# Tool Use Ep 67 - Claude Code and Obsidian with Artum Zuttov  **Channel:** Tool Use **URL:** https://www.youtube.com/watch?v=Wj-505ECkXU **Date processed:** 2026-01-08 --- ## Summary This episode features Artum Zuttov, a physicist PhD exploring coding agents for non-coding workflows. The conversation centers on how to use Claude Code with Obsidian to build a "second brain" that evolves into a powerful personal assistant. Artum shares his journey from Notion to Emacs to Obsidian, ultimately settling on a system where he primarily interacts with his vault through Cursor and Claude Code rather than the Obsidian interface directly. The core philosophy is pragmatic iteration over perfect system design. Rather than architecting an elaborate system upfront, Artum advocates for building minimal viable solutions that address immediate needs, then packaging successful workflows into reproducible custom slash commands. He has hundreds of these commands but only uses about 10 regularly—the usage patterns naturally reveal what's valuable. Key applications discussed include an automated morning routine (reviewing yesterday via parsed notes, answering reflection questions, reviewing goals, and generating a daily schedule), job search automation (where Claude discovers, analyzes, and filters job postings based on preferences), and session history parsing to extract voice notes and conversation summaries. The conversation emphasizes keeping Claude.md lean through file references and building parsers that inject only relevant context rather than reading full files. --- ## Key Points - **Iterate over architect**: Don't design a perfect system upfront. Build MVP solutions for immediate problems, see if they stick, then formalize into slash commands - **Custom slash commands are king**: Package valuable workflows into reproducible commands. Create meta-commands that help create other commands - **Context management matters**: Build parsers that extract only relevant portions (e.g., last 3 logs) rather than reading full files. Use templated formats for consistent parsing - **Claude.md with references**: Keep the main Claude.md lean by referencing other instruction files for specific tasks ("just in time instructions") - **Session history is accessible**: Claude Code stores all conversations locally—you can build parsers to extract voice notes, summaries, and patterns from past sessions - **80% context is good enough**: For personal assistant use cases (vs coding), getting context roughly right works fine. Claude's exploration capabilities fill gaps - **Voice transcription speed matters**: Tools like Whisper Flow with instant processing reduce friction compared to tools with processing delays - **Obsidian Bases for categorization**: Use categories (people, projects, commands) displayed in Obsidian Bases for visual organization --- ## Detailed Notes ### Journey to the Current Setup ([0:00 - 3:15]) Artum's evolution: Notion → Emacs (local files) → Notion again (for databases) → Obsidian with Claude Code. The key breakthrough was Claude Code's ability to run local commands and scripts without API restrictions. He now works primarily in Cursor with Obsidian vault open, rarely using the Obsidian interface directly. ### Philosophy: Build for Immediate Value ([3:15 - 7:00]) The anti-pattern is trying to design the perfect system upfront—this wastes energy without results. Instead: - Solve problems as they arise with MVP solutions - Package workflows that you do more than once - The criteria: **value + repetition = worth automating** - Make the process of workflow creation itself easy (meta-commands) ### Organization: Index Files and Bases ([9:00 - 11:30]) Uses wiki links through **index files** that collect related notes (e.g., all job applications linked from one index). Obsidian Bases display categorized items (people, projects, goals) in visual tables with metadata. ### Morning Routine Workflow ([11:30 - 16:00]) A structured, reproducible daily workflow: 1. Opens Spotify + Pomodoro timer via scripts 2. Reviews yesterday using parsed notes from a note-taking app (exported as JSON) 3. Answers reflection questions (energy levels, excitement) 4. Reviews goals (Claude pulls goal files, asks about one step toward each) 5. Proposes daily schedule based on working rhythm 6. Back-and-forth refinement, then exports to calendar The workflow is the same each day (same commands, same