Obsidian Semantic Search Magic, Claude Code Workflows & Windsurf Adventures
Hey everyone! This week delivered some genuinely sweet discoveries around making Obsidian notes searchable in ways that matter and Claude Code workflows that just work together. The kind of setups that make you wonder why you didn't try them sooner,
On Rediscovering Old Notes
The semantic search setup I shared in my Obsidian MCP Tools video has been even more powerful than I initially realized. I'm three years deep into daily voice journaling now, and having all those notes become instantly searchable and discoverable is genuinely transformative.
Here's what's been blowing my mind: I'll search for something simple like "Based on my journals, what are entertaining activities I like that would help me recharge?" and suddenly I'm rediscovering computer games I added to my wishlist two years ago, films friends recommended that I completely forgot about, or creative projects I abandoned but might want to revisit. The semantic search doesn't just find exact matches – it understands context and intent in ways that folder structures and traditional tags never could.
There's something profound about having three years of notes suddenly become searchable and discoverable again. This is all enabled by daily voice journaling, which has become one of my most valuable habits.
On Models Orchestration
Here's something I've been playing with that's been surprisingly effective: using different AI models for what they're actually good at. Here's a technique that's been incredibly effective: I pipe semantic search results from my Obsidian vault into Gemini CLI. It's been working beautifully for a specific reason – context windows.
Gemini models have massive 1M context windows – about five times larger than Claude Sonnet (200k). So my workflow now looks like this: Claude Sonnet handles the initial orchestration and reasoning, but when I need to process large amounts of retrieved information, I hand it off to Gemini.
For example, when the semantic search pulls up 15 related notes for a my new idea I'm capturing, Gemini can analyze all of them together and suggest much more nuanced connections than Claude could fit in its context window.
The key insight is that Gemini models excel at information manipulation, but they're terrible at autonomous reasoning and tool use. Claude Sonnet is the opposite – incredible at orchestration and decision-making, but hits token limits faster. Together, they cover each other's weaknesses perfectly.
If you want to try this approach yourself, I've shared the /semantic-search command snippet on GitHub that shows how I'm combining Gemini CLI with Claude Code.
On Useful Tools
Dropover: The Missing Piece in File Management
I've been using Dropover more and more lately, and it's one of those simple ideas that just works. The concept is straightforward – it gives you a shelf where you can temporarily store files and access them later through shortcuts. But in practice, it eliminates so much friction from my workflow. No more digging through downloads folders or trying to remember where I saved something five minutes ago.
cpdown: Getting .md Content Out of Any Website
Discovered this open-source extension called cpdown that's been surprisingly handy. It extracts content from any website in markdown format, and it works with ChatGPT conversations too. One of those tools that does exactly what it says, just useful functionality. Sometimes that's exactly what you need.
The fact that it works with ChatGPT chats is particularly useful. Sometimes I'll have a productive conversation in ChatGPT that I want to capture as a structured note in Obsidian, and cpdown makes that seamless.
On Development Tools
I've been experimenting with Windsurf as my latest code editor, and it's been surprisingly solid. I got interested in it because under the hood you can pick fast Alibaba's Qwen3-Coder model, hosted on Cerebrus Inference Cloud. Hardware optimizations push this model to stream tokens at 2000 tps rate. The feeling of the speed if real when working with ~100 tps of Claude Sonnet. There's a 14-day free trial if you want to test it out.
What excites me is seeing open-source models reaching performance levels that genuinely compete with closed-source alternatives. Though I'm still hoping we'll eventually be able to run these powerful models through software and hardware optimizations locally rather than relying on hosted services in a couple of years.
I actually cancelled my Cursor subscription about three months ago and went full-time with Claude Code. The one thing I genuinely missed was proper tab completion. It's one of those features you don't appreciate until it's gone - especially when you're building custom slash commands or handling repetitive structured data. Having that predictive flow back in Windsurf has been more valuable than I expected.
If you're interested in more experiments like these, I document a lot of explorations on my YouTube channel. I'm always pushing myself to discover new patterns in how we think, capture, and connect ideas. Thanks for reading – drop a comment and let me know what's working in your own creative process!