2x, not 10x: coding with LLMs in 2026

There's a case going around that LLM coding tools are plateauing in usefulness, not because the models got worse, but because they already cleared the bar that actually matters. The author's theory, laid out in a July 2026 post, is that LLMs took off this year mainly because they became reliable enough to run inside automated feedback loops: tell one to build a button that does X, let it click around and check whether X happened, and it can iterate without spinning out. That's genuinely useful. The author calls it life-changing, maybe a 2x improvement. But once a model clears that reliability threshold, further gains in raw model performance start mattering a lot less, the same way being tall enough to climb stairs one at a time beats being tall enough to take two or three, but that height advantage tops out fast.

The catch is what LLMs still can't judge well: whether code is structured in a maintainable way, or whether documentation includes the right information and skips the rest. Those aren't things you can verify with a quick pass/fail check, so the models can't reliably predict what a human would accept. The result, in the author's own workflow, is using LLMs to produce a rough draft that then gets heavily reworked until the structure feels right — with an admitted looseness on line-by-line readability along the way. And even accounting for that looseness, the author says they keep underestimating how long the cleanup takes. A working implementation used to mean a task was roughly 80% finished; now it feels closer to 20%.

On documentation specifically, the fix has been blunt: instruct the model to never write READMEs, docstrings, or comments at all, and write them by hand afterward. It's a workaround, not a capability upgrade, and that distinction is the whole point of the piece. The tempting assumption is that a year of rapid improvement means next year's models will simply get much better at documentation and code structure too. The author pushes back on that, arguing climbing a tall staircase doesn't mean you can suddenly swim. Different skill, different problem.

So the prediction is that a 10x productivity leap over what the author calls the dark ages of 2025 won't come from model improvements alone. It'll come from the industry rebuilding workflows and tooling around the capabilities that already exist. The author traces their own path from using LLMs like a Stack Overflow replacement, to chatting with them interactively while coding, to writing declarative specs of the desired end state, with sandboxed environments as a small but real upgrade so permissions don't need approving every 30 seconds. There's also a nod to vibe coding, defined here as generating code without fully reading or understanding it, which the author has tried only on non-work projects and finds intriguing but unproven over any meaningful stretch of time. Maybe better testing practices eventually make black-box LLM code safe even for critical infrastructure. For now, the author is keeping the hand-written READMEs.