# Get ahead of 99% of Claude Code users  **Channel:** Artem Zhutov **URL:** https://www.youtube.com/watch?v=G9S5DgmNBaM&list=PLKbj3lGwh8C3j7xoSVXa-fWB9rhoXZWiJ **Date processed:** 2026-01-16 --- ## Summary This video presents advanced patterns for power users of Claude Code who want to move beyond basic usage. Artem Zhutov, who has spent 800+ hours using Claude Code, shares his methodology for creating feedback loops that allow the agent to self-improve based on analyzing past conversations. He emphasizes that simply adding more tools and skills isn't the path to productivity—most users with 35+ skills only regularly use about 5. The core thesis is that context management is the highest-leverage activity for Claude Code users. This includes understanding how Claude Code works internally (system prompts, tool definitions, context windows), designing efficient skill architectures, and integrating personal context (notes, tasks, goals) into the agent's workspace. The video makes a provocative claim that "MCP is dead and CLI is the future," arguing that MCP tool definitions consume too much of the context window (14%+ for common integrations like Notion and Linear), whereas CLI tools loaded on-demand are more context-efficient. A key insight is treating markdown files not as static documents but as executable context—using Obsidian as both a dashboard for observability and a repository of actionable information that Claude can read and act upon. --- ## Key Points - **Feedback loops for self-improvement**: Analyze past Claude Code conversations (stored as JSON files) to identify friction patterns and automatically improve CLAUDE.md memory - **Context is finite**: The 200k token context window degrades in performance above ~50% usage; every token matters - **MCP vs CLI tools**: MCP tool definitions consume 13-14% of context upfront; CLI tools loaded on-demand via skills are more efficient - **Progressive disclosure**: Don't load all tool definitions upfront—load minimal skill descriptions and expand only when needed - **Skills as routing hubs**: Use skills not just as single-purpose tools but as routers to different workflows (morning routine, weekly review, etc.) - **Obsidian as dashboard**: Use Obsidian Bases for observability—track skill status, manage tasks, review goals - **User comments pattern**: Leave inline comments in code for Claude to find and address rather than explaining everything in chat - **Handoff prompts**: When running out of context, document learnings to a file in Obsidian for the next session to continue from - **Verbose output**: Enable verbose output in settings to understand what the agent is actually reading and doing --- ## Detailed Notes ### Feedback Loops and Session Analysis [0:00-5:00] The differentiator for advanced users is designing feedback loops where the agent can self-improve. All Claude Code conversations are stored as JSON files on your computer. Artem built a "session analysis" skill that: - Locates conversation files in the cache folders - Extracts user messages and searches for friction indicators (keywords like "wrong") - Generates reports showing exact messages and proposed improvements to CLAUDE.md - Identifies repeating patterns that cause frustration He also uses WhisperFlow (voice-to-text app with 1M+ words dictated) and built a skill to analyze dictation patterns across apps and hours. ### Understanding Claude Code Internals [5:00-10:00] **Enable verbose output** (in settings) to see: - Exact prompts sent to the agent - What files are being read - Thinking traces **Claude Code architecture:** - System prompt (~15k tokens) with Anthropic's instructions - Tool definitions (write, edit, read, bash, skill tool) - Skill descriptions loaded into the skill tool - System reminders (todo list status, Obsidian context, open files) - CLAUDE.md memory files - User messages → thinking → responses **Context tool** shows current token usage and remaining space. ### Git Review and User Comments [10:00-12:30] **Always review changes in git** because the agent doesn't always do what you asked. Use a git diff viewer (like in Zed) to see exactly what changed. **User comments pattern**: Instead of explaining in chat, leave inline comments in code using text expansion (e.g., `// USER: get these programmatically, don't hardcode`). Then tell Claude to look for and address user comments. ### Obsidian as Dashboard [12:30-14:30] Use [[Obsidian]] Bases to create dashboards for skills management: - Track skill status (active, archived, not used) - Filter and group skills - Add inline comments in skill files for Claude to parse ### Skill Architecture Patterns [14:30-19:30] Most users have