Software Architecture Blog
The Simplest Second Brain
Whatever your position is in the tech world, it's a key advantage to have a reliable system of record. Aka, a "second brain".
I use Obsidian, and I've probably tried at least half a dozen different community plugins based on LLM technology.
There's a sameness about them. Install the plugin, wrangle with an API key, then discover that it doesn't quite work the way I need it to. Then forget about it until it becomes an error message that pops up when you launch the app (especially on mobile).
There's got to be a better way. And I think I've found it…and it works without any plugins.
The key is using Anthropic's Cowork feature. This is something like Claude Code, but useful for all kinds of general purpose tasks. Including curating a large tree of Markdown files.
I created a top-level folder called Agents/ and in it I put one-file-per agent as well as a README.md. The files are simply natural language prompts. The README goes something like this:
# Agents
This folder contains instructions for Claude Cowork to periodically review and act on.
## How It Works
Each `.md` file in this folder defines an **agent** — a set of instructions that Claude reads and executes against the vault. Agents scan, curate, and report on the vault's health and connections.
Each run produces two outputs:
1. **A timestamped log file** in `Agents/logs/` (e.g., `2026-03-04_Orphan_Scanner.md`) — full findings for that run, one file per run.
2. **A one-line summary** written into the agent's own Working Notes section — a brief running record for quick reference.
This means full history is preserved in `logs/` without bloating the agent files themselves.
## Agent Types
- **Scanner**: Finds things (orphans, stale notes, missing metadata). Read-only.
- **Curator**: Suggests improvements (links, frontmatter, restructuring). Proposes changes but doesn't apply them without approval.
- **Reporter**: Produces periodic summaries and trend data.
## Agent Frontmatter Properties
Each agent file uses these frontmatter properties:
| Property | Meaning |
|----------|---------|
| `agent-type` | One of `scanner`, `curator`, or `reporter` |
| `schedule` | How often it should run: `daily`, `weekly`, or `monthly` |
| `last-run` | Date of the most recent run (`YYYY-MM-DD`). Updated by Claude after each run. |
| `status` | `active` = runs normally; `paused` = skip this agent |
The `Agents.base` file in this folder powers an Obsidian Bases dashboard showing all agents, their schedules, days since last run, and overdue status. Open it to see the overview.
## Running Agents
Ask Claude Cowork to "run the agents" or run a specific one by name (e.g., "run the Orphan Scanner"). Claude will:
1. Read the agent's Instructions section
2. Execute the described analysis against the vault
3. Write full findings to `Agents/logs/YYYY-MM-DD Agent Name.md`
4. Add a one-line summary to the agent's Working Notes section
5. Update the `last-run` frontmatter property
## Modifying Agents
These are just markdown files. Edit the Instructions section to change what an agent does. Add new `.md` files to create new agents — they'll appear automatically in the Bases dashboard.
The **Purpose** section is for your reference. Claude uses the **Instructions** section to determine what to do.
## Example Roster
| Agent | Type | Schedule | What It Does |
|-------|------|----------|-------------|
| Orphan Scanner | scanner | weekly | Finds disconnected notes |
| Link Suggester | curator | weekly | Proposes missing wikilinks |
| Staleness Auditor | scanner | monthly | Flags old untouched notes |
| Frontmatter Enforcer | curator | weekly | Checks metadata compliance |
| TODO Gardener | curator | weekly | Prunes and restructures TODOs |
| Vault Stats Reporter | reporter | monthly | Produces vault health snapshots |
| People Relationship Manager | curator | weekly | CRM-style contact follow-ups |
Pretty self-explanatory, but there's a few things worth calling out.
Still a WIP that I'm actively tweaking and optimizing.
As of this writing, Cowork is available only in the Mac desktop version of Claude. I expect this to expand rapidly.
This is not automated. You still need to manually start a cowork session in your vault directory and tell it to get to work. Since my initial development, Anthropic has added a scheduled task feature that looks promising.
Look for further discussion of this on the Architect+ Digest. Let me know if you try it!
"I put my name on everything I do"