diff --git a/src/content/posts/WIP-getting-good-results-out-of-figma-mcp.md b/src/content/posts/WIP-getting-good-results-out-of-figma-mcp.md
index 37c0ac1..477b437 100644
--- a/src/content/posts/WIP-getting-good-results-out-of-figma-mcp.md
+++ b/src/content/posts/WIP-getting-good-results-out-of-figma-mcp.md
@@ -9,6 +9,7 @@ draft: true
 # Framing
 
 - Disposable prototype for early feedback on a bigger upcoming project where lots of things are not yet set in stone.
+- You might thinkAI is probably able to build things correctly, if given correct specification, proper guardrails and so on. So what we're working on here is an artifact that will allow us to come up with the correct specification of this product. 
 
 I'm personally in favor of low-fidelity designs for this purpose.
 Like a quick idea sketched out onto a napkin. Or rough drawing on a whiteboard. 
@@ -42,5 +43,12 @@ const Button = () => {
 Design context quite bloated, feels very redundant/lossy for large pieces of work
 Get metadata more useful, but is needs augmenting. 
 
+# Running claude in a loop
+
 - combine with the ralph wiggum technique
+Hyped a lot RN, to the point of having a memecoin
+Haven't actually seen anything built with it. Saw Matt Pocock's video, among other things, but he only showcases running a single interactive session with a very simple feature.
+- run CC as a cli https://code.claude.com/docs/en/cli-reference
+You will need a few flags, like --permission-mode acceptEdits and --output-format stream-json (in combination with --verbose)
+TODO: might make sense to write output to terminal (to monitor that stuff's happening) + to file, so that you can later analyze what actually happened.
 - data, results, precision, errors?