many skills but only use a few. Curate ruthlessly. **Pattern 1: Targeted tool** - Single-purpose skills (e.g., Excalidraw for diagrams) **Pattern 2: Router hub** - Skills that route to different workflows: ``` review/ ├── SKILL.md (router) ├── workflows/ │ ├── morning.md │ ├── evening.md │ ├── weekly.md │ └── monthly.md ``` ### Why MCP is Context-Inefficient [16:30-19:30] MCP (Model Context Protocol) loads all tool definitions upfront. Testing with Notion + Linear MCPs showed **13-14% context consumption** before any work begins. **Solution: CLI tools via skills** - Small description in skill definition (minimal context) - Full instructions loaded only when skill is invoked - On-demand loading = progressive disclosure Example: Gmail skill loads only when needed, not consuming context until invoked. ### Integrating Life Context [19:30-24:00] Store all context (notes, tasks, goals) in a single folder (Obsidian vault) that Claude can access. "English is now a programming language" - notes are no longer static but executable context. **Workflow example:** 1. Go for a walk with wireless mic 2. Record brain dump 3. Transcribe with Gemini Flash 4. Claude processes transcript with full context of goals/tasks 5. Execute action items directly in Claude Code Tasks and goals stored in Obsidian Bases can be read by Claude and acted upon. ### Handoff Prompts and Session Continuity [23:00-26:00] When approaching context limits: 1. Create custom handoff command 2. Document current progress and learnings to an Obsidian file 3. Tag the file for future reference 4. Start new session, tag the file, continue with full context ### Getting Started [25:40-end] 1. Turn on verbose output 2. Set up git review for all changes 3. Back up everything 4. Try user comment pattern 5. Create your own skills (high leverage) 6. Analyze sessions weekly **Recommended follows:** Seb Obsidian, Simon Willison, Mario (claude-trace tool creator) --- ## Connections - [[Claude Code]] - the tool being discussed - [[Obsidian]] - used as dashboard and context store - [[MCP]] - Model Context Protocol, criticized as context-inefficient - [[Progressive Disclosure]] - UX pattern applied to agent tools --- ## My Notes *Space for your own thoughts, reactions, and connections after watching* --- ## Full Transcript <details> <summary>Click to expand transcript</summary> [0:00] Everyone is downloading skills, [0:01] installing MCP service, following [0:03] tutorials, more tools, more power, [0:06] right? And that's not how it works. Here [0:08] I have 35 skills and maybe I use only [0:12] five of them. And I've spent 800 hours [0:15] in this thing. I literally live in [0:18] terminal in cloud code and I right now [0:20] analyze my own conversations to find [0:23] where I get frustrated and I feed it [0:25] back to my agent to improve my cloud. MD [0:28] automatically. And here how it looks [0:30] like. And by the end of the video, [0:33] you're going to know exactly what [0:34] separates 1% from everyone else. Let's [0:37] go. Here is a brief outline. What I'm [0:39] going to cover is how to design those [0:42] feedback loops for our agent to [0:44] self-improve our memory. How to improve [0:47] observability of what agent is doing and [0:50] how cloud code works internally. Next, [0:52] we're going to cover how to manage your [0:55] skills. What are the different patterns [0:57] of working with skills? How to curate [1:00] the list of your skills. We're going to [1:01] also cover why MCP is dead and CLI is [1:04] the future. Finally, the most exciting [1:06] part for me is how to integrate all of [1:09] the context in your life, your notes, [1:12] your tasks into cloud code such that you [1:14] can be more efficient. I'm going to show [1:16] how you can just go for a walk, record [1:18] yourself, and then go from a brain dump [1:22] to something that is executable within [1:24] clock. So let's get started. First thing [1:27] is analyzing our feedback loops. How can [1:30] we analyze our past conversations to [1:32] improve our memory for the agent such [1:35] that we can avoid those repetitions [1:38] where we tell the agent do this thing [1:40] and where we can just make the agent [1:43] remember how to behave next time. So the [1:45] key thing here to realize is that all of [1:49] your cloud code conversations are stored [1:51] as a JSON files on your computer. You [1:54] can ask clot to analyze them and I build [1:57] this skill called session analysis to [2:00] analyze cloud code conversations. So [2:03] here you have a location with all of [2:05] your cloud code projects. We have a [2:07] cache conversations stored in these [2:09] folders. That's example of where my [2:12] vault lives. We tell the agent how the [2:15] structure of our messages looks like. [2:17] How to extract today's user messages. [2:19] How to get some friction in indicators [2:21] by searching for the keywords. Then we [2:24] have workflows to analyze the friction. [2:28] Once you execute the skill, once you [2:30] analyze your conversation, your agent [2:32] can come up with a report to you saying [2:35] your exact message which you provided to [2:38] the agent and propose an improvement. [2:42] And you can clearly see what are the [2:44] repeating patterns that happen over and [2:46] over. and you can think together with [2:48] the agent how to address those. That's a [2:52] very very high leverage action which you [2:54] can do now to improve your just overall [2:58] experience working with the agents. If [3:01] you're using voicetoext apps such as [3:03] whisper flow here is mine. Uh I have uh [3:08] already like 1 million words dictated. [3:12] It's really amazing uh app. It just [3:15] gives me so much leverage. So I don't [3:17] have to type and I build a cloud code [3:20] skill uh which helped me to analyze my [3:23] conversations [3:24] across different apps. Here's example of [3:27] analysis which cla did for me. It uh [3:31] counted the number of words I dictated [3:33] per day and it analyzes them by by the [3:37] app. how many words I dictate during the [3:41] different hours and that's one thing we [3:43] can do this analysis and I just run it [3:47] just now to see how it works u I build a [3:50] skill it's it's called whisper flow and [3:53] I just need to analyze my points of [3:55] frustration working with cloth and [3:57] suggest improvements to my clot MD there [3:59] like keywords like wrong and so on and [4:02] let me see the response of of different [4:05] patterns like whisper flow frustration [4:07] analysis this and for example ignoring [4:10] my instructions. Uh that's a common [4:14] pattern. Um uh repeating mistakes um [4:17] hardcoding values and we can like ask [4:20] lot open port and let us see [4:25] and uh we can see it in more details and [4:27] that's uh very useful information which [4:30] you can use to design a better uh memory [4:34] for cloud. Um here cloud is already [4:37] suggesting add three missing rules to [4:40] cloud.md [4:42] that's an example of feedback loop which [4:44] you can engage with. So just like [4:48] understanding this concept that all of [4:50] your data all your conversations is here [4:53] on your computer and you can access it [4:55] and use it uh for improving of how you [4:59] work and how you behave. So next uh [5:02] we're going to cover um how can we [5:04] better understand what agent does. For [5:07] example, if you understand how agent is [5:10] working internally, what goes into the [5:12] context, we can have a better handle of [5:14] this tool. We can really master this [5:17] tool, understand its limitations, why is [5:19] it failing and what are the different [5:21] capabilities which are available to us. [5:24] Now the first thing which I suggest to [5:27] do is to use a verbos output. Let's [5:31] start our new conversation and we have a [5:33] settings [5:35] in the config and we can go to [5:38] uh just search for varos output and we [5:41] can turn it on. Right now it's on. We [5:44] can turn it off and what it does it [5:47] helps me to understand what content is [5:51] read by the agent. Now we can compare [5:53] how it behaves side by side. Here on the [5:56] left you have verbos uh output on and [6:00] here on the right you have uh it to be [6:02] off. Can you um explain the structure of [6:06] the subsidy world? I'm just going to [6:08] give the same tasks to those two agents [6:12] how how it's uh looking. We we can [6:14] already see that here we see the [6:17] thinking trace right but here it just [6:20] gives me um acknowledgement I can see [6:24] exact prompt uh here which it tells to [6:27] the agent to do I can see what it reads [6:31] uh and so on and here it just looks like [6:34] a black box that's just one example [6:37] right and by learning how it works under [6:40] the hood you can u have a better [6:43] understanding meaning what's possible, [6:45] what are the limitation here. Next, it's [6:48] it's very important to understand how [6:50] cloud code is actually working and here [6:52] is a diagram. We have a a system prompt [6:56] in cloud code which which sort of memory [6:58] for cloud and we have a instructions for [7:02] clot um prescribed by entropic uh have [7:06] some useful commands what is the current [7:08] date and so on. Then you have a tools [7:11] definitions uh which code has access to [7:13] with the descriptions name parameters [7:16] and how to use them. You have a write, [7:20] edit, read, bash tool. We have a uh [7:23] skill tool and within this skill tool we [7:27] have a loaded skills their descriptions [7:30] and then what happens is we have our [7:33] conversations with the agent and context [7:35] grow over time. uh agent does some tool [7:38] calls. It gives us responses. Now how [7:41] this looks in practice here uh is exact [7:46] system prompt. We can get it from cloud [7:48] code and we can see what actually [7:51] entropic puts inside of the clo code to [7:55] behaves the way it behaves. uh URL cloud [7:57] code andropic official CLI tool here's [8:00] the tone and the style and those ones [8:03] are very useful to read and understand [8:06] what's going on here we can't change [8:09] those because it's like system prompt [8:13] um but we can understand and better [8:16] understanding would give us better um [8:18] handling of this tool system prompt it's [8:21] um quite lean I would say it shows uh [8:25] also the status of the current files [8:27] which are opened. Um and then I think [8:30] it's very very interesting to see the [8:33] tool definitions. You have a task tool, [8:36] task output, uh bash tool, uh globe [8:40] tool, grap for search. But I think the [8:43] most uh like interesting is is skill [8:46] tool. Uh how exe how can we execute [8:49] skills within the conversation? [8:52] And here is a list of available skills [8:54] which Clot sees and has within its [8:57] memory. And those are exactly the same [9:00] which I have here displayed in my [9:02] obsidian a skill name and description. [9:04] And CL can read and and uh engage in [9:08] those on demand. And finally we have a [9:11] system reminder that uh your to-do list [9:14] is empty. We have a system reminder [9:16] about obsidian context. What I I'm [9:19] selecting here. For example, what page [9:22] do I open in my Obsidian? And and then [9:24] what comes next is the memory this [9:26] cloud. MMD and here I have uh in in my [9:29] memory um like how to create events [9:32] using Apple script. Next I have for [9:35] example who I am my like core values um [9:39] it's a personal assistant for my [9:41] obsidian vault. Finally after all this [9:44] uh memory files we have our user [9:47] message. For example, I just say hi. And [9:50] here we see a thinking phrase and claude [9:53] gets a response. I'm sending this like [9:55] another uh idea uh like to check what's [9:59] in my memory and then um this is the [10:02] response. Uh it it executes a read tool, [10:06] executes uh the output of the read tool [10:10] and there's a thinking trace and then as [10:13] to clean up it file just make it empty. [10:16] just uh cleaned it up and that how this [10:18] tool works internally. Here we are in [10:20] conversation right and we are uh we are [10:22] engaging uh in conversation agent does [10:26] tool calls uh it gets responses. Now [10:28] once you understand how cloud works and [10:31] um that you can have a conversation one [10:33] goes into the system promp to tools [10:35] definitions we start thinking that um [10:38] context is is very important. we have a [10:40] finite uh memory uh finite conversation [10:44] which we can have with cloth until it [10:47] starts underperforming. And here's a [10:50] context tool which enables us to see how [10:53] much um tokens which we have right now [10:56] the free space is 160,000 tokens and [10:59] those ones we can use in in this [11:01] conversation for example and we'll go [11:03] more on how to manage your uh context um [11:07] in the next sections. Another pattern [11:10] which I want to highlight is um review [11:13] every changes in g um because agent [11:16] doesn't always do what I asked and it [11:20] always gets confused and one of the [11:22] patterns which I use uh to work with [11:25] agents use uh a user comments. Now let [11:29] me just open it in Z and how it looks [11:31] like. Instead I have a different um [11:34] uncommitted files and I can look here on [11:38] what files have been changed and uh I [11:41] can actually see what agent introduced [11:44] here is a new content produced by the [11:46] agent [11:48] um here is what's been deleted here is [11:49] what's been changed and that's uh so [11:51] much value because now I clearly [11:53] understand what changes has been made [11:57] and then to iterate on those changes [11:58] what I do is I can leave some comments [12:02] I use those text expansion which enables [12:04] me to uh quickly create this user [12:07] comment here how it looks like uh I [12:09] might uh for example kindly get uh those [12:12] people programmatically right I don't [12:15] want to hardcode that's uh for my uh [12:18] Twitter skill where I um search for [12:21] different tweets from different creators [12:23] what do they tell and here cl just [12:25] hardcoded hardcoded the names uh whereas [12:29] we have this uh JSON file from which you [12:31] can read and then I told uh to look for [12:35] those comments and actually um what it [12:39] did it f it found it and then it just [12:41] wrote uh a few commands to extract those [12:44] programmatically. This is not this is [12:46] some example how can you apply in real [12:48] use cases. Next is using obsidian as a [12:52] dashboard. Obsidian is u is very useful [12:55] for observability. Here is example of [12:58] all of my skills and Obsidian has this [13:01] uh uh basis which enable you to sort [13:06] um you can like group um and present [13:10] information in efficient way. It's it's [13:12] like a notion but locally and this [13:16] becomes very powerful because now I can [13:18] have a clear understanding what is the [13:21] status of those uh skills. I can have [13:24] like status active archived right. We [13:27] can add the status um here is it like um [13:32] okay active for example um this one is [13:35] um not used right and we can filter by [13:40] those and clearly understand um our [13:42] usage patterns. What I also do is um I [13:45] add some comments inside of my skills [13:48] such that the agent can can parse those [13:51] and address them. I can tell okay I can [13:54] just copy those and uh start a new [13:57] session uh can you um [14:01] can can you find comments here and uh [14:05] and let's think how to address those. [14:07] Here we have u comments right and we can [14:11] directly work with clo to figure out how [14:13] to um or how to organize those skills in [14:17] the battle ruin and that's one example [14:19] of using obsidian and there are many [14:21] many many more use cases for it which [14:23] we'll cover later in the video. Next is [14:27] a skill architecture. Now we have this [14:29] dboard with the skills and we want to [14:32] really curate them and um we want to use [14:36] them and figure out uh which one of them [14:39] are working for us and which one aren't [14:42] don't work for us. what reality is that [14:44] you might have 50 skills and yeah you [14:48] feel like okay I'm very productive now I [14:50] have like 50 skills but in reality you [14:53] just use a couple of them and all of the [14:56] rest is just noise let's move on to [14:59] different usage of patterns so as I told [15:02] you that uh what skills are uh they have [15:05] just a name skills is just a maron file [15:08] uh in a folder that's it uh there is no [15:11] need to over complicate it really just a [15:13] markdown file in a folder has a [15:15] description. [15:17] Here are a few different patterns on how [15:20] can we use skill. One of one of the [15:22] usage of skill is using it as a targeted [15:24] tool. Here's example I use excal skill [15:28] when I want to create this diagram for [15:31] my presentations. They are purely [15:33] created by claude using this exact skill [15:36] and the only job of the skill is to [15:38] create diagrams. Another pattern is to [15:40] use a skills as a router hub. I have a [15:45] skill for my uh daily review and when I [15:48] tell um morning routine [15:51] it has a different routers to to my [15:54] different workflows. Cloud understand [15:56] that it needs to run this morning [15:58] routine and we are going through my [15:59] morning routine together with clot. [16:01] Similarly for the weekly review, if you [16:03] want to do weekly review, you're going [16:05] to use the same skill [16:08] and this how it looks. We have this [16:10] folder with the workflows and we have a [16:13] separate markdown files for each of the [16:15] workflows. Uh this is example that's uh [16:18] we have a workflows and we have a skill [16:21] file here in this review skill. You can [16:23] see that it's a routing pattern have a [16:27] morning, evening, um, weekly, monthly [16:30] and we have a template as well. So the [16:32] previous issue which we had is to how [16:35] can we connect cloud code to external [16:37] sources and the solution was MCP. Let's [16:40] just use notion, let's just use linear [16:43] MCP to connect to all of the tools. But [16:46] people realize that it just not context [16:49] efficient. uh what we have is uh 2,000 [16:52] tokens and it's very very important how [16:55] do we fill those what kind of what do we [16:58] put into the context of cloth we