Add some learning prompts and "send to edit session" for inline chat (#267519)

* wip

* update learnings

* clean 1

* clean 2

* commit update-instructions.prompt.md

* swap back to normal model after being in delegate

* adding generation

* wip

* wip

* update

---------

Co-authored-by: Henning Dieterichs <hdieterichs@microsoft.com>
Co-authored-by: Anthony Kim <anthonykim@microsoft.com>
Co-authored-by: SteVen Batten <sbatten@microsoft.com>

.github/instructions/learnings.instructions.md

@@ -0,0 +1,32 @@
+---
+applyTo: **
+description: This document describes how to deal with learnings that you make. (meta instruction)
+---
+
+This document describes how to deal with learnings that you make.
+It is a meta-instruction file.
+
+Structure of learnings:
+* Each instruction file has a "Learnings" section.
+* Each learning has a counter that indicates how often that learning was useful (initially 1).
+* Each learning has a 1-4 sentences description of the learning.
+
+Example:
+```markdown
+## Learnings
+* Prefer `const` over `let` whenever possible (1)
+* Avoid `any` type (3)
+```
+
+When the user tells you "learn!", you should:
+* extract a learning from the recent conversation
+	* identify the problem that you created
+	* identify why it was a problem
+	* identify how you were told to fix it/how the user fixed it
+* create a learning (1-4 sentences) from that
+	* Write this out to the user and reflect over these sentences
+	* then, add the reflected learning to the "Learnings" section of the most appropriate instruction file
+
+
+	Important: Whenever a learning was really useful, increase the counter!!
+	When a learning was not useful and just caused more problems, decrease the counter.