Your notes already contain the patterns. You just need something that can read all of them at once. (Credit: GPT Images v2)
Your AI second brain is set up. Here’s what to actually do with it. Number 4 is my favourite.
If you followed the setup from Part 1, then you’ve got an Obsidian vault with Claude Code pointed at it, a CLAUDE.md telling the agent how to behave, and maybe a /context or /weekly command to get started.
Great! You’ve set up the infrastructure.
Now, this article is about what to build on top of it.
Because the setup alone doesn’t do much. A vault with 20 notes and no links is just…a fancy folder. But once you’ve been writing daily notes for a few weeks, linking them with [[wikilinks]], and letting context accumulate, a different category of workflow becomes possible.
These are the ones worth building.
1. Trace how an idea evolved
Picture this: you’ve been writing about something for months. Maybe it’s a technical approach at work, a side project that keeps changing shape, or some recurring frustration. The notes are scattered across dozens of daily logs. You couldn’t reconstruct the timeline yourself without hours of scrolling.
/trace does it in minutes.
Create .claude/commands/trace.md:
---
description: Track how a specific idea evolved across the vault
argument-hint: <topic to trace>
---
Search the entire vault for notes mentioning or related to: $ARGUMENTS
Build a chronological timeline of how this idea developed:
- When it first appeared
- How the language and framing shifted over time
- Key turning points (when I changed my mind, added nuance, or dropped it)
- What other topics it connects to through the link graph
- Where it stands now based on the most recent mentions
Quote directly from my notes where relevant. I want to see my own
words, not your paraphrasing.
Do not create or edit any files.
Run it: /trace remote work, /trace the API migration project, /trace my gym habit
Vin Vashishta ran this live on the Startup Ideas Podcast and traced his relationship with Obsidian across 13 months of notes. The output showed four distinct phases: skepticism, adoption, heavy building, and a shift toward delegation.
He said he wouldn’t have been able to see those phases himself. As humans, we’re bad at noticing our own gradual shifts in thinking. Fortunately, the vault doesn’t forget what you were saying in January.
2. Find connections between unrelated topics
This one depends heavily on your link graph, which is why backlinking matters so much. /connect takes two topics you think are unrelated and walks the graph between them looking for shared nodes, overlapping references, and structural parallels.
---
description: Find hidden connections between two domains
argument-hint: <topic A> and <topic B>
---
I want to connect two domains: $ARGUMENTS
Read notes related to each domain. Use the link graph to find:
- Shared nodes (notes that both domains link to or through)
- Structural parallels (similar patterns in different contexts)
- People, tools, or concepts that appear in both domains
Build 2-3 "bridges" between these domains. Each bridge should
reference specific notes and explain the connection in a way
that's genuinely non-obvious, not just surface-level keyword overlap.
Do not create or edit any files.
Run it: /connect data engineering and coffee brewing, /connect the hiring pipeline and the onboarding project
Vashishta connected filmmaking and world-building in his demo and got back bridges he described as genuinely surprising. The output referenced specific notes he’d written months apart, in completely different contexts, that shared a structural idea he hadn’t named. That’s the kind of thing that only works when the agent can see hundreds of notes at once and walk the links between them.
3. Surface ideas you never explicitly stated
I’d say this is the weirdest command but arguably the most valuable. /emerge scans your vault for conclusions that your notes imply but you never actually wrote down. Recurring themes that don't have their own note yet. Patterns sitting across five different daily logs that you haven't connected.
Create .claude/commands/emerge.md:
---
description: Surface ideas the vault implies but never states
---
Scan daily notes and wiki pages from the past 60 days.
Look for:
- Recurring themes that appear 3+ times but don't have a
standalone note
- Conclusions my notes point toward without stating directly
- Unnamed patterns across different projects or life areas
- Questions I keep circling without answering
For each finding, show the evidence: which notes, which phrases,
how many times. Label your confidence (strong signal vs. weak hint).
Separate what my notes actually say from what you're inferring.
I need to see the difference clearly.
Do not create or edit any files.
Run it: /emerge (no arguments needed — it scans everything)
Just a heads up: this only becomes useful after several weeks of notes. With five daily logs it’ll produce nothing interesting. With sixty, it starts catching things like: “ You’ve mentioned latency problems in three different projects this month but never connected them. You might have a general architectural preference you haven’t articulated.”
4. Compare what you said you’d do with what you actually did
/drift is the accountability command. It reads your stated intentions (goals, plans, priorities from earlier notes) and compares them against what you actually spent time on based on more recent daily notes.
Create .claude/commands/drift.md:
---
description: Compare stated intentions vs actual behavior
argument-hint: [optional: time window, defaults to 30 days]
---
Read daily notes from the past 30 days (or $ARGUMENTS if specified).
Find notes where I stated intentions, goals, priorities, or plans.
Then look at what I actually wrote about doing day-to-day.
Identify:
- Things I said I'd focus on but barely mentioned again
- Things I never planned to focus on but kept returning to
- Patterns in what I avoid (topics that get planned but never executed)
Be blunt. The whole point of this command is to catch self-deception.
Do not create or edit any files.
