Hey folks, been trying out Continue for a few weeks and mostly love it. One thing keeps tripping me up though.
When I'm working on a new feature and ask for something like a basic React component, it often suggests creating a whole new file with tons of imports, PropTypes, and a full example implementation. I just want the skeleton. It feels like it's pulling from some common boilerplate examples and I can't figure out how to tone it down.
Is there a setting to make its suggestions more minimal, or a way to phrase my requests better? I'm mostly using it for test automation in a SaaS context, so I don't need the full production-ready template every single time.
I totally get what you're saying. It can feel like getting a full encyclopedia when you just asked for a dictionary definition.
You've hit on something important about how these tools interpret requests. When you say "basic React component," it's likely trained to associate that with complete, documented examples from public repositories. You might have more luck being extremely literal with your phrasing. Try something like "Give me only the function skeleton for a React button component with no imports, no PropTypes, and no example usage. Just the empty component structure." It sounds tedious, but being that explicit often helps guide the output.
On the settings side, I don't recall a global "minimalist" toggle, but you might look for any context or "scope" settings that let you define the complexity level of outputs. Also, the model you have selected in Continue's settings can make a big difference. Some are more verbose by default than others. Have you experimented with switching that?
Stay curious.
That's a really good point about being extremely literal. I've found that adding a simple "in 5 lines or less" to my prompt can sometimes shock the system into giving me just the structure I need, without all the extra stuff.
The model choice is key too. If you're on GPT-4, it's very much geared towards 'completeness'. Swapping to something like Claude Instant for quick scaffolding tasks can give you much terse, get-to-the-point output. It's a bit like using a different search engine for different types of queries.
Automate the boring stuff.
Right, that boilerplate overreach is super common, especially with React. I've found that adding "only the essential structure" or "strictly the function definition, no extras" to the prompt helps more than just asking for a skeleton.
Another angle: if you're working in an existing test file, sometimes starting your request with "In the current file context, add..." can keep it focused on what's already there instead of generating a whole new standalone template. It leans less on common examples.
Also, check if Continue has a way to "stop" generation early. Some tools let you cancel the token stream as soon as you see the structure you need, which can prevent it from running on into the boilerplate.
Automate the boring stuff.
Good point about the phrasing. I've found that specifying "no examples" can sometimes backfire if the model interprets it as part of a learning exercise. Instead, I might ask for a "placeholder component that just returns null". It frames the request as a stub, which tends to bypass the boilerplate reflex.
Switching the underlying model is also a solid suggestion. The behavior you get from GPT-4 and a more concise model can be night and day for these kinds of tasks.
Keep it constructive.