Ask HN: Anyone struggling to get value out of coding LLMs?

Where LLMs Help Today

• Strong for boilerplate and small, self‑contained tasks: CRUD endpoints, React components, regexes, scripts, simple SQL, Dockerfiles, migration of queries between DBs, etc.
• Useful “rubber duck” / research tool: explaining libraries, APIs, math, or unfamiliar stacks; summarizing bad docs; locating likely bug areas in new repos.
• Good for scaffolding greenfield MVPs and throwaway utilities: many report building landing pages, small apps, internal tools, and data‑munging scripts they’d never have had time to write themselves.
• Helpful for tests, refactors, and polishing: suggesting better names, formatting, JUnit tests, minor refactors, basic security reviews.

Where They Struggle

• Reliability and trust: non‑determinism, hallucinated APIs, subtle bugs, broken invariants, and regressions when modifying existing codebases. Everything must be reviewed; many find that slower than writing code themselves.
• Larger, evolving projects: models lose track across files, undo prior decisions, re‑introduce removed patterns, and collapse after enough iterations. Context‑window limits and weak codebase understanding are recurring complaints.
• Complex or novel domains (compilers, intricate SQL, legacy systems, highly constrained data structures) often yield shallow or simply wrong solutions.

Workflow, Tools, and “Using Them Right”

• Best results come from: tight scoping, incremental changes, heavy use of tests, explicit specs and rules files, and treating the model like a bright but inexperienced junior.
• Several report big gains only after reorganizing projects around LLMs (spec directories, ticketing, MCP/RAG for targeted context, strict conventions).

Productivity, Quality, and Jobs

• Reported impact ranges from negative to “1.25–2x” to “100x” (mostly for non‑experts or greenfield work). Many note: LLMs raise the floor more than the ceiling.
• Common tension: they produce “working code” quickly, but often low‑quality or hard to maintain; good engineers still spend most time on design, domain understanding, and debugging.
• Broad agreement that LLMs are not a silver bullet or autonomous replacement yet, but are already meaningful accelerators for certain tasks.