21:21Mark Kashef · Tutorial
How I Built an AI Council with Claude Code Subagents
A 21-minute walkthrough of spinning up three competing AI agents that debate your ideas and write their reasoning to a shared file.
January 23rdThe argument in one line.
Dynamic workflows earn their token cost through verification, not volume: a second wave of agents that adversarially checks the first wave is what makes Claude Code outputs trustworthy rather than merely fast.
Read if. Skip if.
READ IF YOU ARE…
- You use Claude Code daily and have accumulated more than a handful of sessions, skills, or CLAUDE.md files.
- You have been burned by a Claude research output that sounded confident but turned out to be ungrounded.
- A new model just dropped and you want an upgrade guide calibrated to your actual prompting patterns, not a generic changelog.
- Your ~/.claude folder has grown organically for months and you suspect it contains dead weight, contradictions, or duplicate skills.
SKIP IF…
- You have not used Claude Code yet -- multi-agent workflows compound on top of a working baseline you do not have yet.
- You are looking for no-code AI automation; this requires writing workflow prompts and understanding JSONL file structure.
The full version, fast.
Most people use Claude Code as a single-thread chatbot. Dynamic workflows let one Claude instance spawn and coordinate a fleet of sub-agents in a single window. The video walks through three patterns: mining your own JSONL session logs to generate a personalized model-migration guide; using Apify to pull real X threads, then running 200 adversarial agents to fact-check 170 claims (116 survived); and auditing your .claude folders to find duplicated skills, contradictory rules, and at least one hardcoded API key. The lesson across all three: the verification pass -- the second wave of agents that checks the first -- is what earns the token spend.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:46
01 · The report it built in an hour
Opens on the live HTML upgrade guide personalized to the host actual session history. Hook is the output, not the process.
00:46 – 01:07
02 · The 3 use cases
Promises three workflow patterns, teases the third as most useful, retention hook to end.
01:07 – 01:38
03 · Use case 1: Opus 4.8 upgrade guide
Walks through the report output -- prompting pattern changes, checklist, what stays vs. what shifts.
01:38 – 02:48
04 · The model-is-insane trap
Shows a grid of INSANE thumbnail mockups -- the problem the workflow solves. Generate your own tutorial from your own data instead.
02:48 – 05:03
05 · The prompt (JSONL + claude-code-guide)
Shows the full 3-step workflow prompt live on screen: data analysis, model comparison via claude-code-guide, synthesis and HTML output.
05:03 – 06:14
06 · Make a 2-min video + a skill
Pipes the HTML report into Hyperframes to auto-generate a tutorial video. Saves the full pipeline as a reusable /model-migration skill.
06:14 – 07:40
07 · Use case 2: self-checking deep research
Introduces adversarial verification as the key differentiator over single-agent research. Shows live fact-check report on dynamic workflows claims.
07:40 – 09:06
08 · The research prompt (Apify X scraper)
Shows the full research prompt -- Apify setup, angle fan-out, adversarial verify pass, claim-survival scoring.
09:06 – 11:01
09 · Use case 3: audit your .claude
Fans out agents across global and project .claude folders. Shows live audit report: duplicated skills, contradictory rules, hardcoded API key, stale demo skills.
11:01 – 12:07
10 · More uses + grab the prompts
Extends to other domains like travel booking. CTAs for Living Course and free prompts pack.
Lines worth screenshotting.
- Dynamic workflows are most useful for verification, not volume -- the second agent pass that checks the first is what makes outputs trustworthy.
- Your Claude Code JSONL files contain every session you have ever had; a workflow can mine all of them for personalized prompting-pattern analysis in under an hour.
- The claude-code-guide agent is a native Claude Code sub-agent trained on Anthropic internals -- it acts as a live fact-checker against the actual release notes, not your memory of them.
- 170 X claims about dynamic workflows were examined adversarially; 116 survived, 14 were cut, 40 could not be confirmed -- the viral framing almost never matches the underlying GitHub PR.
- The Bun port is real and the demo genuinely happened; the specific numbers were inflated -- the actual merged PR added ~1.07M lines over 6 days, not 750K lines in 11 days.
