The fenbs blogAll articles
Page 20 of 48
- Claude Projects Best PracticesScope each project to one body of work, give it only the knowledge every chat needs, write instructions a stranger could follow, and write conclusions back so the next chat starts where the last one ended. · 7 min read
- Multi-Agent Workflows: Coordinating Several AI Agents on One BoardThree patterns for putting several AI agents on one job, the three ways it goes wrong, and the handful of rules that keep them from doing the same work twice: one task per piece of work, one owner per task, and the result written on the task. · 7 min read
- Claude Code Subagents vs Agent TeamsSubagents are workers inside one session that report back to it. Agent teams are separate Claude Code sessions that share a task list and talk to each other. What each one is today, how to turn teams on, and which to reach for. · 7 min read
- Giving Claude Code a Memory That Survives the SessionClaude Code forgets everything when a session ends, unless something was written down. What CLAUDE.md, auto memory and a resumed conversation each carry forward, who can see each one, and where shared work should live instead. · 7 min read
- Agent Memory vs RAG: What Each Stores, Who Writes It, When It Is ReadRAG fetches passages from documents that existed before the agent arrived. Agent memory holds what was learned while working. They fail in different ways, need different owners, and most long-running agents need both. · 7 min read
- Agent Memory Over MCP: Options and Trade-offsServing memory from an MCP server lets any assistant read and write the same notes. Three ways to do it, and what each means for who can read the memory, who can change it, whether you can see who did, and where private data ends up. · 7 min read
- Copilot Agent vs Ask vs Plan: Which Mode for Which Task?VS Code’s Copilot chat has three built-in roles: Ask answers, Plan proposes, Agent changes things. Pick by how big the task is and how much it would cost to get wrong, not by habit. · 7 min read
- Vibe Coding for Product Managers Without Losing the PlanBuilding a working prototype by describing it to an AI is one of the most useful things a product manager can now do. It goes wrong when the prototype quietly becomes the spec, the decisions stay in the chat, and the follow-ups never reach the team. · 7 min read
- Vibe Coding With a Backlog: Structure Without Slowing DownVibe coding is building software by describing what you want to an AI and letting it write the code. Here is where the term came from, what it means now, and how to do it without ending up with a project nobody understands. · 7 min read
- Agentic Workflows Explained With Everyday ExamplesAn agentic workflow is a job where an AI model does some of the steps and decides some of the path. Here is how a workflow differs from an agent, three ordinary jobs worked through step by step, and where a person should say yes. · 7 min read
- Claude Code Hooks: Update a Board Automatically When Work FinishesWhat Claude Code hooks are, how they answer back, and two small hooks that make sure a commit never ends a session without a line on the task board. · 7 min read
- Claude Projects vs ChatGPT ProjectsBoth keep files, instructions and chats together around one piece of work. They differ in who can share, what collaborators see, how memory is walled off, and which tools come along. A side-by-side, feature by feature. · 7 min read
- Claude Code vs GitHub Copilot Agent Mode: How They Handle TasksTwo agents that take a coding task and work through it with tools. Where each runs, how it reads your standing instructions, how it reaches other tools, how you approve and undo its work, and how each runs several tasks at once. · 7 min read
- Security Controls for AI Coding AgentsA coding agent runs commands with your shell, your files and your network. Seven controls, each enforced by software rather than by a polite instruction, keep what it can reach small and what it did visible. · 8 min read
- Running a Security Review of an AI Agent’s AccessOnce a quarter, list every credential an AI agent holds in every system, find out who issued it, when it was last used and when it ends, and remove what nobody can justify. A procedure, a register and a checklist. · 7 min read
- When an AI Agent Makes a Mess: An Incident ChecklistAn agent deleted the wrong thing, pushed what it should not have, or read something it was never meant to see. Contain, assess, restore, communicate and learn, in that order, mapped to NIST’s current incident response guidance. · 8 min read
