If there is one pattern across every question I received this month, it is this: more people are asking how to build an AI system they can actually trust.
Trust comes from reliability. Reliability comes from a proper system built inside the agent.
That’s why I wrote two posts about it this week:
The first question came from Atit, and it was the clearest version of it.
He wanted McKinsey-style research and decks, at a real quality bar, and he was trying to work out which tool would get him there. But the tool was never going to be the answer. Without a defined source, a report standard, reference decks, and an output format the agent can see, every deck starts from nothing and lands somewhere different.
That is where the babysitting begins. You give the agent a task. It gives you something that looks mostly right. Then you notice the weak research, the generic email, the missing task, or the visual that has nothing to do with the post. You explain the problem, ask it to try again, and keep watching because you do not fully trust what happens next.
Six AI Maker Lab members sent questions this month, and all six were a version of the same frustration:
Atit wanted reliable McKinsey-style research and presentation decks, plus a clearer way to decide between local models and hosted ones.
Mei wanted an agent that could follow a real workflow with less hand-holding.
Karen wanted to avoid getting trapped inside a single AI tool and asked how she could work her project across Claude Code, Codex, and Hermes.
Thomas stuck between Cursor, Claude Code, and Things.
Julia was overwhelmed with so many saved posts she had not touched and consumed at all
Lee spent two hours to create his newsletter banner.
All of them is different jobs, but the real problem underneath them them is identical.
So, in this post, I tackled all of them.
In this video, I also showed the working versions behind the folder structure for a research agent, my Obsidian–Todoist planning loop, the AI News Intel Skill that read 27 newsletter emails, the X bookmark workflow that has been collecting everything I’ve consumed over the past week, how to import your Claude setup into ChatGPT/Codex, the Glif process that generates thumbnails and infographics for my newsletter, and how to create a design system for generating banners with Claude Design.
Let’s dive in.
Atit: How do I build reliable McKinsey-style research and decks?
Atit wanted to create deep research and presentation decks with a specific quality bar. He was considering Claude Code, Codex, NotebookLM, and Gemini. He also asked how I would think about the build-versus-buy decision for an AI agent.
For research work, I keep coming back to a file-based agent like Claude Code or Codex, because I want control over the process and the final format.
NotebookLM is useful when the main job is working through a large pile of sources. But, where it stops working for me is customization. You cannot really shape the workflow, the agent, or the output beyond what NotebookLM already gives you. For a job where I need to define the research sequence, enforce my own report standard, and produce several exact deliverables, I can’t have that ceiling in my workflow.
So, the tool you choose matters here, but having more control over how you work and what you can generate is more important.
If you want to build a research agent, here is the rough structure I’d build in my project folder:
