AI token spiral — stop runaway credit burn in Cursor, Bolt, and Lovable

1. Regression loops where the AI "fixes" working code. Cursor’s Composer and Lovable’s chat both default to a proactive stance — they will volunteer refactors you never asked for, especially on files they perceive as messy. You ask for a bug fix in one handler; the model rewrites three neighbouring utility files "for consistency". One of those utilities was imported by something else. That caller now breaks. You prompt the model to fix the new break, it rewrites the utility again and breaks the original handler. Each iteration costs roughly the same input-plus-output tokens as the last, but produces negative progress. The pattern is visible in git: three commits in a row where files oscillate between states. Once you see oscillation, no amount of further prompting reverses it — only a hard reset does.

2. Context window overflow forces re-explanation.Every chat turn in Cursor, Lovable, and Bolt carries the full conversation history plus any auto-attached files back to the model as input tokens. On turn 1 you pay for 500 tokens of prompt. On turn 15 you’re paying for 15,000 tokens of prior context plus the 500 of your new message, on every turn, even though the model has already seen 90% of it. Worse, the model starts hallucinating from stale context: it will "remember" a function signature from an earlier turn that you’ve since rewritten, and regenerate broken code that references the old shape. The more you chat to clarify, the more stale context accumulates, the more hallucination creeps in, the more you chat. The only exit is a fresh session with a clean context.

3. Amplifier prompts trigger wholesale rewrites. Phrases like make it better, clean this up, add more features, polish the UI, or production-readyhave no file scope and no acceptance criterion, so the model maximises surface area to prove effort — it rewrites everything in view. A single make the landing page better prompt in Lovable has been observed to consume 80+ credits (roughly $4 at standard rates) by rewriting the hero, nav, three feature cards, pricing table, footer, and an unrelated shared Button component. None of that work was requested. Founders who would never accept a contractor touching six files on a one-line ticket accept it from an AI because the meter is silent until the invoice arrives.

4. Missing constraint guardrails.A well-scoped prompt names the file, names the function, names the change, and explicitly forbids touching anything else. Almost nobody writes prompts like this without training, and none of the AI tools enforce it by default. When you ask Bolt.new to "fix the login bug", Bolt reads every file in the project (charging input tokens for all of them), picks the three it thinks are related, and rewrites them. Two of the three are fine. The rewrites of those two introduce new bugs. Without a guardrail like only modify app/login/page.tsx in the prompt, the model cannot know the scope, and cost scales with project size instead of change size. This single guardrail cuts typical burn by 60-80%.

5. Agent tool-use loops without backoff.Cursor Agent, Lovable’s auto-run mode, and Bolt’s build-and-verify loop all let the model call tools (run tests, run the dev server, read the filesystem) on its own. When a tool call fails — the test fails, the build errors, the file isn’t found — the agent retries. There is no exponential backoff and usually no global retry cap. A failing build can trigger 8-12 auto-retries, each of which includes the full prior context plus the failing output as new input tokens. The usage graph shows this as a flat high-rate plateau over two or three minutes, not a normal spiky pattern. Disable auto-run for anything mutating until you trust the scope; then re-enable selectively.