- Token burn from multi-agent workflows can be scoped down interactively -- Claude will ask how deep to go, and you can cut from 70 agents to 20 before it starts.
- Your ~/.claude folder accumulates dead weight faster than you think: 10 stock Anthropic skills duplicated byte-for-byte, a demo skill never used again, and contradictory model rules are all common findings.
- The /model-migration skill pattern is model-agnostic -- the same workflow built for Opus 4.7-to-4.8 will work for any future model drop with no changes.
- Hyperframes (heygen-com/hyperframes) is an open-source library that converts HTML to video -- wiring it to Claude Code lets a workflow auto-generate a 2-minute tutorial from the report it just produced.
- Dynamic workflows are not a daily tool -- there are only a handful of tasks that justify the token burn, and monthly is roughly the right cadence for ecosystem audits.
Verification is the workflow, not a step.
WHAT TO LEARN
A second wave of agents that argues against the first is what earns the token cost of multi-agent workflows -- and it applies to research, model migrations, and your own tool setup alike.
03Use case 1: Opus 4.8 upgrade guide
- Your Claude Code JSONL files are a complete record of how you actually prompt -- mining them gives you a model-transition guide calibrated to your patterns, not a generic changelog.
- Reusable skills are the compounding return on agentic work -- saving a completed workflow as a skill means the next model drop costs one command instead of rebuilding from scratch.
07Use case 2: self-checking deep research
- Single-agent research cannot fact-check itself; there is no internal devil advocate, so confident-sounding outputs are still unverified until a second agent explicitly tries to refute them.
- Adversarial verification is a structural choice, not a prompt tweak -- it requires a separate verification pass with access to primary sources, not just asking the first agent to double-check.
- Of 170 X claims about dynamic workflows examined adversarially, 116 survived and 14 were cut -- the viral framing of real events is frequently inflated even when the underlying feature is genuine.
09Use case 3: audit your .claude
- A Claude Code setup that has grown organically for months reliably contains duplicate skills, contradictory rules, and at least one hardcoded secret; an audit workflow surfaces these faster than manual review.
- Token burn from multi-agent workflows can be negotiated interactively -- Claude will ask how deep to go, and scoping down from 70 agents to 20 before the run starts is a legitimate cost-control strategy.
- The value ceiling of dynamic workflows is set by task type, not ambition -- they suit large, messy, multi-source problems that need validation; they are wasteful for tasks a single sharp prompt handles.
Terms worth knowing.
- Dynamic workflow
- A Claude Code feature where the model writes and executes its own orchestration script, spawning and coordinating multiple sub-agents in parallel within a single session window.
- JSONL file
- The conversation log format Claude Code writes to disk -- one JSON object per line, capturing every message, tool call, and model response from a session. Stored at ~/.claude/projects/.
- claude-code-guide agent
- A built-in Claude Code sub-agent specialized in Anthropic internal documentation, release notes, and the Claude Code codebase. Invoked by name to cross-reference official sources.
- Adversarial verification pass
- A second wave of agents assigned to argue against the findings of the first wave. Each claim requires a 3-lens majority vote (primary source, corroborating evidence, refuting evidence) to survive.
- Ultracode
- An extended Claude Code mode that allows higher agent concurrency and longer-horizon tasks, at a substantially higher token cost per run.
- Hyperframes
- An open-source library by Heygen (github.com/heygen-com/hyperframes) designed to convert self-contained HTML files into rendered video -- built explicitly for agent-generated content pipelines.
- Apify actor
- A packaged scraping or automation task on the Apify platform, runnable via API. Used here to pull real X threads rather than relying on web articles about X posts.
- Model migration skill
- A reusable Claude Code skill (/model-migration) that encodes the full JSONL-mining pipeline so it can be re-run with one command on any future model release.
Things they pointed at.
03:33toolclaude-code-guide agent
Lines you could clip.
02:51
“You upgraded on evidence, not vibes.”
06:35
“There is no devil advocate outside of that one running session.”
08:51
“The Bun port is real. The viral framing is not.”