question templates), but context changes as goals progress. Randomized question databases add variety. ### Updating Workflows ([16:00 - 17:30]) Uses another custom slash command: "update custom slash command—introduce minimal change to address the issue discussed in this chat." The emphasis on **minimal changes** prevents the agent from going off-track. ### Job Search Automation ([17:50 - 22:00]) Claude Code's internal explorer agents fetch job posting details from websites. The workflow: 1. Provide job posting links 2. Sub-agents gather information from each 3. Creates files with metadata for each job 4. View in Obsidian Base with filtering 5. Claude adds comments on job fit based on general preferences 6. "I want more like this" refines future searches ### Claude.md and Context Management ([22:00 - 27:30]) Keep Claude.md minimal with references to task-specific instruction files. Example: "If adding a video idea, use this file for instructions." For reading multiple goal/log files that blow up context: - Standardize log format with consistent headers - Build a **parser** that extracts only last 3 logs from each file - Another script updates logs without reading the full file (faster, less token generation) ### Session History Parsing ([27:30 - 30:00]) Claude Code stores all conversations locally. Artum built a parser to: - List all sessions for a given day - Extract voice notes from specific sessions - Copy to clipboard for use elsewhere Uses Whisper Flow for voice transcription—instant processing is critical for maintaining flow. ### MCP Servers and Reliability ([30:00 - 38:00]) Experience with MCP servers: they add tooling to context which can misdirect Claude. Preference for local scripts over MCP when possible. For reliability: - Shell scripts over Python for simple tasks (more reliable with Claude) - Break complex scripts into smaller, composable pieces - Use `set -e` to fail fast - Validate inputs at script boundaries ### Cost Management ([38:00 - 42:00]) Uses Haiku for lower-stakes tasks (parsing, simple operations). Configures which model to use in slash command definitions. Morning routine mostly uses Haiku except for goal reflection which uses Sonnet. Monitors costs through the Claude Code dashboard. Typical spend: roughly $3/day with heavy usage. ### Sub-agents and Task Delegation ([42:00 - 47:00]) Uses general-purpose sub-agents for: - Web fetching (job postings, research) - File exploration - Parallel processing of multiple items The key is letting the agent decide how to accomplish goals rather than micromanaging steps. --- ## Connections *Highly relevant to my own Claude Code + Obsidian setup. Several ideas to explore:* - Session history parsing for voice notes - Parser approach for goal/log context management - Meta-commands for workflow creation - Haiku for routine tasks to manage costs --- ## My Notes *Space for your own thoughts, reactions, and connections after watching* --- ## Full Transcript <details> <summary>Click to expand transcript</summary> [0:00] Your notes can do so much more. Just [0:02] imagine your notes become your second [0:04] brain which then evolves into a powerful [0:06] assistant. But how do we do it? On [0:08] episode 67 tool use, we're joined by [0:10] Artum Zuttov. He's a physicist PhD [0:13] exploring coding agents for non-coding [0:15] workflows. And we're going to discuss [0:16] how you can use cloud code in Obsidian [0:19] to optimize your day, how you can build [0:21] helpful automations for your general [0:22] tasks, and how to make your system more [0:24] reliable and affordable. I love using [0:27] Obsidian with AI. I use it almost every [0:28] single day. But Artum brings it to a [0:30] whole new level. So please enjoy this [0:32] conversation with Artum Zutav. Why did [0:34] you choose Obsidian for your notes? [0:36] >> I've been uh trying to uh find a perfect [0:40] workflow for me. I've been doing this um [0:43] not taking for a very long time. I think [0:45] um or like five years ago I start using [0:48] notion [0:50] and it was like a big thing. I think uh [0:52] during the co everyone starts like uh I [0:55] need to now change something in my life [0:57] and now I'll start uh uh setting goals [1:00] taking notes and so on and uh I kind of [1:03] start uh from there from notion but then [1:07] I um I kind of um went on uh on track [1:12] about how can I store my notes locally. [1:15] I