have [17:01] this um like system prompt uh it has [17:06] approximately 15,000 tokens right with [17:08] all the definitions and all the tools [17:10] and then what we do is we load al our [17:15] MCP tools definitions and here I just [17:18] tested it with a notion and linear your [17:20] MCPS and they fill up 14% of our [17:23] conversation which we could have spent [17:25] on interacting with the agent and this [17:27] is important because the performance of [17:30] a clot uh degrades after you just after [17:35] you are here after you're above uh uh [17:38] 50% right you don't have much uh space [17:41] to to do some useful work and uh the [17:44] solution is using C light tools so here [17:47] is example I just tested it with [17:51] linear MCP server all of this tone of [17:54] the definitions and as well as the [17:56] notion one you can see here the MCP [17:59] tools is 13 14% of of our of our context [18:05] window [18:06] they are here and solution to that is to [18:09] use CI tools and CI tools is um [18:12] integrated with with uh skills for [18:15] example you want to connect your agent [18:17] to a Gmail mail right so previously what [18:20] you would do is uh okay we need a Gmail [18:24] MCP or something like that now what we [18:27] do is uh we have a skill for that and [18:31] here is example I just tell okay okay I [18:33] want to um test I just want to test this [18:37] uh uh Gmail how can I draft the email to [18:39] to myself all right let's see what agent [18:42] does uh it it loads the skill [18:45] automatically it even like doesn't need [18:47] to to be explained and uh it loads it on [18:51] demand right just by understanding the [18:56] descriptions which we put into the [19:00] system prompt for the skill. Yeah. So [19:02] let's uh let's test this draft. Yeah. [19:05] And open it uh in my browser. Okay. [19:08] Let's uh let's try to test. Uh now we [19:11] created this draft and um hopefully [19:13] Claude can figure out how to open it. [19:16] Great. And uh there you go. Um here we [19:19] can just send emails. How easy it is. [19:22] Now we don't need any MCP for that. Just [19:25] using CLI tools which will on demand is [19:28] the way to go. That's the current [19:29] solution to this um problem. And as I [19:32] told you there's a concept of this [19:34] progressive disclosure which we [19:36] partially covered. We you don't load all [19:38] the definition up front. just only load [19:41] the small small description of what [19:44] skill does and load only what's needed [19:47] when it's needed. The next part which I [19:49] want to cover is how can we use clot [19:51] code not only for coding but for just [19:54] managing what's going on in our life. So [19:56] when we are storing uh the information, [20:00] the context, the notes about ourself, it [20:03] gives us leverage. Typically what people [20:06] do to have a tasks in one app like right [20:09] you use notion, you use uh Apple notes [20:12] um like everything is just scattered. [20:14] But now when you put all of your context [20:18] in your uh folder uh I just have a [20:21] single folder on my computer and uh here [20:24] how it looks like. That's my Obsidian. I [20:26] have all these nodes which I've been [20:28] doing for the last uh couple years so [20:31] far on a daily basis trying to trying to [20:35] build a system and all this my nodes are [20:38] here and I can just use them and execute [20:42] them right because with a cloud code [20:45] like everything right now becomes an ex [20:47] executable English right now is an [20:50] programming language that's uh the tweet [20:53] from Kipano that the CEO of Obsidian And [20:56] I think that's a great concept. It took [20:58] me like a while to realize it that uh [21:01] the the nodes right now aren't static, [21:04] right? Uh it's a useful context which [21:06] Asian can act upon. Now here's example [21:09] how I apply this concept. I just uh like [21:13] went for a walk and then um after a walk [21:16] I just transcribed everything I was I [21:18] was telling. I use this uh microphone. [21:21] It's like wireless mic which you can put [21:24] on yourself. It's like very handy. And [21:26] then here is my transcript. I I got this [21:29] recording from my phone. Uh and then I [21:32] use Gemini Flash to transcribe it. And [21:34] then CL did it right. And here is all of [21:38] my uh brain appalysis with different [21:41] action items which I could do directly [21:44] within the clot. It has all of the [21:46] context about my goals and my tasks [21:49] which are available. Now the example [21:51] here would be is uh okay. So can you [21:54] tell me what