- Approval Workflows for AI Agent ActionsAn approval is only useful if it pins down what was approved, by whom, and for how long. How to design an AI agent approval workflow: which actions need one, whether to approve before or after, who signs, and how to keep the record. · 7 min read
- Shadow AI Agents: Finding the Assistants Your Team Already ConnectedYour team has almost certainly connected AI assistants to work tools already, one sign-in at a time. Where to look for them, in the identity provider, the tools, the repositories and the key lists, and what to do next without making anyone regret telling you. · 7 min read
- MCP Governance: Who in a Team May Connect WhatA lightweight MCP policy a team can run without a platform group: an approved servers list, who may add to it, the scopes each connection gets, the client settings that enforce it, and how often it is reviewed. · 8 min read
- How to Debug MCP Tools When an Assistant Gets It WrongWhen an assistant calls the wrong tool, passes the wrong arguments or shrugs at an error, the fault is in one of four places. How to find which, using the client’s own status and logs, a replay in the MCP Inspector, and a careful read of what the model was actually shown. · 8 min read
- Jira MCP With GitHub Copilot in VS CodeConnect GitHub Copilot’s agent mode in VS Code to Jira through the Atlassian Rovo MCP Server: the gallery install or a .vscode/mcp.json entry, the browser sign-in, tool approvals, token access, and what admins on both sides control. · 7 min read
- How to Use Linear MCP in CursorConnect Cursor’s agent to Linear through Linear’s hosted MCP server: the one-click install or an mcp.json entry, the browser sign-in, a read-only connection, which tools to allowlist, and how it differs from Cursor’s own Linear integration. · 6 min read
- Connecting Trello to ChatGPT With MCPTrello’s official MCP server is listed in ChatGPT, so connecting takes a few clicks in Settings. The steps, what to tick on the consent screen, how ChatGPT asks before changing cards, the developer-mode alternative, and the limits worth knowing. · 6 min read
- Jira MCP vs the Jira API: Which Should Your AI Use?Atlassian’s MCP server and the Jira Cloud REST API reach the same data by different doors. How each one signs in, what it exposes, how it is limited, and when a plain script beats an assistant. · 8 min read
- Microsoft To Do and MCP: Can Claude Manage Your List?Microsoft ships several MCP servers, but none of them is a To Do server. What Microsoft does offer, what Claude’s Microsoft 365 connector covers, the Graph API that community servers wrap, the risks of those servers, and when a shared board is the better fit. · 7 min read
- ChatGPT Connectors for Jira and Other TrackersJira reaches ChatGPT through Atlassian’s own app, Atlassian Rovo. What it does, how to connect it, who controls it on each side, and which other trackers have official listings in ChatGPT today. · 6 min read
- How to Automate Claude Code Tasks SafelyHeadless runs, machine-readable output, tool lists decided in advance, hard limits on turns and spend, and a GitHub workflow that ends in a pull request a person reviews. The pieces, and the guardrails that make them safe to leave alone. · 7 min read
- Claude Code Subagents: Examples Worth CopyingFour subagent files you can drop into .claude/agents/ today: a reviewer that cannot edit, a test runner that only reports, a docs checker, and a board recorder that can read and comment and nothing else. Each with the reasoning behind its tool list. · 6 min read
- Claude Code in VS Code: Tasks, Terminals and TrackingThe Claude Code extension gives you a chat panel with diffs, @-mentions and checkpoints; the CLI in the integrated terminal gives you everything else. How to use both, how to turn repeat jobs into VS Code tasks, and how to keep track of work across sessions. · 6 min read
- Codex CLI: Setup and Keeping Its Work on a BoardInstall Codex CLI, sign in, give it an AGENTS.md and a board over MCP, and set how much it may do without asking. Then every session ends as a list of tasks you can check, not a scrollback you cannot. · 6 min read
- Generating Release Notes With AI From Finished TasksPull requests say what changed in the code. Finished tasks say what changed for the people using it. Here is how to have an AI assistant draft release notes from the second, and what a person still has to do. · 6 min read
- Using MCP Without Writing CodeConnecting Claude or ChatGPT to another service over MCP needs a web address and a sign-in, not a terminal. What you click, what the sign-in screen is asking, and how to stay safe while you learn. · 6 min read
- Claude Skills for Project ManagementFive skills worth writing if you run projects with Claude: the weekly status, triage, meeting notes to tasks, release notes and the stale-work nudge. With a complete SKILL.md you can copy and adapt. · 7 min read