used Emacs. Uh yeah, if you know [1:17] Emacs, yeah, it allows you to kind of [1:20] work locally with your files. And um [1:23] then um Evan went back again on on [1:27] notion track. I think they used um they [1:30] created a very nice uh kind of way to um [1:36] like I think the the kind of the by [1:39] point for notion is databases [1:42] >> where you [1:44] can visually see and sort uh your items [1:49] and that's kind of uh something that I [1:51] never could achieve in other systems. I [1:54] started exploring uh coding agents and [1:57] um I I start uh storing my notes [2:00] locally. I think um maybe like this year [2:03] I start using obsidian once again. And [2:06] then uh I think during like in May uh [2:10] when the cloud code became a big thing [2:11] and everyone start using it for kind of [2:14] for for coding I I wanted to create sort [2:17] of operational system for myself on I [2:20] started like creating different scripts [2:22] uh uh for different automations and I've [2:25] been considering also different options [2:27] maybe there is like N8N maybe there is [2:29] something else and uh I think um the key [2:33] criteria for me was to be able to not to [2:37] click buttons. [2:40] So let the agent to manage um uh like [2:44] all of the scripts and I think um like [2:47] doing this locally uh where the agent [2:50] can run commands can execute scripts is [2:53] something that really kind of uh enables [2:56] me to to take like full um power of that [3:00] and not be being restricted by APIs [3:03] uh or that sort of stuff like I think I [3:06] I removed all the complexity and now I [3:08] kind of work locally. deleted all of my [3:10] notes with uh coding agents such as [3:13] cloud code. [3:14] >> Was there a uh like an adoption curve [3:17] for you when you started using a coding [3:19] agent in non-coding manners? Was it like [3:21] very intuitive or how did that process [3:24] get started? [3:25] >> You need to learn it for me. Um you need [3:28] to kind of uh I need to uh iterate on [3:31] that. And I think uh for me one of the [3:35] big problems which I had is trying to [3:37] create a perfect system to manage [3:40] everything. [3:41] uh and I think that's just uh something [3:43] that um you waste a lot of energy but [3:46] then you don't see result of the system [3:48] working because it's too complex and [3:50] what I end up doing I think is just uh [3:53] building uh something right now what uh [3:56] what brings immediate value [4:00] um and um I think um it uh kind of um [4:05] should not be sort of uh um in a way [4:09] that you try to design perfect system [4:11] but rather I was trying to uh like later [4:14] on I learned that I should be doing this [4:16] approach where I I need to solve the [4:18] problem and uh let me design some MP [4:20] solution for that and see if it sticks. [4:23] uh and uh it's very easy to see if if [4:26] something that you like or not. to just [4:28] like see if are you using it or not and [4:32] uh I think I kind of uh just start [4:35] building the system and over time it [4:37] grows and grows where you have uh like [4:41] kind of workflows emerging and uh I [4:44] think uh for me I'm a big fan of uh like [4:48] uh custom flesh commands like something [4:51] that you can package as an instructions [4:54] and make it uh reliable and [4:56] reproducible. [4:58] U and that's what I've been doing. I [5:00] kind of collect start collecting those [5:01] workflows which uh I found valuable and [5:06] I think right now uh in my Obsidian [5:08] world I have like hundreds of those [5:09] custom flesh commands but I use maybe [5:12] only uh 10 of them and uh [5:17] and that's you can kind of see like what [5:19] you use and what you not and um [5:22] uh I think the patterning are just [5:24] naturally [5:25] >> very cool. Before we get into the actual [5:26] like implementation details of how you [5:28] build these and orchestrate it and [5:30] whatnot, could you just kind of touch on [5:32] your thought process for how how do you [5:35] go about doing the the planning for a [5:38] workflow? I know a lot of people say, [5:40] "Oh, I want to apply automations, but I [5:42] just don't have a use case." And I [5:43] really feel like anyone that does any [5:45] type of knowledge work, any type of work [5:46] under computer, there are workflows that [5:48] can be automated, but