are my tasks uh for today? [21:58] What's going on? Here we can show that [22:01] claude can get information uh about our [22:05] goals about our task. Uh here is a task [22:07] list. Can you open this dashboard for me [22:11] uh where you get my tasks from? Great. [22:14] And clude here opened my my tasks in my [22:17] Absidian uh which are exactly as it read [22:21] from my Obsidian. Right. And here all of [22:24] my tasks you have all of the context. I [22:26] also do the same for my goals here. They [22:29] are listed in um in in obsidian basis [22:32] that enables me to render this [22:34] information efficiently and clo can read [22:36] this and that's very very handy. Um yeah [22:39] for example can you read my goals which [22:41] I need to review every every morning. Uh [22:44] great. Uh let's see how it goes. Okay. [22:47] And it uh reads all of my goals. There [22:51] you go. And that's sorry and that's like [22:54] exactly the same ones which I have here [22:56] which is very very handy right you can [22:58] just like go and execute uh whatever you [23:02] want uh because all the context is here [23:05] within a clot code that's extremely [23:07] powerful now we are moving from this [23:10] markdown files [23:12] uh to make them executable with clot [23:14] code and I believe that's uh the real [23:17] differentiator that uh your context uh [23:20] how do you store your context how do to [23:21] feed it to your agent. That's the [23:24] highest leverage you can imagine. Now, [23:27] let's switch gears a bit. I think uh [23:29] also what's been very useful I just like [23:33] quickly quickly go about that. Um I [23:36] think for handoff prompts it's something [23:38] that um we need to um minimize the [23:42] friction and um have a standardized [23:45] processes on how to transfer context [23:48] between sessions. Now we have this [23:50] concept of u context window and context [23:53] budget. Now when you are here at this [23:57] point we need to compact our [23:59] conversation and the default compaction [24:02] is not great. What I typically do is I [24:04] have my custom uh handoff command for [24:07] another agent to continue conversation [24:10] from scratch and I also document um the [24:14] current progress the current learnings [24:16] in my Obsidian so that I can see and [24:19] track my sessions working sessions with [24:21] an agent. Uh here I have a dashboard for [24:24] my immigration where I have my my tasks [24:27] I need to apply for this postgraduate [24:28] work permit um etc. I have like those [24:32] sessions which essentially Claude wrote. [24:35] Uh here is the like immigration analysis [24:38] file which I asked CL to create when we [24:41] were running out of context. I want to [24:43] document uh what we learned so far. Uh [24:46] and this is reusable for the next [24:47] session. I can just tag this file uh [24:50] later like agent um agent u this right [24:56] and I can just tag it and I can just [24:58] continue working within all this [24:59] context. This is great. That's for the [25:02] hand of prompts. So you can just save a [25:04] session lock in Obsidian tag later, [25:06] return it, return back to it when you're [25:09] ready. And also I recommend um to curate [25:12] your list of sources, people who you [25:14] follow really matters. Uh um for me, I [25:18] just have had a couple um Seb Obsidian, [25:22] Simon Williamson, and Mario who created [25:25] this clot trace tool. that's really [25:28] helps to learn from the best uh to like [25:30] understand uh what's what's possible [25:32] what are the useful cont concepts which [25:34] you can apply uh and of course there's a [25:37] lot of experimentation [25:39] um which is happening here so now if you [25:42] want to start I would start very small [25:44] uh you can just turn on your uh verbos [25:46] output to understand what's happening [25:48] set up g review changes back up [25:51] everything um maybe you find useful like [25:55] user comment pattern which uh which I [25:57] showed uh definitely create your skills [26:00] that's a high leverage action analyze [26:02] sessions once per week there's a lot of [26:05] insights in information within your data [26:08] so I'm also share the Swiss preflow [26:10] skill under description and those are [26:12] the patterns which took me to figure out [26:14] 800 hours I run uh this obsidian and [26:18] cloud code workshop this weekend if [26:20] you're interested on integration of [26:22] personal contacts and we have some great [26:24] community forming around cloud code and [26:26] obsidian I had some great feedback uh [26:29] from from people. Yeah. </details>