- Linear MCP for Product Management WorkflowsFour jobs a product manager can hand to an assistant through Linear’s MCP server: triaging new issues, writing requests up as issues, summarizing a cycle, and drafting project and initiative updates. What each one needs in access, and where a person has to look before anything changes. · 7 min read
- GitHub Copilot Agent Mode With a Task TrackerA five-step loop for driving Copilot agent mode in VS Code from the work items in Jira or any other tracker: pull the task in, scope it, let the agent work, review, and write the result back where the team will see it. · 7 min read
- Handing Work Between Coding Agents and People Mid-TaskPausing Claude Code is one key. Handing its half-finished work to a colleague, or a colleague’s half-finished work to an agent, takes a note written for whoever comes next. Here is how to do both directions. · 7 min read
- MCP for Beginners: Connect Your First Tool in 10 MinutesA first, hands-on Model Context Protocol tutorial: add one remote MCP server to the assistant you already use, sign in, try three prompts that read, write and check, and see exactly what the assistant did. · 8 min read
- Context Engineering for Claude Code: The Levers That Decide What It SeesClaude Code gives you about eight separate controls over what is in its context window: CLAUDE.md files and their scopes, imports, rules, memory, compaction, subagents, skills and hooks. Here is what each one loads, when, and what survives a long session. · 8 min read
- Context Engineering With GitHub Copilot: Every Lever, and When to Use ItGitHub Copilot decides what the model sees from instructions files, prompt files, custom agents, skills, #-references, MCP servers and the workspace index. Here is what each one does, where it lives, and which to reach for. · 7 min read
- Agent Memory vs Context: What Each HoldsContext is what the model can see in this one call, and it ends with the call. Memory is what is kept somewhere else and brought back when it is needed. How the two meet, three everyday examples, and a way to decide what belongs in each. · 7 min read
- Claude Projects for Teams: What Works and What Does NotA shared Claude project gives a team one set of knowledge and instructions that every chat starts from. It does not tell anyone what is being worked on, by whom, or who changed what. What to use it for, where it stops, and how teams pair it with a board. · 7 min read
- Agentic Workflows in GitHub: Issues, Agents and ReviewIn GitHub, an agentic workflow is a loop: an issue describes the work, an agent picks it up, a pull request comes back, and a person reviews it before anything merges. The pieces GitHub gives you for each step, including the feature it actually calls Agentic Workflows. · 8 min read
- Where Claude Code Task Output Goes, and How to Keep ItClaude Code’s output ends up in six different places: the terminal, the session transcript, background task files, subagent transcripts, headless output and whatever it was told to write. Most of it is temporary. Where each one lives, how long it lasts, and how to keep the part that matters. · 7 min read
- Which Tasks Can AI Agents Automate, and Which Need a Person?Four questions sort any task on your list: can it be undone, can something other than a person check it, does it need judgment or authority, and does it touch money, people or production. The answers put it in one of three piles. · 7 min read
- Human Oversight of AI Under the EU AI Act: A Plain SummaryArticle 14 of the EU AI Act puts a human in the loop for high-risk AI systems: providers design for oversight, deployers assign competent people to do it. What that covers, what it does not, when it applies, and what a small team can sensibly do anyway. · 8 min read
- Human in the Loop for Generative AI ContentMarketing copy, help pages, support replies and code comments written by AI all need a person before they reach their readers. What to check for each, who signs off, how to record it, when to label it. · 6 min read
- Prompts for Project Management With AI AssistantsA copyable prompt library for the jobs a project lead repeats every week — planning, triage, status, risks, meetings, retros and stakeholder updates — written to run against a connected board and to read before they change anything. · 8 min read
- The Steps an AI Agent Takes to Complete a TaskRead the task, plan, choose a tool, act, look at what happened, check it against the goal, go round again, and report. What happens at each step, how each one fails, and where a person should step in. · 7 min read