10:45
“There are only a handful of tasks that really deserve the level of token burn that you get by spinning up all these agents.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphor
00:00Alright. So take a look at this. This is a web page that Claude's new dynamic workflows built for me in under an hour.
00:06And, no, these aren't some generic release notes or a summary from someone else's perspective on how to use Opus 4.8. This workflow spun up tens of agents to go through all 1,500 conversations that I've ever had with Cloud Code to tailor a full report on exactly how I should use this brand new model and what it should change around my approach.
00:27In addition to the report, it also created a full length two minute tutorial walking through every single feature that I should care about, how it works under the hood so I never have to go and rely on some generic tutorial that's pretty much a summary of the documentation. And this is just one of three different use cases that I'm gonna walk through in this video to show you where dynamic workflows can really help move the needle in your day to day.
00:50So if you wanna learn the use cases, especially the last one, which I think you'll really like, then you wanna stick with me till the end of the video, and I'll save you from having to burn those tokens yourself. Let's dive in. Alright.
01:01So like I said, this report walks through how I should alter my shift to using this model, how I should change my prompting patterns, what should stay the same, and what should differ or be slightly adjusted for this new model and the way it behaves. So one of the key things was being more literal with it. It's better at dealing with direct prompts without having to beat around the bush.
01:21And if I just wanted a overall summary of every single thing that I should do in my day to day based on all the conversations it's gone through and what it's seen in terms of patterns, this would be my checklist. And before I walk you through the exact prompt that led to that report and the video, it's important to understand what we're solving for in this first use case.
01:41So if you're like me, you will see a brand new model come out and you'll be curious. What's different? And then you'll go on YouTube and you will inevitably see the that say, this brand new thing is insane.
01:55Even though this model could be one incremental step change, which 90% are, where there are some pros and cons, but overall, it's not a night and day difference. So instead of clicking on Opus 4.8 is insane, Opus seven, or Opus 492, what if you could just create your own tutorials that are tailored to what you care about?
02:15So the way this works is that you ask Cloud Code to go through all the JSONL files that have all your conversations, and all of these are stored on your local computer if you use Cloud Code on your personal laptop.
02:27And then you can have all of these agents spin up and mind through all of the sessions and look for all the patterns. And once it understands the patterns of how you use ClaudeCode, how you prompt it, then you'll get tailored advice whether it's 4.8 or five or even you can use this for codex or Gemini.
02:46You can really tailor your own education and your own upskilling to the things that actually move the needle forward. When it comes to the prompt, you could send over something like this to generate a very similar report. So you would say, would like you to build a workflow.
02:59These would be the magic words. Again, when you write the word workflow, you'll get a multicolored name for the workflow, which means you're using the function that analyzes all JSONL files.
03:09A JSONL file is a fancy JSON file, basically, restructured in a different way in folder path here. Now if you don't know the path, then one, you can ask Claude code. If I show you a little preview here, if we go to this folder, this has every single conversation that I've had in my YouTube folder.
03:28So you'll see that they are JSONL files, like I said, and you can use something like Sublime Text or Notepad plus plus to take a look at them. The TLDR is you will see a combination of assistant and you. Assistant and you.
03:41So that basically summarizes all the tool calls. It's very token heavy. Having all these agents spin up outside of your context window using dynamic workflows will help them not bloat your existing conversation.
03:53Then I give it the overall goal, which is to help me understand how to extract maximum value from this new model. Again, you can change this to whatever model you want. Then I walk through all the criteria.
04:03So step one is data analysis. So going through every single JSONL file and looking for my token usage patterns. Then for number two, model comparison.
04:13I like to invoke this Claude code guide agent. This is a sub agent that is native to Claude code. So its entire role is to be the PhD scientist in the latest documents from Anthropic.
04:26So when you ask this a question, it's a . They'll go and double check everything that is on the website. So we'll be able to see the release notes and then compare it to step one's output, which is analyzing your core workflow.
04:40And then in step three, I basically walk through all the core synthesis and analysis that I want. So number one is some form of executive summary. Number two, a usage profile analysis, then a direct comparison of both competencies, then prompting patterns, then skill and tool recommendations, and then dynamic workflows that I could put together based on this new feature and advanced opportunities.
