Ask HN: How do you get into a flow state when using AI to code?

Hacker News Discussion ↗

Programmers experimenting with AI-assisted and agentic coding tools report that traditional “flow state” is hard to maintain when work becomes stop‑start prompting, waiting, and reviewing. Many describe a loss of joy and deep engagement as coding shifts toward project management, multitasking across multiple agents, and higher‑level planning, while others say they’ve found new kinds of flow by focusing on architecture, experimentation, or tightly scoped, fast feedback loops. Across perspectives, key variables are model speed and reliability, UI/UX (chat vs more integrated tools), task chunk size, and how much control and understanding the human retains over the resulting code.

Overall sentiment on flow state with AI coding

  • Many say they cannot reach traditional flow with agentic/AI coding; experience is “stop-and-go,” like project management or babysitting a junior dev.
  • Some feel coding joy is gone; they’d avoid AI if it weren’t required, and a few have stepped back from AI entirely in new roles and report better satisfaction.
  • Others report new kinds of flow: shorter, repeated bursts of focus, or flow in higher‑level design and learning rather than typing code.

Main blockers to flow

  • Waiting for responses breaks concentration; chat UIs encourage passivity and web/YouTube distraction.
  • Slow but capable models vs fast but dumber models both disrupt flow differently (trust vs latency trade‑off).
  • Multitasking across many agents or tasks taxes working memory; context switching is widely cited as anti‑flow.
  • AI output can be hard to read, architecturally weak, or wrong, creating extra review/debug burden.

Workflows people use to cope

  • Multitask across 2–10+ agents, repos, or worktrees so something is always running; some frame this as “selective multitasking” or “prioritized context switching.”
  • Use AI only for narrow, boring, or boilerplate tasks (bug hunting, small scripts, visualization, search in deps); code interesting parts by hand.
  • Treat AI as autocomplete/pair programmer: comment-driven development, small snippets, incremental diffs, and tight verification loops.
  • Shift flow to planning/design: write design docs, plans, and prompts; AI does execution and tests; engineer stays in review and orchestration.
  • Build custom tooling (tmux/TUI/task managers/orchestration layers) to handle multiple sessions asynchronously and protect attention.

Alternative attitudes and reframing

  • Some embrace AI as delegatable “junior devs” and accept that deep uninterrupted flow is rarer at senior/staff levels anyway.
  • A few lean into experimentation and research, using agents to explore many ideas; “flow” comes from running and interpreting experiments.
  • Others see the solution more in UX than models: background reviewers, non‑chat interfaces, better notifications, and terser outputs.

Coping strategies outside tooling

  • Suggestions include meditation during waits, keeping scratch/ideas docs, listening to music, or even psychedelics; effectiveness is subjective and mixed.