Run it: /drift for the default 30-day window, or /drift 60 days to go wider
This one can sting (or well, depending on how brutal you want Claude to be). It’s also the command that several people in the community describe as the most practically useful, because the gap between stated priorities and actual behavior is where most of the wasted time hides.
5. Pressure-test what you currently believe
/challenge picks a belief or position you've expressed in the vault and tries to dismantle it using your own writing. Not generic counterarguments from the internet. Evidence from your own notes that contradicts, qualifies, or undermines what you think right now.
Create .claude/commands/challenge.md:
---
description: Pressure-test a current belief using vault history
argument-hint: <belief or position to challenge>
---
I believe: $ARGUMENTS
Search the vault for:
- Past notes where I expressed doubt about this, or leaned
a different direction
- Evidence or experiences I recorded that contradict this belief
- Times I changed my mind on this or related topics
- Assumptions baked into this belief that my notes don't support
Build the strongest case against my current position, using
only material from the vault. If the vault doesn't contain
counterevidence, say so honestly rather than fabricating it.
Do not create or edit any files.
Run it: /challenge we should migrate to microservices, /challenge remote work makes me more productive, /challenge buying property right now is the right call
Eugeniu Ghelbur’s framing of this one is good: “Your vault holds you accountable.” If you wrote something six months ago that directly contradicts what you believe today, and you never noticed, that’s worth knowing.
6. Generate ideas grounded in your actual context
/ideas is the command that went viral from Vashishta's demo. It does a deep scan of the vault and produces suggestions that aren't generic productivity advice. They reference your projects, your people, your patterns, your gaps.
Create .claude/commands/ideas.md:
---
description: Deep vault scan for actionable ideas
---
Read CLAUDE.md, context files, all daily notes from the past 30 days,
all active project pages, and open loops.
Also check for orphan notes (notes with no inbound or outbound links)
and dead-end notes (linked to but empty or minimal).
Generate ideas across categories:
- Tools or workflows I should build
- Connections between projects I'm not exploiting
- Things I keep writing about but haven't acted on
- Conversations I should have (with specific people, if named in vault)
- Stale projects that should be killed or revived
Ground every suggestion in specific vault evidence.
No generic advice. If it could apply to anyone, don't include it.
Do not create or edit any files.
Run it: /ideas (no arguments — it reads broadly and generates from everything)
In the live demo, this command ran for about five minutes against months of notes and produced a report suggesting specific people to meet, tools to build, essays to write, and projects to kill. Greg Eisenberg compared the experience to a coaching session.
The key is the last line of the skill: “ If it could apply to anyone, don’t include it.” That’s what separates vault-grounded ideation from generic brainstorming.
7. Ingest external sources into your knowledge base
This one comes from Andrej Karpathy’s “LLM Wiki” pattern, which went massively viral in April 2026. Instead of dumping articles and transcripts into a folder, you have Claude read the source and update existing wiki pages with what it learned. People mentioned in the source get updated. Stale claims get revised. New topics get their own pages.
Create .claude/commands/ingest.md:
---
description: Process a source into the wiki knowledge base
argument-hint: <path to file or URL>
---
Read the source: $ARGUMENTS
Then read the current state of relevant wiki/ pages.
Update the wiki to reflect what this source adds:
- Update existing pages with new information
- Create new pages for topics that don't have one yet
- Flag contradictions between this source and existing wiki content
using [!contradiction] callouts
- Add the source to a "Sources" section at the bottom of each
touched page
Show me every proposed change before writing. I approve each one
individually.
Preserve all [[wikilinks]]. Use Obsidian-flavored markdown.
Run it: /ingest raw/sources/karpathy-llm-wiki.md, /ingest https://random-url.com
The power move from AgriciDaniel’s claude-obsidian project: point any other Claude Code project at your vault by referencing its path in that project's CLAUDE.md. Now your coding projects, your writing workflows, and your daily reviews all draw from the same knowledge base. One source ingested into the vault propagates everywhere.
Building your own commands
You’ll notice every command above follows the same skeleton: read specific files, apply a specific lens, produce structured output, don’t touch anything without permission. Once you see the pattern, writing a new command takes about two minutes.
The better approach is to let commands emerge from practice. Run a workflow manually in a Claude Code session. If it produces something useful, ask Claude to write it up as a reusable command and save it to .claude/commands/. After a couple of months, you'll have a library that grew from how you actually think rather than from someone else's template.
Before you start running these commands…
Every command in this article gets more useful over time and most of them are nearly useless in the first week. /trace needs months of notes to find an arc. /drift needs 30 days minimum. /emerge needs enough volume that patterns can actually form. /connect needs a link graph with enough edges to traverse.
So, this is the honest part: the setup from Part 1 takes an evening.
Getting real value from these commands takes weeks of writing daily notes, backlinking them, and building up enough context that the agent has something meaningful to work with. There’s no shortcut past the accumulation phase.
But the accumulation phase is just writing. Writing about what you’re working on, what you’re thinking, what you decided and why. The agent handles the rest: the pattern detection, the cross-referencing, the accountability checks, the idea generation.
Your job is to keep feeding it honest notes. Its job is to read all of them at once, which is the one thing you can’t do yourself.
Disclaimer: AI was used to assist with the research and structural refinement of this article.