?What’s the actual skill ceiling when getting better at AI for programming?

On Hacker News, commenter JKHeadley described their multi-year arc of getting better with LLM-assisted coding: From early on it became clear to me that the most important factor for successful development with LLMs is context management. It started with rich, natural language comments to guide the tab completions, then moved to more complete documentation markdown files to inform the coding agents, then agent specific files like Cursor rules, AGENTS.md, and CLAUDE.md to instruct agents on how to navigate and maintain the existing documentation. Each step was an architectural progression with the single goal of allowing a fresh LLM agent to get up to speed and start contributing effectively without me having to provide any context beyond the task at hand. Pattern: getting better at AI for programming is not learning prompts. It’s a multi-year stack of progressively more durable context artifacts — inline comments → doc files → tool-specific instruction files. Each level reduces the per-task setup cost. The OP’s idiomatic-code problem is downstream of a missing CLAUDE.md / AGENTS.md spelling out idiomatic SvelteKit in this codebase means X.

From multiple 2026 guides on AI agent config files (Augment Code, HumanLayer, deployhq): AGENTS.md is a dedicated, predictable place to provide context and instructions to help AI coding agents work on your project. In December 2025, AGENTS.md was donated to the Agentic AI Foundation (AAIF) under the Linux Foundation, alongside Anthropic’s Model Context Protocol (MCP) donation. But the harder finding is on size: If your config file is over 500 lines, most of it is being ignored — a focused 50-line file outperforms a sprawling 1,000-line one. Frontier thinking LLMs can follow approximately 150-200 instructions with reasonable consistency. Claude Code’s system prompt contains about 50 individual instructions, which is nearly a third of the instructions your agent can reliably follow already. Pattern: the leverage move for OP is not write a giant style guide but write a 30-50 line AGENTS.md / CLAUDE.md that names the 5 idiomatic patterns the LLM keeps violating (no div-spamming, prefer Svelte’s $state over useState patterns, use tailwind not bootstrap classes, etc.). Specificity over volume.

From the 2026 AI agent config guides: LLM-generated files give negative returns with worse performance at higher cost, while human-curated files yield roughly a 4-percentage-point improvement, making writing manually worth the overhead. Rules should respond to observed failure, not be generated speculatively. Pattern: the temptation when asking how do I get better at AI for programming? is to ask the LLM. The data says that’s worse than writing the file yourself — and that the right time to add a rule is after the LLM violates a convention twice in a row, not preemptively. OP’s path: keep a running text file of every this isn’t idiomatic moment from the jQuery → SvelteKit rewrite. After 2 weeks the file becomes the personal AGENTS.md, hand-curated, ~50 lines, that the next LLM session reads first.