how do you [5:49] identify it? And then how do you go from [5:51] like idea to actually putting it into [5:54] practice? Well, um I think I do this [5:58] step very gradually. Um that's something [6:02] that um if I see that something I'm [6:06] going to be doing um more than once, I [6:09] try to uh package it into a workflow and [6:12] uh I have a specific command where I can [6:16] execute it uh mid session. Um so in [6:19] cloud code you have uh those like chats [6:21] continuous chats and then you have a [6:24] custom f command and then I kind of tell [6:26] the agent okay let's uh package this [6:29] into a that's kind of let let's package [6:32] this um workflow like sort of this [6:35] interaction which you did into something [6:38] that is reproducible and next time you [6:41] can uh execute it without like one one [6:43] shot I think um if I see the value and I [6:47] want to capture that um state and make [6:50] it reproducible. I I I try to um package [6:55] it and if I do it more than once that [6:57] kind of the criteria for me the value [7:00] plus doing it more than once. One way [7:03] also to um adopt those workflows is to [7:07] make it easy of process like make the [7:10] process of those workflow creation is [7:13] easy for you. And that's example we have [7:16] another custom flesh command that kind [7:18] of okay we are midshot and let's say [7:21] okay I just run it and let's capture [7:23] that into a workflow. And in terms of [7:25] identifying, I think um that's something [7:28] that I want to do. I want to just uh [7:31] track uh like you know like do [7:32] experiment where like for a couple days [7:34] I I just record uh uh what I do [7:39] like what what what what actually I do [7:42] like what kind of work do I do and what [7:45] kind of actions does it require for me [7:48] and see are there any like uh patterns [7:51] uh that can emerge. Is there anything I [7:54] can make uh reproducible? And uh um I [7:58] think yeah, I've been exploring that and [8:01] u I prepared some stuff. Excellent. [8:03] Yeah, being uh cognizant of the work [8:05] that you're actually doing, making [8:07] notes, which I love that idea, just you [8:08] know, observing yourself and then trying [8:10] to convert that into a workflow. Um [8:13] yeah, do you want to start showing us [8:14] the Obsidian vault and what you've got [8:17] going on and some of your setups? [8:19] >> Yeah. Okay. So, let's do that. Um I [8:21] think for me it's a bit unconventional. [8:23] Uh right now I um I kind of uh in a way [8:28] don't interact much with obsidian. Um [8:31] and uh what I do instead I so I have [8:35] basically um cursor open and that's kind [8:38] of how I work nowadays. Yeah. And uh [8:42] basically that's um the chat right and [8:46] then you have all of that all of this uh [8:49] kind of um notes which you have here and [8:53] that's uh my obsidian vault uh opened in [8:57] cursor and then then I have a bunch of [9:00] folders and there are sort of some [9:03] structure but um I didn't try like to [9:08] have a perfect system it more or less uh [9:11] uh in a way like self self-organizing uh [9:15] um structure. [9:18] >> And sorry just on that note of of the [9:20] structure, do you do like wiki links to [9:23] connect different notes to each other? [9:24] Do you do tags? How how do you help the [9:26] system become aware of connected bits of [9:28] information? [9:29] >> Yeah. So for me I do index indexing like [9:33] for example uh I use index files and [9:36] here uh what I do is for example if [9:40] during the workflow I created a multiple [9:44] nodes like during the chat with the [9:46] cloud then I try to uh kind of connect [9:50] those into something that uh can be [9:53] packaged right and here's a I do index [9:56] index files where you have one file And [9:59] then inside of this file you have [10:01] different links such that you don't lose [10:04] uh like h how how those nodes are [10:08] related right [10:09] >> uh in this example right I've been doing [10:11] job applications and here like some [10:14] companies which I've been considering [10:16] and like those are in index files and um [10:21] that's the way I try to keep track of [10:23] what's going on and um another thing [10:27] which I do is um using obsidian basis is [10:31] to have um sort of um way to have to to [10:35] categorize your items and here I have [10:38] categories for different um items uh [10:42] which I have in my world. I try to [10:44] create a categories. Is it like uh uh [10:48] people? Is it like uh projects? is like [10:51] custom/ command and I think what you can [10:54] do with that once you have those [10:56] categories you can display them in a [10:58] obsidian basis and uh I think here is a [11:02] example [11:03] um yes for the uh like for the goal for [11:07] example right and uh like what you do [11:11] here I think uh so you have categories [11:14] uh goals [11:16] and this way you can display all of your [11:18] goals in in one place in a nice kind of [11:21] format that helps me to see uh like [11:25] what's going on really. [11:27] >> Yeah, I know. I think that's really cool [11:28] uh and a really good approach for being [11:31] able to keep the information connected. [11:33] So, you you set up your categories, you [11:35] set up different notes. How how do you [11:37] actually leverage cloud code to generate [11:39] value out of that information? [11:41] >> I think there are different ways. I I [11:44] try to use it as a capilot [11:48] where for example I try to like work uh [11:52] together with a cloud code and I think [11:54] uh one example uh is uh my morning [11:57] routine where um clo code guides me uh [12:02] throughout my morning process. I think [12:05] that's a very important thing that like [12:08] that um each day we make a decision on [12:11] like what to do right and I try to bring [12:14] a structure into this process and here [12:17] um example of this process where a cloud [12:21] guides me through my morning routine and [12:24] uh what I have here is my uh main file [12:28] which is executes this workflow and then [12:32] uh it opens my Spotify and And it opens [12:34] my Podora timer. Um, I can actually try [12:38] to run it. Uh, let me see. Yeah, it [12:41] needs to cache. Okay. Yeah, I don't know [12:44] what's going to happen, but let's see. [12:46] Okay, so it opens my Spotify app and it [12:49] opens um my Pora timer [12:54] and um you kind of start this morning [12:57] routine. So, sorry. I I'll pause it. [12:59] Right now, I have like music playing. [13:01] It's to pause it. Yeah. And then what [13:05] cloud does so it runs this command um [13:08] review yesterday and it reconstructs my [13:11] day um kind of uh using [13:15] the notes which I take throughout the [13:18] day and uh it kind of uh ables to [13:22] reconstruct uh like three days patterns [13:26] uh and the way you do that the way I [13:30] collect this data is using um this app [13:33] just like a test test do test. So here [13:36] you have notes and for example if I [13:39] start working on something I can just [13:40] record it and uh like my lessons like [13:43] learnings and then it's going to appear [13:46] here and then I at the end of the day I [13:49] export those notes uh into a JSON file [13:53] and then cloud just reads it and it able [13:56] to reconstruct uh uh like what happened [13:59] yesterday like uh hour by hour and um [14:02] like I do this over three case. [14:06] Now, uh what I do next uh is I answer [14:09] questions. [14:11] Each day I answer questions um like [14:13] about my energy levels uh uh what I'm [14:17] excited about today. And this is um all [14:21] of the context. Now, [14:24] next step, you review your your goals. [14:27] Here's the claude um pulls all of my [14:30] goals. [14:32] I think I showed those like think this [14:35] are my goals and claude uh reads those [14:38] files and then it asks me a question [14:41] like what what is the one step you can [14:43] do uh towards achieving that goal and [14:46] then you do this uh answering those [14:48] questions and in the end of the day you [14:52] you plan your your day and the clock [14:54] kind of proposes a schedule based on [14:57] your working rhythm and uh now here you [15:01] have a schedule and then finally you [15:04] kind of do back and forth and then uh [15:08] you put it into your calendar. I think [15:10] that's one way how you can like create a [15:14] structured workflows and um create value [15:18] because uh I think that's um something [15:21] that I've been doing for quite a long [15:23] time and uh it's been really helpful [15:26] like gathering context and understanding [15:28] like what's going on. [... transcript continues - truncated for display ...] </details>