You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Other tools try to "guess" what pieces of the codebase matter for the given task, often struggling to get it right. They workaround this by overloading context or making excessive number of API calls.
44
+
Coding agents are designed to offload the burden of relevant context selection from the programmer. With today's advancements in LLMs capabilities, the technology is very successful in its work bringing undeniable gains in speed software development can be done. For some it is too fast too keep up though. Over time they lose confidence in their own knowledge about the codebase they work on, veryfing AI-suggested changes becomes challenging, bugs and unintended behaviors are much more likely to be introduced.
46
45
47
-
CWC works differently:
46
+
CWC does not prompt-engineer its functionality and expects programmer to be constantly in the loop with the evolving codebase by requiring accurate context selectios to the task.
48
47
49
-
-**You select** which folders and files are relevant to the task
50
-
-**You know** exactly how many tokens you have in the context
51
-
52
-
The result? Unmatched in accuracy and cost AI assistance.
53
-
54
-
> Too many tokens fighting for attention may _decrease_ performance due to being too "distracting", diffusing attention too broadly and decreasing a signal to noise ratio in the features. ~Andrej Karpathy
48
+
This friction rewards in building mental model about the codebase, faster code generations and AI-related cost savings.
@@ -65,13 +59,9 @@ The Connector extension is available in [Chrome Web Store](https://chromewebstor
65
59
66
60
> <small>**Legal Disclaimer:** After chat initialization, the extension does not read the incoming message. The injected _Apply response_ button is not a means of automatic output extraction, it's an alias for the original _copy to clipboard_ button.</small>
67
61
68
-
**TIP: Practice single-turn interactions**
69
-
70
-
Chat conversations are only a construct of product interfaces, they hurt the quality of responses from the model and once your context is "poisoned" it won't recover. Whenever you're not satisfied with a reponse, **the best practice is to alawys refine your initial instructions and reinitialize**.
0 commit comments