My AI-Assisted Workflow
Developers are sharing increasingly elaborate “AI-assisted workflows” that use LLMs for rubber-ducking, drafting specs, planning tasks, and generating code, often orchestrated through tools like GitHub, Linear, or Notion. Many find value in treating AI as a spec-driven implementation aid, but others argue this adds process overhead, risks “productivity theater,” and still leaves humans doing the hardest work: clarifying requirements, ensuring correctness, and testing. Underneath is a broader split between those who see LLMs as a powerful but narrow tool and those who view current hype, reliability issues, and ad‑hoc “skills” as evidence that AI is far from the hands-off coding assistant it’s marketed to be.
Meta‑discussion: “AI workflow” articles
- Several see these posts as repetitive “influencer-style” content, akin to people sharing
.vimrcfiles, but with fewer concrete takeaways and more vibes. - Others defend them as useful for newcomers, friends, or small audiences; even if workflows are similar, one good article may be the only one a reader sees.
- Some criticize posting such pieces to HN/Reddit as implicit self‑promotion or clout seeking.
Spec‑driven / “waterfall-ish” workflows
- Many describe workflows that start with problem discussion, then detailed specs/PRDs, then task breakdown, then implementation and review.
- Some observe this strongly resembles a scaled‑down waterfall model; others stress that frequent feedback and changeability distinguish it from classic waterfall.
- There is agreement that “thinking first, coding second” is core software engineering, regardless of AI.
Where AI helps vs. fails
- Supporters: AI is helpful as a “rubber duck” to clarify problems, draft specs, generate boilerplate, mutate code for test coverage, and act as a verbose search engine.
- Critics: AI often produces wrong, overcomplicated, or unsafe code, especially for concurrency or non‑trivial logic; manual review is mandatory.
- Consensus: AI is better for one‑off, iterative tasks than for fully automated, repeatable, high‑stakes workflows.
Prompting, “skills,” and meta‑evaluation
- Some describe elaborate agent “skills” and orchestrators (e.g., PRD‑to‑issues pipelines, status workflows, sub‑agents).
- One camp argues you should feed skills back into an LLM for critique; they see this as a kind of “lint” or introspection.
- Another camp calls this unreliable “vibes scoring,” noting LLMs optimize for plausible text, not truth, and that repeated runs would yield inconsistent scores.
Productivity, hype, and skepticism
- Several worry current AI workflows are “productivity theater”: more process and overhead, not clearly better outcomes.
- Strong skeptics say LLMs underdeliver compared to their marketing (“nation of PhDs,” mass unemployment predictions) and that reliance on them may be harmful.
- Others argue that, despite overhype, LLMs are still useful tools if used cautiously and combined with solid testing and human judgment.