05:01And in terms of the video that I put together, all I did was hook up something called Hyperframes, which is an open source library that is created by Heygen and then linked it to Claude code and told it to take this HTML page it put together and convert it to a short two minute video on two x speed.
05:18And the best part of this is once all of these agents that have been fanned out go and figure out how to accomplish the task, you can always materialize this entire path and pattern by creating a skill. So in this case, I created something called model migration.
05:33So in the future, when we have 4.9, five, six, etcetera, I can now run this every single time and quickly get my explainer HTML and my video so I can quickly get up to speed with what matters to me.
05:47And by the way, if you want access to exclusive skills just like this one, along with all the exclusive content that we keep adding to our Claude Code Living course, then you'll wanna check out the first link down below. As these agentic systems mature, you're gonna want to mature your skill set as well. So if you're done sitting on the sidelines and just passively watching YouTube videos and you wanna go hands on and become a master in a matter of weeks, then check out the link down below, and I'll see you inside.
06:12Alright. Back to video. And for use case number two, I found that using dynamic workflows for deep research where you actually want all the agents to spin up and check each other's work versus spending $10.15, 30 minutes having an individual agent go and check a series of sources.
06:28Many times it tells you 200 or 300 sources, But in reality, it doesn't actually fact check itself. There is no devil's advocate outside of that one running session. So running deep research, especially to look into how to leverage these features like dynamic workflows becomes really handy.
06:46This is a full report on every single claim it found on x on what dynamic workflows seem to be able to do and not do. And with each one, it has either a link to a GitHub repo or a link to an existing x post or a link to a report.
07:02So I can scan through all of this and see you can see right here, it says the word survives, meaning it had a series of agents double check and fact check these claims from x to make sure that by the time it comes to us, it's distilled and validated. And then you have things like this, where obviously on x, people will make some bold claims on what something can or can't do.
07:23And then you have other posts that are unconfirmed. So now you can look at this with a second set of eyes and double check and see if you can trust any of these claims that you see. Then you can see every single claim that seems to be cut.
07:35So this is amazing for double checking and basically peer reviewing via agents different types of research claims and information in general. This is the prompt that I used to execute this. So I said, I want you to research all the findings on dynamic workflows associated with Opus 4.8, pull together everything that people are actually reporting about it, the real results, the claims, what holds up, and what is just hype.
07:57And then I go through and I tell it to use the Apify x scraper. So you can use all kinds of means to scrape tweets or go through tweets. They have their own API on x now, but I like to use Apify for all kinds of things including pulling YouTube analytics.
08:12So I just tell it to go and use the skill to find this what's called an actor, basically an app that you can use on Apify. Once it pulls all of those, then it goes through and it spins up all of these agents. And as it calibrated how it's going to execute the task, what angles it will look at, it asked me how deep do you wanna go.
08:29So one thing that most people are not talking about is if you use something like workflows or ultra code, you can intervene and tell it, I don't want you to go to the level of depth of having 60 or 70 agents. So this feature does burn tokens, at a a huge rate for big tasks. But you can also measure the task and scope it down if you need to.
08:50And twenty minutes later, it came back with the full response that linked to the very web page that I showed you, walked through the fact that it went through a 170 claims, spun up almost 200 agents to go through and adversarially verify all the passes on the features that seem to be legitimate versus the ones that seem to be more hype.
09:08And for the last use case, you can also use workflows to audit your existing Claude code ecosystem to look for opportunities where you can use it better. Specifically, you could have it take a look and audit your dot Claude usage.
09:21And these are basically folders that have all the core assets at the project level and the global level. And these are things like your agents, your skills, your rules, your hooks, and commands. So as you keep building and as you keep adding more skills, over time, things will go stale.
09:38And inevitably you'll have some inefficiencies. So you can have these agents spin up and then tell you exactly what are the core things you should fix and why. Then as you go through, you could see if you have overlapping skills, duplicate skills, and ideally what you should do about them.
09:54So this one here is telling me that I have 10 stock anthropic skills that seem to be duplicated. So perhaps during a demo for a client, I copy pasted it more than once. So now that is bloating my context against my will.
