Ask HN: What is your (AI) dev tech stack / workflow?

Hacker News Discussion ↗

AI-assisted software development is fragmenting into a wide range of toolchains, from simple “text box plus terminal” setups to complex multi-agent factories orchestrating specs, tests, and deployments. Developers report that the real leverage comes less from specific tools (Claude Code, Codex, OpenCode, Pi, etc.) and more from practices like precise problem descriptions, spec-driven development, TDD, and careful context management. Many caution against over-automation and opaque agents, emphasizing sandboxing, small feedback loops, and starting beginners on minimal, transparent workflows rather than elaborate harnesses.

Common Tools and Stacks

  • Most use a small core: Claude Code, Codex (OpenAI), OpenCode, Pi, VS Code/Zed/Neovim, Git, Docker, FastAPI/TypeScript/Python.
  • Several build custom harnesses or terminal-first setups (tmux/zellij/Ghostty, TUI launchers, custom Vim/Neovim integrations).
  • Others rely on cloud agent platforms and orchestrators (Conductor, exe.dev, Pi, Mecha-AI, various MCP/skills-based systems).
  • A minority stick to traditional IDEs (PyCharm, IntelliJ, plain vim) or refuse AI tools entirely.

Workflows and Roles for AI

  • Common patterns: architect → implementer → reviewer agents; or “sword and shield” (one model codes, another audits).
  • Many use AI mainly for debugging, explanation, refactoring, and test generation, not raw code writing.
  • Some go “full factory”: multi-agent swarms, worktrees per feature, CI/CD-heavy pipelines, and continuous automated test runs.
  • Others deliberately keep it simple: one agent, one repo, request minimal diffs, manual review and testing.

Context, Specs, and Planning

  • Context loss is widely seen as a bigger problem than model quality.
  • Popular mitigations: spec-driven development, detailed markdown requirements, Gherkin stories, todo.md / ticket hierarchies, and immutable phase docs (discovery → plan → implementation → verification → review).
  • Several keep separate repos or folders just for plans and AI artifacts to preserve context across sessions.
  • Persistent “memory” features are often disabled due to unpredictability; explicit written artifacts are preferred.

Automation vs Control & Quality

  • Strong emphasis on TDD, linters, type systems, and CI as guardrails against sloppy or unsafe AI output.
  • Some run agents in VMs/containers with restricted credentials to avoid destructive actions.
  • Views differ on full automation: some claim thousands of AI-driven commits; others only trust AI for adjacent tooling or small utilities.

Teaching and Newcomers

  • For beginners, many recommend: minimal tooling (Claude/Codex + simple editor), tight feedback loops, and focus on learning to describe problems precisely.
  • Several warn that over-complex multi-agent setups cause confusion; start with “slow code,” strong specs, and manual review.

Skepticism and Concerns

  • A visible minority avoid AI on principle or for learning-quality reasons, stressing the value of struggle and deep understanding.
  • Others worry about “vibe-coded” messes, context mismanagement, and long-term maintainability of AI-generated code.