one go, and a bug is something that does not work as intended. Where each sits, how each is written, and a simpler scheme if the hierarchy is more than you need. · 8 min read
- Bug Severity vs Priority: How to Set BothSeverity is how bad a bug’s effect is. Priority is how soon the team will fix it. They are set by different people from different evidence, and the interesting bugs are the ones where they disagree. · 7 min read
- Feature Request Template: A Form Users Will Actually Fill InA feature request is most useful when it describes the problem, not the solution. The fields that matter, a template to copy, a four-question form for people outside the team, good and bad examples, and what to do with a request once it arrives. · 6 min read
- Product Backlog Examples and a TemplateThree real-looking product backlogs, for a SaaS app, a mobile app and an internal tool, each ordered top to bottom with the reason for the order. Then a template to copy, one fully written backlog item, and the habits that keep a backlog short. · 8 min read
- Backlog Refinement: How to Run It, and Why It Stopped Being GroomingBacklog refinement is the ongoing work of splitting, clarifying, sizing and ordering what is near the top of the backlog. What the Scrum Guide says, where the old name went, and a one-hour agenda with the inputs, outputs and people it needs. · 8 min read
- Sprint Retrospective Questions and Formats That Get Honest AnswersA retrospective is only as good as the questions in it and the actions that come out of it. Six formats, a question bank grouped by what you are trying to learn, how to make honesty safe, and how to run a retro on work an AI assistant did. · 8 min read
- Kanban Board Examples for Eight Kinds of TeamEight example kanban boards, from a software team to a building site to a board shared with AI agents. For each: the columns, a handful of sample cards, and the one rule that makes it work. · 8 min read
- Personal Kanban: Using a Board for Your Own WorkPersonal kanban is a board for one person, run on two rules: see all your work, and limit how much of it you start. Where the idea comes from, a setup that takes ten minutes, a weekly review, and an honest word on ADHD. · 7 min read
- How to Write a Changelog: Format, Examples and Keep a ChangelogA changelog is a curated list of notable changes, newest first, grouped by kind. The Keep a Changelog format, how Semantic Versioning fits, what generated changelogs get right and wrong, and a good and a bad example side by side. · 7 min read
- Claude Cowork vs Claude Code: Which One for Which JobBoth come from Anthropic, share the same agent machinery and are on the same paid plans. Cowork is for knowledge work that ends in a document, a deck or a sorted folder; Claude Code is for work that ends in a commit. Who each suits, where each runs, what each can touch, and five jobs sorted between them. · 7 min read
- Claude Team Plan: What You Get and How to Set It UpThe Team plan is Anthropic’s plan for a group of people working in one Claude organization: shared projects, Claude Code and Cowork on every seat, admin controls, single sign-on, and no training on your content by default. What it includes, how it differs from Pro and Enterprise, and how to set it up in an afternoon. · 7 min read
- Claude Desktop App vs Web: What the Desktop App AddsThe Claude desktop app for macOS and Windows is claude.ai with a way into your computer: local MCP servers and desktop extensions, the local half of Cowork, quick entry on a Mac, and a bridge for Claude in Chrome. What it adds, how to install it on each system, what stays the same, and when the browser is all you need. · 7 min read
- Claude Prompting Best Practices, With Before-and-After ExamplesAnthropic’s own prompting guidance comes down to a handful of habits: say exactly what you want, give the reason, show an example, separate the parts, put long material first, and say what to do rather than what to avoid. Here they are, applied to seven ordinary work prompts, with what has changed for the newest models. · 8 min read
- Claude Skills Examples: Twelve Skills Worth CopyingEight skills Anthropic publishes and four of our own, each with the lines of its SKILL.md that do the work and the request that makes it trigger. Then what the good ones have in common, and where to find more without installing something you have not read. · 8 min read
- Claude Artifacts: Twelve Examples Worth BuildingTwelve artifacts you can ask Claude for today, from a status dashboard to a quiz that calls Claude itself, each with a one-line prompt and what you get back. Then what artifacts can and cannot do as documented, and how to iterate on one without starting again. · 7 min read
