Ask HN: How is AI-assisted coding going for you professionally?
AI-assisted coding is reshaping professional software work in uneven ways: some engineers report 2–10x speed-ups on boilerplate, legacy code analysis, tests and greenfield features, while others say it just shifts effort into exhausting review and cleanup of unreliable “slop” code. Experiences vary by codebase, tooling and process, with the best results coming from tight human-in-the-loop workflows (spec → plan → implement → test → review) and AI‑friendly repos, rather than unchecked “vibe coding.” Beneath the productivity gains runs a deeper anxiety about skill atrophy, inflated management expectations, the hollowing out of junior and mid-level roles, and what software engineering looks like once most code is machine-generated.
Overall sentiment & divergence
- Experiences range from “game‑changer, 5–10x faster” to “net negative, I want to quit.”
- Many describe simultaneously feeling empowered on side projects and burned out or anxious at work.
- Strong split between people who feel they’ve learned to harness tools and those who find them inconsistent, myopic, or unusable on complex systems.
Where AI tools help
- Greenfield / small, well‑scoped tasks: scaffolding apps, CRUD APIs, CI/CD, glue scripts, data munging, infra boilerplate, tests, refactors, local tools.
- Understanding large or legacy codebases: “what touches X?”, “how is auth done?”, summarizing flows, exploring unfamiliar repos, generating diagrams.
- Debugging: pinpointing bugs, reading logs and traces, fixing test fallout after refactors, triaging build failures.
- Non‑coding: design docs, ticket drafting, documentation clean‑up, research and architecture brainstorming.
Where they fail or cause harm
- Complex, interconnected systems: client/server interactions, legacy monoliths, niche domains, performance‑critical or numerical code.
- Architecture and design: tendency to over‑engineer, duplicate logic, add layers and custom parsers, drift from existing patterns.
- Reliability: hallucinated APIs, wrong docs, subtle security bugs, timing issues, and brittle refactors.
- “AI slop” PRs: huge diffs that superficially look fine but are conceptually wrong, bloated, or unmaintainable.
Team and organizational dynamics
- Some orgs mandate “AI‑first” or even “100% AI‑generated code,” others ban or heavily discourage it.
- Senior engineers report becoming “code janitors,” cleaning up AI‑generated mess from managers or peers.
- Code review is a new bottleneck: more and bigger PRs, reviewers overwhelmed, tension over quality vs velocity.
- Management sometimes uses AI to mass‑produce design docs, tickets, and performance text that nobody really reads.
Impact on careers & skills
- Many fear skill atrophy, loss of “craft,” and hollowing out of mid‑level roles; others lean into design/architecture and accept less typing.
- Juniors can now produce large change sets without understanding them, making mentorship and review harder.
- Some see solo‑dev and small‑team opportunities exploding; others anticipate layoffs or a sharp rise in expectations without matching rewards.
Emerging best practices
- “Spec → plan → critique → implement → review” workflows; plan mode before code.
- Strong tests, CI, and e2e coverage as guardrails; reject AI code that doesn’t move tests from red to green.
- Repo hygiene for “AI‑native” development: AGENTS.md/CLAUDE.md, coding style guides, skills, and scripts for common workflows.
- Use AI heavily for understanding, small increments, and boring work; keep humans fully responsible for design and final quality.