How I code with AI changed a lot

Video thumbnail: How I code with AI changed a lot
May 27, 202647m 34s video lengthTheo - t3․gg

The Signal

Theo, a software engineer and AI product builder, reports that his workflow has shifted entirely since late 2025. He now rejects terminal-heavy or plan-centric agent flows in favor of short, simple prompts and persistent context steering. He views his current 'agent-native' development process as a battle between bloated, overly complex toolchains and simpler, remote-hosted app interfaces that treat the AI as a conversational collaborator rather than a rigid instruction-follower. He asserts that his preferred stack for these workflows, centering on GPT-5.5 and stable remote-app UIs, is now superior for productivity, though this remains an experience-based claim rather than an industry standard.

The Case

  • Theo reports that his most significant practical gain is not improved model intelligence, but aggressive thread-resetting: he started over 100 isolated threads in five days to prevent irrelevant context bias, rather than relying on parallel worktrees.30:44
  • He characterizes his previous reliance on 'plan mode' and complex skill sets as counterproductive, finding that simple, two-sentence prompts coupled with clear glossary definitions for terms like 'developer' versus 'agent' yield faster, higher-quality results.33:25
  • He notes a frustrating divide in remote coding experiences, citing a 30-second to two-minute input lag in Codeex desktop-remote sessions compared to the 'pleasant, stable' performance of his own T3 Code tool running on a remote Mac Mini.12:45
  • He frames his own project, Lakebed, as an 'agent-native app substrate,' noting that the model successfully 'roasted' the project’s initial bloated feature list and forced him to sharpen its thesis into a narrower tool for building small full-stack applications.23:27
  • He emphasizes that agents must perform their own verification to be reliable, favoring workflows where the AI uses built-in tools like CLI tests, browser-based computer use, and screenshot analysis to debug its own PRs before he reviews them.39:29
  • He flags that his experience with usage limits—specifically the $200-per-month credit threshold for non-native integrations in the Anthropic/Claude ecosystem—is shaped by atypical event-related multipliers, making his personal cost-efficiency observations potentially inapplicable to others.4:20

The 1 Minute Signal Take

This video is a useful manual for developers struggling with 'agent bloat' and tool-induced friction. While Theo’s bias toward T3 Code is clear and his preference for app-based interfaces over CLI-native workflows is a subjective stance, his emphasis on context hygiene and aggressive thread-resetting is a compelling, evidence-backed productivity shift. Watch this if you want a granular look at steering models toward self-verification; skip it if you are purely looking for a neutral feature comparison of current agentic IDEs.

Pro Analysis

Strategic Significance:

  • This marks a shift toward 'agent-native' development where the IDE is a secondary consideration to the agentic harness. It signifies the maturation of AI tools from simple 'autocomplete' utilities to autonomous co-workers capable of full-stack delivery.

Who Should Care:

  • Senior developers and technical leads managing projects where AI integration is slow or error-prone.
  • Tooling engineers building the next generation of developer-focused AI products.

Contrarian Takeaway:

  • Formal planning is often a distraction. The most effective way to ship is not to map out every step in a Jira ticket, but to start a, discuss, and let the model iteratively 'roast' and refine your thesis in real-time.
Time saved:44m 56s

Share this

Tags

Written by: 1 Minute Signal Editorial Team