How I'm Actually Using AI Right Now (Including Claude Duet)

The host frames Claude Duet’s launch as a reason to explain his broader AI stack rather than review Duet in isolation. His main work, spanning the AI Daily Brief and Superintelligent, is deciding what to create next: content strategy for the podcast and business strategy for Superintelligent, including identifying AI users’ learning gaps. Because he thinks best by externalizing ideas, he keeps extensive “walk notes” in Notion—thoughts developed while walking, driving, or sitting and planning. Across all his AI conversations, he estimates that roughly 60% by time spent involves strategic thinking, even though coding consumes more tokens. He says every major new model is tested against this strategy use case to see whether it can replace his current tool one-to-one. The most important quality is not generic intelligence but the ability to push back without becoming unhelpfully contrarian or sycophantic. He wants a model that is creative and adaptable but also confident enough to distinguish a genuinely good idea from one that merely sounds good. He currently prefers Opus 5.5, after using Fable 5.1 and, before that, OpenAI models that he found relatively flat for this work.

Research is another major use case because the results become public on the AI Daily Brief or the host’s wife’s true-crime podcast, Love Murder. He needs models that cite sources accurately, avoid hallucinations, understand source quality, and apply the right emphasis to different kinds of evidence. He uses Astra and Opus about equally, and often has them review one another’s work: Astra checks Opus’s research, Opus checks Astra’s, and the combined results provide an additional confidence layer before publication. For copywriting, he does not use AI to script the AI Daily Brief. Each evening he and his researcher, Scott, survey the day’s AI news and discourse; around 8 or 9 p.m. Eastern, they decide the next show’s coverage, and Scott prepares an overnight dossier. The host uses AI after recording for episode descriptions, social posts, and titles. He always brings an initial title, but lets Astra argue for alternatives; about half the time he keeps his original, while the other half produces a different title through back-and-forth. He attributes Astra’s advantage partly to long-established context about what he likes and dislikes, making a switch difficult even if another model might be equally capable.

Image generation is part of the host’s daily workflow, especially for episode cover art and visuals shown during the program. He mostly uses GPT image models because it is convenient to create the cover in the same thread where he finalized the headline, but he still subscribes to Ideogram and Midjourney. Ideogram is useful for iteration because it produces four images at a time and offers controls over variation. Midjourney remains valuable when he does not know precisely what he wants and is looking for something visually arresting, unusual, or artistically distinctive. For data analysis—including organizing financial information such as tax expenses and analyzing podcast performance—he generally uses Astra inside Codex. He says Astra gives him the best results for this work and has the “condition of thought” he finds useful. His point is not that one model wins universally, but that each model’s value depends on where it fits in his workflow: Astra is strong for structured analysis and copywriting, while other models are better suited to strategy, research, or construction.

For software and workflow construction, the host has tried OpenClaw, Codex, Claude Code in both its app and command-line forms, and third-party tools such as Cursor. His stable default is Claude Code with Opus 5.5, usually using high or extra effort. He sometimes builds in Codex with Astra to get native access to image-generation models, particularly for visually oriented show companions, but Claude Code handles most of the organizing, narrative architecture, and orchestration. He can also give Claude Code access to GPT image models through an API, though he sees a difference between using those models natively inside Codex and accessing them through the API. He uses Claude Code to maintain the AI Daily Brief website and says Fable 5.1 changed his expectations about completion. Earlier tools made it easy to start projects but often left them at 80% or 90%; Fable reduced the “completion energy” required to finish. On the first afternoon he had access, he built an initial version of a new website and production pipeline before Fable was taken offline during an export-control dispute with the White House. The pipeline accepts a transcript, extracts relevant information, and turns it into readable, shareable web cards using text and color rather than image generation. A second pipeline turns those cards into tweets and a LinkedIn post scheduled through Typefully.

The host later extended the website pipeline into video. Claude reviews each episode transcript, selects what it judges to be the strongest self-contained 60-to-90-second segment, extracts the clip, adds visual assets matching the website’s aesthetic, and packages it for distribution. He says recent Claude models have become so capable at video editing that even people without an immediate video need may benefit from experimenting with them. He still has human video editors, who make a first pass and then integrate AI for animation and direct connection to the clip pipeline. These systems became the foundation for How We Got Here, a weekly 45-minute-to-one-hour show about the social, cultural, technological, and economic decisions that produced major current phenomena. The first episodes cover America’s loss of faith in the technology industry, why many things became cheap while necessities became expensive, and the rise of prediction markets. The host uses Claude for strategic and distribution decisions, Claude and Astra for research dossiers, and Claude to turn his arguments into key beats. After recording, Opus 5.5 receives his raw transcript and the original research and drafts a 5,000-to-6,000-word companion essay. He then puts that essay into Codex with Astra and GPT image generation to create its visual design, though he says he rewrites roughly half of the first three essays during an exhaustive edit.

