When to Start a New Chat Instead of Continuing

The one-sentence version: a full conversation competes with your current request

Every message in a conversation, yours and the agent's, stays in context for everything that follows, at least until a platform's own management kicks in. That's genuinely useful early on, since it means later prompts don't have to re-explain what's already been established. But it has a real cost too: a long conversation accumulates abandoned approaches, superseded instructions and dead-end attempts that never actually get removed, all of it still sitting there, competing with your current, actual request for the same limited attention. The practical effect isn't usually a hard error, it's a gradual decline: the agent gets slightly worse at holding your current intent clearly, well before any formal limit is reached.

The qualitative signals, not a token count

Counting prompts or tokens isn't a reliable rule, since one exchange involving a large file crowds a conversation far more than ten short, simple ones. Two signals matter more in practice. First, the conversation's own title drifting away from what you're currently working on, since most platforms auto-generate or let you see a summary title, and a stale one is a sign the thread has genuinely moved on from its original purpose. Second, being able to name two or three abandoned approaches or superseded instructions still sitting in that history, the same dynamic this site's own Prompting hub covers under "why prompts decay as a project grows." Either signal showing up is the actual cue to restart, not an arbitrary count.

What actually differs by platform

Bolt.new's own mechanism is genuinely distinct and worth naming specifically: it resends your entire project's files, along with its own special instructions, with every single prompt, to fill the underlying model's context window, not just your conversation history. That means a large project can approach a context limit faster than conversation length alone would suggest, since the project itself, not just the chat, is what's filling the window each time. Claude Artifacts has two separate limits worth telling apart, a filled context window and a completely different, exhausted usage quota, covered in full in this site's own guide to Claude Artifact token limits. Lovable's and Base44's own exact context-management mechanics aren't publicly documented in the same specific detail as Bolt.new's project-resend behaviour, which is itself worth knowing honestly: on those two platforms, the qualitative signals above matter more than any platform-specific rule, since no confirmed one exists to lean on.

What to actually carry into the new one

A short, current-state summary, not the old conversation copied back in. What's already built, what's still pending, and any decision that genuinely has to be respected going forward, two or three sentences is usually enough. Pasting the entire previous conversation defeats the purpose entirely, recreating the same crowded context a fresh start was meant to escape. This is the same discipline this site's own guide to giving your AI builder project context covers for a project's persistent context, applied here to a single conversation's opening message instead.

Not every long conversation is a problem

Staying in one thread has real value worth weighing against the cost above: later prompts genuinely can be shorter and more direct, since the agent already holds the project's established framework without you re-explaining it each time. The point isn't to restart constantly out of caution, it's to notice the actual signals, drift and pile-up, rather than either extreme: needlessly restarting a clean, productive conversation, or pushing a genuinely crowded one long past where it's still helping you well.