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

The host’s AI use is concentrated on strategy, idea development, research, and content creation rather than monitoring an inbox, email, or Slack. Strategy accounts for an estimated 60% of AI conversations and sessions, and the host wants a model that challenges ideas thoughtfully instead of flattering or reflexively disagreeing. Opus 5.5 is the current preference, following Fable 5.1 and earlier OpenAI models; GPT models felt too flat and insufficiently dynamic for much of this year. For research, the host uses Astra and Opus about equally, often having them check each other, and emphasizes careful source evaluation when preparing public statements because hallucinations remain a risk.

AI supports the AI Daily Brief’s production without writing its unscripted show: Astra helps create episode descriptions and social posts after recording, and it changes the host’s proposed title about half the time by making a case for an alternative. GPT image models produce most images, including the daily cover once the headline is settled; Ideogram makes four images per iteration, while ChatGPT produces one at a time. Midjourney is useful when the desired image is unclear and creative latitude is welcome. For data analysis, Astra inside Codex gives the host the best results and concise reasoning.

Claude Code with Opus 5.5, usually at high or extra effort, is the default for building. The host also uses Codex with Astra for some visual companion experiences to get native access to image-generation models, and gives Claude Code access to GPT image models through the API, though that differs from using the models inside Codex. Fable stood out for reducing the effort needed to finish projects, not merely to start them. Claude also maintains the AI Daily Brief website and its pipeline from transcripts to shareable web assets; text and color, rather than generated imagery, keep the branding consistent, clear, and readable. Another pipeline turns episode assets into social posts and schedules them through Typefully, though automation cannot replace engaging with people.

Claude Code also helps select and package self-contained transcript segments as 60–90-second video clips, adding visual assets for upload. The host says Claude models, especially Opus 5.5, have become very good at video editing and encourages experimenting even without an immediate video use. These pipelines predated How We Got Here, a planned weekly show lasting 45 minutes to an hour: Claude helps develop ideas and key beats, while the host sources and arranges visuals and talks over them. Opus 5.5 drafts a companion essay of typically 5,000–6,000 words from the raw transcript and original research; the host develops it in Codex with Atlas and GPT Image. Video editors make the first pass, with AI used for animations and to connect editing to the clip pipeline.

Duet is the host’s latest personal-agent experiment, and its form factor is still unsettled: it might act as a chief of staff, a proactive proxy, or something else. After only a few days, the host is more optimistic about its proactivity than about earlier agents. Duet is work-oriented, asks to connect Slack, Gmail, and Drive, and suggests projects. It has proposed replies to emails the host would usually star and revisit; the host has let it send some, including one substantive draft that was about 90% complete. Its web interface, separate channels, editable interactive artifacts, and task states appealed to the host. But it repeatedly warned about a possibly duplicate sponsor booking because it did not understand the host’s slot system, requiring the host to tell it several times that the issue did not matter. Duet also performed much worse than Claude’s native design experience on a UI and brand-design task.

Earlier OpenClaw experiments included heartbeat and cron-job agents, a coordinating chief-of-staff agent, and a persistent research agent. The host stopped the research agent after the project changed and dropped the chief-of-staff agent because its value did not justify the upkeep. Personal agents may help others, but so far their setup costs can outweigh their benefits; booking flights can be especially impractical when explaining mileage-program constraints takes as long as booking directly. The host sees promise in proactive work, such as monitoring email and drafting replies, or in a “mono thread” that carries context across tasks and can dispatch subagents. Better context compaction and context-window management make that approach more practical, though the host is currently more inclined to use agents professionally than personally.

ChatGPT’s Quasar serves as a mono-thread context holder for important professional work, while simple queries and daily podcast copywriting stay outside that workflow. After receiving a Claude planning conversation spanning about six hours, Quasar acknowledged saving it, but the host found its personality flat and its strategic engagement weak. The host’s initial view is that such tools add little to reactive chat or proactive work but could retain context and act on it; that assessment may change. ChatGPT and Claude matter because the host’s professional context already lives there, while Grokbot remains relevant for its X integration. The host is not investing heavily in Google’s AI tools or evaluating Microsoft’s tools as a solo operator, and expects useful features and interface ideas to spread between AI ecosystems within days or weeks.