{"slug": "the-power-of-focus-in-agentic-coding", "title": "The Power of Focus in Agentic Coding", "summary": "Alex Graveley, the creator of GitHub Copilot, demonstrated a focused approach to agentic coding that prioritizes a fixed ten-step checklist over parallel sessions and excessive tooling, enabling a week-long task to be completed in hours. Graveley's method uses two Claude Code sessions, a short CLAUDE.md file built through annoyance-driven development, and tight scope discipline to avoid what he calls \"fake work\" that doesn't move metrics or unblock users. The approach helped ship a self-serve onboarding feature in a single day by enforcing explicit done-states and preventing architectural drift from over-scoped agents.", "body_md": "Alex Graveley, the guy behind GitHub Copilot, showed me how I’d been thinking about agentic coding wrong.\n\nHe used pen and paper.\n\nAlex dropped by our Mission Street office on a Tuesday. I told him we had about a week of work left to get self-serve working.\n\n“No. That should take 4 hours.”\n\nHe pulled out a sheet of paper, wrote down ten steps, and handed it to me. That was the plan.\n\nI’ve been through every phase: ChatGPT copy/paste, Cursor, Claude Code, Codex, back to Claude Code, subagents, orchestration frameworks, 10 parallel sessions. The assumption was that more tooling meant more output.\n\nThat assumption is half-right. The real constraint is undirected scope, not the model or the tools. Without a fixed list, parallel agents multiply your WIP, not your throughput. You end up with ten sessions, five half-finished features, and nothing shipped.\n\nParallelism only helps when every agent is pointed at the right thing.\n\nAgentic coding makes hard things look like low-hanging fruit. A feature will only take an hour, so you fire off a session. Then another. Suddenly you’re ten commits deep in a niche optimization nobody asked for. Alex calls this “fake work”—all the motion that doesn’t move a metric or unblock a user. In startups, the only thing that matters is getting users and removing blockers to that goal.\n\nAlex’s process is blunt: ten steps on paper, two Claude Code sessions, nothing else. Work through the list until done. The checklist is a contract with yourself about what ships today.\n\nHe starts every project with a fresh CLAUDE.md. He only adds lines when something annoys him. Claude keeps asking whether to commit? Update the file. He calls it annoyance-driven development: you don’t pre-load rules, you earn them through friction. Here’s where we ended up by the end of our session:\n\n- keep everything as concise as possible\n- no unnecessary abstractions\n- no unnecessary comments\n- when the next step is obvious, do it without asking\n- always be as general as possible\n- parallelize independent steps whenever possible\n\nOn prompting: I stuff agents with context, design docs, previous sessions. He keeps prompts short. For tasks a thousand teams have already solved (ECS, auth, DB migrations), more context means more noise. The model already knows the generic path; what it doesn’t know are your sharp constraints, which is where you spend your context budget.\n\nOn early optimizations: I default to thinking about reliability and scaling way ahead of when it matters. “Building for scale with a handful of users is avoidance.”\n\nThis changes scope discipline, not the quality bar. The ten-step list is a forcing function, not a shortcut around quality. When you ship a step, you verify it works. The discipline is in writing the list tightly enough that each step has a clear done-state: something you can check, not just declare.\n\nWhat it does kill is architectural drift from over-scoped agents. Short prompts, tight scope, explicit review at each step. That’s the governance. Simple, but it holds.\n\nBy the end of our session we were on step 6 of 10.\n\n“Don’t stop until you finish the list. Tomorrow, rinse and repeat.”\n\nBy night, anyone who requested access could onboard end-to-end.\n\nIn the age of agentic coding, focus is the superpower. Scope creep that *feels* like velocity is eating most of your AI-coding gains. More sessions, more output, less of what actually matters shipped.\n\nI came into that conversation thinking the bottleneck was tooling. More parallelism, richer context, smarter orchestration.\n\nI left with a crossed-off list on a sheet of paper. The milestone shipped.\n\nLook at what you’re working on right now. Can you write it down in clear steps on a sheet of paper? If not, you’re running on a vibe, not an explicit plan.", "url": "https://wpnews.pro/news/the-power-of-focus-in-agentic-coding", "canonical_source": "https://tenzinwangdhen.com/posts/copilot-creator-ai-coding-wrong/", "published_at": "2026-07-28 08:49:59+00:00", "updated_at": "2026-07-28 08:52:27.775282+00:00", "lang": "en", "topics": ["ai-tools", "developer-tools", "ai-agents", "generative-ai", "large-language-models"], "entities": ["Alex Graveley", "GitHub Copilot", "Claude Code", "Mission Street"], "alternates": {"html": "https://wpnews.pro/news/the-power-of-focus-in-agentic-coding", "markdown": "https://wpnews.pro/news/the-power-of-focus-in-agentic-coding.md", "text": "https://wpnews.pro/news/the-power-of-focus-in-agentic-coding.txt", "jsonld": "https://wpnews.pro/news/the-power-of-focus-in-agentic-coding.jsonld"}}