Before Duet, the host had not made proactive AI a major part of his workflow. He does not use agents to monitor his inbox, watch Slack, send routine updates, or draft recurring responses, because he had judged the setup and maintenance effort greater than the benefit. OpenClaw was his main experiment: individual Claude agents managed specific tasks through heartbeats and scheduled jobs, while a chief-of-staff agent coordinated them. He also ran an always-on research agent tracking AI studies, surveys, and adoption for the AI Daily Brief and Superintelligent. He eventually stopped the research project and found that the chief-of-staff agent did not add enough value to justify upkeep. He stresses that this result may be specific to his habits. He tightly controls his calendar, with about 90% of his calls unscheduled, and handles email by starring important messages and clearing the inbox every few days. He therefore does not dismiss proactive agents generally, and says Hermes may be an excellent product even though it does not fit his work. He also expects to use personal agents professionally rather than for flight booking, restaurant reservations, shopping, or similar personal tasks. For example, he usually flies with an airline where he has miles, so explaining all the relevant contingencies to an agent would take as long as using that airline’s site.

The host sees two possible purposes for personal agents: genuinely new proactive work, such as monitoring email and drafting responses, or better coordination of existing AI work. The second possibility resembles his preferred “monothread” model: instead of many chats and projects with fragmented context, one ongoing thread carries the relevant history, retrieves what it needs, and launches subagents when necessary. Improvements in context-window management and compaction have made this more practical, and he says it fits his stream-of-consciousness working style better than manually maintaining a hierarchy of chats and folders. His experience with ChatGPT Dots is limited. He uses Quasar—the name of his Dot and also a character from Dungeon Crawler Carl—as a monothread space for important ChatGPT work such as research and podcast-data analysis, while simpler queries and copywriting remain elsewhere. But Dots has not helped with strategy: after he shared a six-hour Claude and Opus 5.5 strategy conversation, Quasar reportedly responded, “Great, this information has been saved for the future,” rather than engaging with any of its ideas. He describes its personality as flat, its proactivity as weak, and its main potential value as holding context. Duet is more explicitly work-focused and immediately asks to connect Slack, Gmail, and Drive. It proposed proactive projects from the start, though the host thinks it sometimes overvalues a task merely because it can act on it. It repeatedly warned him that two sponsors appeared to have been offered the same advertising date, without understanding that multiple slots existed; he had to tell it that the entire workstream was irrelevant. On the other hand, Duet has identified several emails he would normally star and revisit later, drafted responses he could edit or authorize, and once produced a substantive response that got him about 90% of the way there. He also likes its web interface, separate channels, interactive Claude artifacts, and task categories such as active, up next, scheduled, and resolved.

After only a few days with Duet, the host is more optimistic about its approach than about Dots, especially for proactive work. He does not expect to do everything in Claude through Duet: a UI and brand-visual-design task was much worse there than in Claude’s native design experience, which he expects to be incorporated into Artifacts later. For him, Duet’s value will depend on whether it can cheaply perform the small amount of proactive work he currently leaves undone, rather than replace his preferred tools for strategy, design, coding, or research. He also discusses the broader ecosystem. Third-party tools such as Instinct and Town are lower priority because he does not want to move his professional context away from ChatGPT and Claude. Grokbot is useful because of its integration with X, and he stays engaged enough not to start from zero if it becomes compelling, but he would rather not maintain another major ecosystem. He is tracking Google and Microsoft, while noting that Google currently does not seem worth additional investment for him and Microsoft’s enterprise orientation does not match his solo workflow. He concludes that users probably do not need to make a permanent ecosystem bet: features successful in one product will be copied by the others within days or weeks. His current view is that personal agents are less a new primary interface to AI than “an easy-to-set-up little baby version of myself” that handles unwanted tasks. That is valuable, but he is not yet convinced it represents the future of the industry.