Tech • AI • Robotics • Game
Claude Duet preview lands as Base 44 and Codex reshape AI app work
Briefing
Claude Duet debuts proactive workspace
Claude Duet emerged in preview as a more autonomous assistant built around persistent work rather than one-off chat. The product organizes activity into channels, groups, sessions, artifacts and tasks, signaling a project layer above a standard assistant. Anthropic positions it closer to a proactive agent that keeps work moving 24/7. Early hands-on testing, however, found the experience slower and more fragmented than higher-control agent systems.
Claude Duet leans on connectors
A central part of Claude Duet is its reliance on external app connections including Gmail, Google Calendar, Slack, Google Drive, Linear, Canva and Figma. The promise is a shared workspace that can monitor context and act across tools. In practice, setup was described as cumbersome because each integration required separate authorization. That repeated consent flow weakens the seamless assistant narrative Anthropic is aiming for.
Agent hierarchy sharpens around Codex
The discussion drew a clear line between three layers of assistant software: chatbots, execution agents and proactive agents. Tools such as Claude Code and Codex sit in the execution tier, where they can carry out concrete tasks on command. Claude Duet was framed as a step beyond that model by initiating work and tracking projects over time. The comparison underscores how fast the market is moving from prompt-response systems toward ambient operator software.
Base 44 pushes planning first
Base 44 was highlighted as an example of AI app builders shifting emphasis from coding to specification. Its plan mode can ask follow-up questions, map user flows, define pages and outline data relationships before generation begins. That workflow treats product planning as the cheapest place to catch mistakes, since every extra prompt consumes time or credits. The broader message is that structure now matters more than syntax for many first-pass applications.
Four-question brief becomes essential
A strong AI app brief was reduced to four core questions: who the users are, what they need to do, what data must be stored and how the experience should feel. A community-club example showed how adding details such as California organizers, event fields and RSVP logic sharply improved output quality. The contrast was with vague one-line prompts that force the model to guess requirements. For builders, better prompting now looks increasingly like lightweight product management.
Iteration beats one-shot generation
The strongest workflow advice was to build in small increments rather than ask for a complete app in one pass. Early planning and stepwise review reduce rework once interfaces, databases and business logic start compounding. That is especially important in no-code and AI-generated environments where assumptions can silently propagate through a project. The lesson is straightforward: iteration has become a cost-control mechanism as much as a quality method.
Testing remains the missing safeguard
Even as AI tools generate interfaces and backend logic, testing was presented as the final gate before publication. The warning is that polished first drafts can still hide structural or data-handling errors if the original brief was incomplete. Builders were urged to verify flows, stored fields and user permissions before shipping. In effect, AI lowers the barrier to creation but not the obligation to validate.
La Belle Vie offers product lesson
Outside software, La Belle Vie provided a business case study in disciplined systems design. The Paris grocery delivery company survived quick-commerce's collapse by focusing on dense urban logistics, proprietary software and large baskets near €100, rather than chasing €10 to €15 impulse orders. It reportedly raised only €33 million to €35 million, far below many rivals that expanded across Europe in 2021. The takeaway for AI builders is familiar: tight operating assumptions and clear target users still matter more than speed alone.
Videos covered
The Ultimate Beginner’s Guide to Building Apps with AI
- •Communication over coding
- •Four questions shape the brief
- •Planning cuts cost and confusion
La startup qui a survécu au massacre à plusieurs MILLIARDS
- •A survivor in a market wipeout
- •A different customer model
- •Margins leave little room for error
Claude Duet : mon avis honnête après l'avoir testé pour de vrai
- •Three tiers of AI agents
- •Preview interface adds a dedicated workspace
- •App connections are central but cumbersome