Context Drift

Context drift is what happens when an AI coding agent starts a session respecting your design tokens, component library, and spacing rules then gradually forgets them. By turn twelve it is inventing new colors and breaking your vertical rhythm. The problem is not the model. It is the accumulating noise in the context window that dilutes your original instructions.

Every designer hits context drift in week two of serious vibe coding. The first three prompts look perfect. The next five start to slip. By the time you are deep in a refactor the output looks like it came from a different brand. This is a workflow problem that feels like an AI problem.

Context drift is not the same as model hallucination. Hallucination invents facts that do not exist. Drift slowly erodes the constraints you set at the beginning. The model still makes sense. It just no longer matches your system. That is harder to catch in the moment.

The fix is pinning. You keep the tokens.json or globals.css file referenced in every single prompt. Cursor rules and Claude Skills can re inject the full constraint list automatically. Treat context efficiency as its own discipline instead of hoping the model remembers.

Concrete example. A designer spent four hours building a dashboard in Cursor. The first components used the correct semantic tokens and radius scale. By the time they reached the settings panel the shadows were wrong, the type scale had drifted, and the layout grid was ignored. The agent had seen too many new requirements and the original design system instructions fell out of the effective context. The designer learned to start every new session with a pinned system prompt that restated all token rules before any feature work began.

Another team using Claude Code for a large migration hit drift across forty files. The early files respected accessibility standards. The later files reverted to generic focus states. They solved it by creating a dedicated constraints.md file that every prompt was forced to read first. The drift disappeared.

Use techniques to fight context drift when you are working on large codebases or long sessions that span multiple hours. Pinning, rules files, and one shot spec prompts keep the model honest. These tactics are table stakes for anyone serious about vibe coding at scale.

Do not bother fighting context drift during early ideation or tiny throwaway prototypes. The overhead of constant pinning slows you down when speed of exploration matters more than pixel precision. Know the phase. Early exploration tolerates drift. Production code does not.

The best teams in 2026 treated context drift like a known law of nature instead of an occasional bug. They built their entire prompt library around the assumption that the model would forget. That assumption led to tighter specs, better seeding, and shorter sessions. The output quality became predictable instead of magical.

Context drift reveals the difference between amateurs and professionals in vibe coding. Amateurs hope the model remembers. Professionals build systems that do not require memory.