**** Beware, this can burn tokens ****
I created a new, utterly experimental skill for (max) subscribers of Claude to use within Obsidian. It’s /deep-recon. Where Deep Research focuses in on one vector, Deep Recon goes out to look for new ideas both in your vault and on the net.
This is a multi-agent skill.
Four agents - Explorer, Associator, Critic, Synthesizer - run in parallel rounds. The Explorer casts a wide net across the web and your notes. The Associator finds structural parallels you hadn’t seen. The Critic stress-tests everything. The Synthesizer maps the territory: not one answer, but competing framings and the tensions between them. The work ok.
The output is an Obsidian-native document structured around 3-5 competing framings of your topic. The goal is to broaden perspectives about a topic.
Go to GitHub - kvarnelis/deep-recon: Multi-agent reconnaissance skill for Claude Code + Obsidian for the skill. Drop the deep-recon folder into claude/skills/ and invoke it with /deep-recon
It needs Opus model access - the Synthesizer does the hardest integrative work and needs the headroom. And, of course, it needs to have access to the net as well as an Obsidian instance with some version of the Claude code installed.
A two-round session runs roughly 800k-1M tokens across all agents and takes 20-30 minutes.
So far, I’ve been a little surprised by some of the inferences Claude has drawn, positively of course.