Most people use AI coding tools the slow way. They type a prompt, wait for the agent, review the output, then prompt again. Round and round, babysitting the whole time. But there’s a faster way, and I just watched a breakdown that lays it out perfectly. This video comes from Matthew Berman, an AI creator who tests every agentic coding tool out there as part of his job. He walks through exactly what the top-tier developers do differently, and honestly, a lot of it clicked for me right away.
The big shift? Experts stop being the bottleneck. Instead of sitting in the loop, they build systems that run without them. Let me break down what the creator shared.
The tools the expert actually reaches for
The author runs several coding agents daily, but two are his go-tos:
- Cursor: He loves it because you can swap models from different companies (OpenAI, Anthropic, even Cursor’s own). It was also one of the first to ship cloud agents.
- Codex: His pick for the clean design and short, concise summaries. He can’t stand reading essays about what an agent just did. One or two sentences, then move on.
He also gives nods to Claude Code, Devin, and Factory. His advice is simple: go try them and find what fits your style.
Rules and skills do the heavy lifting
Here’s the contrast that stood out. Beginners re-explain themselves constantly. The expert writes it down once.
That’s what agents.md (and claude.md for Claude Code) is for. It’s where you define your workflow, commit style, coding preferences, and even the personality you want from the model. Write it once, reuse it forever.
Then there are skills, which the original poster calls one of the most important things to master. His rule of thumb is blunt: anything you do more than once should have been a skill. Instead of copy-pasting the same prompt, you type a slash command and the skill runs.
The creator lists where skills shine:
- Repeated tasks you’d otherwise paste over and over
- Domain-specific rules, like your company’s writing style
- Tool instructions, so the agent already knows your APIs, CLIs, and how you run tests
- Quality gates, like “run all tests and hit 100% pass before opening a PR”
And the neat part: agents can discover which skill to use on their own at runtime. He also points to off-the-shelf libraries, like one popular skills repo with 61,000 GitHub stars covering the full dev cycle. Just grab the URL, tell your tool to install it, done.
Automations and loops, the real secret
This is the section the author hammers hardest. Two related ideas:
- Automations prompt your agent automatically when some trigger fires.
- Loops let an agent keep running until it hits a defined goal.
He shows a great automation example using Greptile (a code review tool that’s also the video’s sponsor). The setup: when a pull request opens, the agent waits for Greptile’s review comments, addresses each one, then pushes the fixed code back. You wake up to cleaner PRs without lifting a finger.
Loops are where it gets wild. The mind behind the video even announced a free Loop Library for the first time in this video. A loop needs three things: a trigger to start, an action it repeats, and a goal so it eventually stops. His real examples:
- Overnight docs sweep: Every night it reviews the codebase and updates any stale documentation, then opens a PR.
- Sub-50ms page load loop: It loads every page, modal, and sidebar, and keeps optimizing until everything loads under 50 milliseconds. He let it run for hours and the app ended up lightning fast.
- Production error sweep: Each night an agent scans production logs, finds errors, writes fixes, and submits PRs. He wakes up to fixes already waiting.
The flywheel: tests, docs, logs
The expert’s best-practice takeaway is a tidy one. There’s no excuse for sloppy code anymore because you can automate the boring parts:
- 100% test coverage, kept full by an automation that writes missing tests
- Always-fresh documentation, updated daily as the app changes
- Exhaustive logging, so agents can hunt and fix any error that appears
He calls this trio a flywheel, and I think that’s the right word for it.
Cloud vs local agents
The creator weighs both. ☁️ Cloud agents win on parallelism (no draining your laptop’s CPU running 20 agents), accessibility from mobile, and isolated environments so agents don’t clash on the same files. Cursor’s cloud agents even hand you a video and screenshots of the changes, no asking required.
Local agents win on speed (an environment is always ready), more hands-on control, and earlier access to new features. His own plan is to move most work to the cloud. He also recommends git worktrees so each agent gets its own copy of the repo and conflicts only get resolved at merge time.
Multi-model and the one unsolved problem
A smart trick the author uses: assign different models to different jobs inside a single skill. Plan with a heavyweight model, write code with a fast coding model, then review with yet another model for a fresh perspective. It saves tokens and finishes faster.
He’s honest about the catch, though. Merging and deploying a dozen parallel agents into production is still broken. They stumble over each other, lock the commit and deploy process, and restart constantly. His workaround is batching PRs and merging together. He notes Cursor just announced a Git alternative built for agent-scale deployment, so a fix may be coming.
My take
If you only grab one thing here, make it automations and loops. That’s the jump from babysitting AI to actually leading it. The full video has the live walkthroughs and the Loop Library link, so go watch it and start turning your repeat tasks into skills.