?Is it just me, or did vibe coding stop working after the first two months?

On Medium’s Vibe Coding publication, an article documents: Anthropic published a postmortem confirming 3 bugs broke Claude Code for a month. The issues included hidden reasoning drops, memory wipes, and a 25 word response cap. Pattern: the two months in, it stopped working feeling is partially real and partially shifting expectations. If your two months happened to overlap with a known incident window — the postmortem covers a specific month — your output really did degrade. The lesson is to keep a personal log of this used to work last week, now it doesn’t moments tied to dates, so when the next platform incident hits you have actual evidence rather than vibes. Right now the tools have no SLA you can call them on; your log is the only audit trail.

From a 2026 Vibe Coding overview: AI models hallucinate by making up function names that do not exist, referencing deprecated APIs, and inventing configuration options that never existed, resulting in bugs that are particularly insidious because the code looks correct. Pattern: month 1 of vibe coding, your reviewer reflex catches obvious garbage (syntax errors, made-up libraries you’ve never heard of). Month 2, the model gets better at producing plausible code and your reflex relaxes — but it’s still inventing things, just at a level your skim doesn’t catch. The subtle missing the mark feeling is the bug class that survived your filter. The fix isn’t a better model; it’s restoring a stricter manual verification step (run every snippet, check every import, type-check every function signature) that you let lapse when month 1 felt easy.

Multiple HN discussions on AI coding cite the same shape: the early adopters who plateau around 2 months are the ones who stopped investing in their own workflow infrastructure. The ones who keep getting faster invested in context management artifacts — CLAUDE.md / AGENTS.md, structured task lists, separate planning sessions. Pattern: the addictive easy mode of month 1 is the no structure needed, model figures it out phase. Once you’ve used the tool enough that your tasks involve more state and history than fits in a single chat session, the model stops figuring it out — and you experience that as the model getting worse. The skill that distinguishes month-3 users from month-2 users is willingness to do the boring structural work (write the spec first, break into tasks, version the prompts) that the month-1 honeymoon let you skip.