AI lets you design software through prompts instead of typing every line. The challenge moved from writing code to framing problems, reviewing outputs, and orchestrating agent workflows—experience still matters, just upstream.
The story
AI is making “software development experience” less of a gate and more of a multiplier. The skill is shifting from typing code to steering a small swarm of very competent robots who never get bored and occasionally get confused in creative ways. So, can anyone write good software now? Potentially, yes, if you can think clearly and prompt like you mean it.
1] Experience moved upstream
Devin and Cognition-style agents can go from natural language to planning, coding, and troubleshooting. That doesn’t delete craft; it relocates it into problem framing, constraints, and review.
2] Prompting is architecture, in disguise
If you ask, “Break this into modular services with clear APIs, SOLID boundaries, tests, and guardrails,” you’re doing design. AI can fill gaps in modularization, but you still choose what “good” means.
3] Model choice is a strategy, not a religion
Use domain-fit tools and verify them. People compare models with benchmarks like SWE-bench, and pick reasoning-heavy options like Gemini 2.5 Pro or agent-control-friendly ones like Claude when orchestration matters.
4] Treat agents like a team, not a vending machine
Multi-agent setups like AutoCodeRover and OpenHands, plus desktop-control systems such as OpenAI’s CUA or Anthropic’s ComputerUse, can run workflows end-to-end. Your job is the spec, the acceptance criteria, and the sanity check.
5] Quality is still a consent issue
If you want safe, traceable output, ask for it, then audit it. Review in JetBrains AI or GitHub Copilot flows, run SonarQube, and keep traceability in mind, EU AI Act style. Goldman Sachs fine-tuned internal AI on their codebase for context-aware generation, and surveys say 84% of developers use or plan AI tools, so the norm is shifting.
You can ship with vibe, or you can ship with proof. Prompt for a plan, then test what it ran, that’s the span.
Recent Comments