- Claude Artifacts vs ChatGPT CanvasChatGPT’s canvas has given way to writing blocks, code blocks and Sites, while Claude’s artifacts have grown templates, storage and sharing. A comparison of editing, running code, sharing, collaboration and plans, as each vendor documents it today, and which fits which job. · 7 min read
- Vibe Coding Security: The Mistakes That Leak DataSeven security issues turn up again and again in apps built by prompting: secrets in the browser, open API routes, tables without row-level security, loose CORS, trusted client input, invented packages and chatty errors. Each one, where it comes from, and the fix. · 8 min read
- AI Code Review Tools Compared, and a Checklist for Using ThemGitHub Copilot code review, Claude Code, CodeRabbit, Cursor Bugbot, Graphite and the open-source PR-Agent, compared by where each runs, what it comments on, how you configure it and whether you can read its source. Then a checklist for the person who still has to approve. · 8 min read
- AI Coding Agent Benchmarks: What SWE-bench and Terminal-Bench MeasureHow SWE-bench and its variants, Terminal-Bench, Aider’s polyglot benchmark and LiveCodeBench are built and scored, how to read a leaderboard without being misled by the scaffold or contamination, and how to run a small eval on your own repository. · 8 min read
- Database MCP Servers: Letting an Agent Query Postgres or MySQL SafelyWhich database MCP servers exist today, from Supabase, Neon, Google and Microsoft, how they are set up, and the controls that make them safe: a read-only database login, a replica, row and time limits, no production writes, and treating what comes back as untrusted. · 8 min read
- Securing Claude for Teams and Enterprise: Admin Settings That MatterThe admin controls that decide how safe a Claude organization is: sign-in and provisioning, roles, connectors, skills and plugins, retention, audit logs and the Compliance API, Claude Code policy and network limits. Which plan has each one, what to set, and a checklist for the first week. · 8 min read
- Claude Sonnet vs Opus: Which Model to Use for Which JobOpus is Anthropic’s default recommendation and Claude Code’s default model; Sonnet is faster and costs less per token; Haiku is the fastest and cheapest; Fable sits above them all. What each is for, how effort changes the picture, and how to switch in Claude Code. · 7 min read
- Claude Code Statusline: Show Model, Cost, Branch and ContextThe status line is one command in your settings that turns session data into a line under the prompt. What Claude Code sends it, when it refreshes, and four tested scripts for bash, plus PowerShell and Node versions for Windows. · 7 min read
- Running Claude Code on Amazon BedrockClaude Code can send every request to Claude models in your own AWS account instead of Anthropic’s API. Why teams do it, the setup as Anthropic documents it, the IAM policy, what does not work on Bedrock, and what to check when it fails. · 8 min read
- Prompt Engineering vs RAG vs Fine-Tuning: Which Fixes Your ProblemThree ways to make a model do better, and each one fixes a different failure. Start from the symptom you can see, and the choice between prompting, retrieval and training mostly makes itself. · 8 min read
- Product Roadmap Best Practices for Small TeamsA good roadmap says what you are trying to achieve, in what order, and how sure you are, without promising dates you cannot keep. Here are the habits that make one work for a small team, a template to copy, and the mistakes to avoid. · 7 min read
- Can Claude Code Build a Website? Where It Needs YouYes: Claude Code can write, run and test a website, a web app or a mobile app. What it can do on its own, what still needs a person (accounts, domains, store signing, payments, taste and review), and a worked small site from first prompt to deploy. · 7 min read
- Claude Code Deleted My Files: Settings That Prevent ItWhen Claude Code deletes files it is almost always through a shell command, which checkpoints do not cover. The settings that stop it (deny and ask rules, the sandbox, worktrees), what /rewind can and cannot bring back, and a recovery path for when it has already happened. · 8 min read
- Claude vs ChatGPT for Coding: The Chat Apps ComparedCan ChatGPT code like Claude? In the chat apps, mostly yes. How the two differ at running code, previewing it, reading your repository and choosing a model, and the point at which a chat stops being the right tool and a coding agent is. · 7 min read
- Windsurf vs Claude Code, Now That Windsurf Is Devin DesktopWindsurf was renamed Devin Desktop in June 2026, and its Cascade agent was removed in September in favor of Devin Local. How the IDE and its agent compare with Claude Code on models, context, rules files, permissions and MCP, and who each suits. · 7 min read