10:06If I scroll the bottom, here's another one where I seem to create this designer designer skill specifically for a YouTube video as a demo, but I never use it again. And these are the things that you don't have time to revisit. So if you're building things up like your agentic OS, having the way to look at your low, medium, and high level issues or repeated contradictory rules is a great way to keep on optimizing maybe once a month.
10:30Because like I said before, this is not something designed to be used every single day. This is something you can use sporadically every couple weeks, maybe once a month. And there are only a handful of tasks that really deserve the level of token burn that you'll get by spinning up all these agents.
10:45So the goal here is that you're able to constantly audit things like your clot MD, your rules and skills, etcetera, to look for ways that you can keep stepping up your game and making sure they don't have these areas of bloat or isolated folders where you spin something up, use it a couple times, and then you completely forget about it.
11:03And that's pretty much it. It's just three of multiple ways that I've been using dynamic workflows, especially for personal usage for day to day tasks.
11:10So hopefully, gives you some inspiration for the different types of things you can use workflows for. Beyond just the business use cases, there are different things that I've tried with workflows like booking travel, looking through bookings and hotel sites and travel prices for airlines. You can apply this to all kinds of very large meaty problems that need a lot of research and validation.
11:31If you wanna grab all the prompts that I showed you in this video, you'll find them in the second link down below. And if you wanna go much deeper on things like Claude code and become a master of it in just a couple weeks, then you'll wanna check out the first link down below for my early adopters community. I drop at least one new module a week, and I'm gonna walk through how I built the hyperframes workflow and skill that I alluded to earlier in this video exclusively in there.
11:56So if you wanna check it out, then maybe I'll see you inside. And for the rest of you, if you found this video helpful, useful, and novel, then I'd super appreciate a like on the video and a comment if you so choose. I'll see you in the next one.
The Hook
The bait, then the rug-pull.
A generated HTML report -- personalized to 1,500 real Claude Code sessions, not pulled from a changelog -- is the opening image. The contrast is immediate: this is what you get when you stop watching other people tutorials and point the model at your own work.
Named ideas worth stealing.
02:48model
3-Step Dynamic Workflow Prompt Structure
- Step 1 Data Analysis -- recursively read all JSONL files, categorize prompting patterns, skill invocations, tool uses, agentic behaviors
- Step 2 Model Comparison -- use claude-code-guide agent to cross-reference official release notes against your actual usage patterns
- Step 3 Synthesis and Output -- produce self-contained HTML with executive summary, usage profile, macro/micro differences, prompting upgrades, skill recommendations, migration checklist
The exact 3-step prompt structure for generating a personalized model-migration guide from your own JSONL logs.
Steal forAny time a new Claude model drops -- run this once, get a tailored upgrade guide instead of five generic tutorials
06:14model
Adversarial Verification Pass
- Pull claims from primary sources via X scraper or GitHub
- Fan out research agents across multiple angles simultaneously
- Run a 3-lens verification panel on every claim: primary source / corroborating / refuting
- Require majority vote to survive; tag SURVIVED, CUT, or COULD NOT CONFIRM
- Every surviving claim carries an inline source link
The pattern for turning a fast research answer into a verified one by having a second wave of agents argue against the first.
Steal forAny research task where you need to distinguish real capabilities from hype
09:06list
Ecosystem Audit Severity Tiers
- HIGH -- Duplicate skills, hardcoded secrets, stale banned-section titles
- MEDIUM -- Overlapping skills that are supersets of each other, conflicting model defaults across CLAUDE.md files
- LOW -- ASCII-split skills that should be one template, follow-up rules duplicated across skill files
Three-tier severity classification for .claude cleanup actions.
Steal forMonthly maintenance pass on any grown-organically Claude Code setup
How they asked for the click.
VERBAL ASK
05:42product
“If you want access to exclusive skills just like this one, along with all the exclusive content that we keep adding to our Claude Code Living course, then check out the first link down below.”
Mid-roll placement before use case 2, at the natural break after the model-migration skill demo. Aggressive but timed to a high-value moment. Secondary CTA at 11:25 for free prompts pack.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
