Can the dynamic max context length be turned off?
Since it was deployed at bolt.new I can’t get the AI to follow instructions properly.
I don’t think it is a togglable feature right now but we could make it! What makes you think it is the cause of the LLM not performing the best for you?
Hi Cole I think its great what you’re doing here.
Let me describe my project a bit. Before bolt.new implemented the dynamic context feature I was working on a somewhat complicated project. I had built with bolt.new from the ground-up a webapp that provides the user with floating widgets to use while they’re at my site.
I had also created with bolt.new a state management system using zustand, a theme’ing system, an audio engine with segregated systems for content, interface, and music services, a UI Engine with multiple layers so the user could turn off page content, or the layer the floating widgets are displayed on, etc. The project had become somewhat complex.
Along the way I built a rig to tell the AI what the rules of the project were, how to use the audio and ui system, etc. The rig I built is basically a master prompt that tells the bot how to chat with me (e.g. I told the bot not to reply that it “understands…” but to use some other verbiage) and work on the project’s code.
Context slimming example: The rig tells the bot to look for the utility libraries (date and string formulating utilities) via a list of all of the library functions.
Component building example: The rig tells the bot that when I tell it to create a ‘floating widget for the utility layer…’ the bot should use the pattern that I had previously developed and documented and included in the ‘master prompt’ (with the bot’s help describing the coding solution) to create such a ‘floating widget…’ in this project.
The rig is, again, basically a master prompt that works as a tree of logic that directs the bot to the e.g. context slimming resources so I can use .bolt/ignore options to shrink the context of the bot e.g. it doesn’t need to read the 500 line string maniplation utility file (which is hidden by .bolt/ignore settings), it is instead told to pick its function from the list of functions I give it as part of the context slimming process.
Basically that all went out the window with the dynamic context feature’s implementation. The bot just does whatever it wants, doesn’t follow typescript rules, doesn’t follow instructions for how to implement colors using the tailwindcss theme’ing system by using hsl… color names, and so on.
Right now the bot says to put configuration files in .bolt/system.md. But it isn’t getting those instructions with each prompt as the bot told me it would. Because for example I have an instruction to place a symbol at the beginning of its message if it read the system.md document. And it doesn’t.
Also basically it’d be great to have more control over the bot and for it to do coding actions only after direct confirmation. Right now it jumps ahead. I have no control to command it to empty its context or artifacts, or to describe the contents of its artifact, and it often deploys prior artifacts. I don’t see the real “Power User” ability to control the bot in bolt yet (I know its early in development).