How to Actually Use Claude Code Dynamic Workflows
The video explains Claude Code's new dynamic workflows feature, comparing it to skills, sub-agents, agent teams, and the /goal command. The presenter demonstrates a real workflow that spun up 41 parallel Haiku agents to audit his skills, consuming ~5 million input tokens. He advises careful, targeted use of workflows due to their high token costs.
Summary
The presenter opens by demonstrating a dynamic workflow that spun up 41 parallel Haiku scoring agents to audit all of his Claude Code skills, feeding results into a single Opus synthesis agent. The output was an HTML file ranking skills worst-to-best with improvement feedback. This example consumed approximately 5 million input tokens, though output tokens were minimal, keeping costs somewhat reasonable.
The video then systematically distinguishes between four related but distinct Claude Code features: Skills (reusable recipes or automation prompts), Sub-agents (parallel agents that don't share context with each other but report back to the main session), Agent Teams (a small crew of agents that can communicate with each other, share task lists, and have individual roles), and Dynamic Workflows (Claude writing a JavaScript script that orchestrates many sub-agents in parallel, with all results merged at the end). The key distinction is who holds the plan — in workflows, Claude generates a JS file that drives execution, rather than keeping the plan in the main session context.
The presenter introduces a 'ladder' metaphor: as you move from simple chat → skills → sub-agents → agent teams → dynamic workflows, complexity, functionality, risk, and cost all increase. He also contrasts workflows with the /goal command: /goal is a depth play (looping until a done-equals-true criterion is met), while workflows are a width play (many agents executing simultaneously on a pre-established plan without checking a completion criterion).
On cost, the presenter warns that workflows can be extremely expensive. In one session he burned through roughly half of his $200/month subscription with a single prompt that had agents crawl his entire desktop and repos. He advises bounding scope, naming deliverables explicitly, and assigning sub-agents to the cheaper Haiku model to manage costs.
Practical tips include: using /workflows to monitor running workflows, saving workflow JS files into your project directory (they default to a global location), and invoking workflows explicitly by saying 'set me up a dynamic workflow' rather than just using the word 'workflow' casually. The presenter also mentions the /deep research command as a built-in workflow that runs parallel research agents, votes on claims, and produces a cited report.
Finally, the video introduces 'Ultra Code' mode (/effort max), which combines extra-high reasoning effort with automatic workflow invocation on every prompt — described as the most powerful but most expensive mode. The presenter closes with a one-slide decision framework: quick task → ask Claude; repeatable process → skill; messy side task → sub-agent; small collaborative crew → agent team; run until objective met → /goal; giant parallel job → dynamic workflow.
Key Insights
- The presenter argues that dynamic workflows differ from agent teams in a fundamental architectural way: workflows have Claude generate a JavaScript file that drives execution, meaning the plan lives in the script rather than in the main Claude session context.
- The presenter distinguishes /goal from workflows as a depth-vs-width tradeoff — /goal loops through multiple passes until a done-equals-true criterion is met (potentially 24+ hours), while workflows deploy many agents horizontally on a fixed plan with no completion-checking loop.
- The presenter reports burning through roughly half of his $200/month Claude subscription in a single workflow prompt that crawled his entire desktop and local repos, illustrating how input-token-heavy parallel agent jobs can be extremely costly even when output tokens are small.
- The presenter reveals that 'Ultra Code' mode (/effort max) effectively combines extra-high reasoning effort with automatic workflow invocation on every prompt, meaning it bypasses many permissions and defaults to orchestrating workflows — making it the most expensive available mode.
- The presenter notes that workflow JS files are saved globally by default rather than inside the current project directory, and that users must explicitly instruct Claude to save them inside the project folder (e.g., in .claude/workflows) to keep them organized and reusable per project.
Topics
Transcript
[0:00] So, I just ran this cloud code dynamic workflow that analyzed all of my skills. You can see here that it spun up 41 Haiku scoring agents because I have 41 skills here. And then it's feeding all of that analysis into an opus synthesis agent. Now, because all of those agents were working in parallel, which is pretty cool. It didn't take that long, but it did take a ton of input tokens. It ate up basically 5 million input tokens, but there wasn't that much to output, so it wasn't super expensive. And I ended up with this HTML file where I could now look at all of my skills. It's ranked them worst to best and…
Full transcript available for MurmurCast members
Sign Up to AccessMore from Nate Herk | AI Automation
Fable 5 + Karpathy’s LLM Wiki is Basically Cheating
The video demonstrates how to build a personal LLM-powered knowledge base using Obsidian and Claude, inspired by Andrej Karpathy's approach. By ingesting various data sources (YouTube transcripts, PDFs, URLs) into an interconnected wiki structure, users can create a 'second brain' that helps AI agents understand context and relationships across knowledge domains.
How Claude is Creating a New Generation of Millionaires
Claude, an AI built by Anthropic, is enabling non-technical founders to build million-dollar software companies by automating development work. The transcript outlines why Claude is outpacing competitors, provides real examples like Vulcan, and offers a four-step framework for anyone to start building with Claude today.
How Anthropic Engineers Actually Prompt Fable 5
A guide to effectively prompting Claude Fable 5, covering six key techniques including providing context/intent, using negative prompting, letting the model act once ready, verifying outputs, avoiding reasoning requests, and keeping instructions concise. The speaker emphasizes that Fable 5 is expensive and should only be used 5-15% of the time, with pricing at $10 per million input tokens and $50 per million output tokens.
Stanford's Method Turns Claude Into a PHD Level Research Team
A researcher demonstrates the STORM method from Stanford, which uses five expert perspectives (practitioner, academic, skeptic, economist, historian) to create verified research reports. The method produces 25% more organized articles than competing approaches and is packaged as a reusable Claude skill that generates HTML briefings with peer-reviewed citations.
Is Claude Mythos Coming?
A YouTuber analyzes the brief appearance of 'Claude Mythos' on Anthropic's API, arguing it signals a marketing move rather than an imminent public launch. Despite competitive pressure from OpenAI and IPO timing, the creator believes Mythos capabilities will quietly fold into future Opus models rather than release publicly under that name. Three possible scenarios are outlined ranging from a limited gated release to Mythos remaining permanently restricted to vetted security partners.