One Prompt Burned Half My Claude Plan
Claude Opus 4.8 shipped a new feature called dynamic workflows.
I ran one prompt and it ate half of my $200 a month plan in about 30 minutes.
So I spent the week figuring out what these actually are, how they stack up against everything Claude Code already has, and when they're worth the spend.
TL;DR
- A dynamic workflow is Claude Code writing a script that spins up a ton of agents in parallel, then merges every result into one answer
- One prompt cost me half my monthly plan in 30 minutes, so scope is everything
- The ladder: ask Claude → Skill → Subagent → Agent Team → /goal → workflow
- /goal is depth (loop until done = true), a workflow is width (50 agents side by side)
- ultracode mode and /deep-research are the hidden pieces worth knowing
- Most knowledge work doesn't need a workflow, and that's completely fine
What A Dynamic Workflow Actually Is
I ran one that audited all 41 of my Claude Code skills.
It spun up 41 Haiku agents at once, one per skill, scored them in parallel, then fed everything into a single Opus synthesis agent.
Because the agents ran side by side, it was fast. It just ate a ton of input tokens. Roughly 5 million.
Not much came back out, so it wasn't crazy expensive. And I ended up with an HTML file that ranked every skill worst to best, gave me feedback on how to fix each one, and flagged patterns to watch as I build more.
Here's the part that makes it different from everything else in Claude Code.
Normally the plan lives with Claude in your main session. With a workflow, Claude writes a JavaScript file, and that file is what gets executed.
You describe the end goal. Claude turns it into a script that spins up and delegates all the agents on its own.
And those scripts can be saved and rerun whenever you want.
The Ladder: From Asking Claude To Workflows
Skills, subagents, agent teams, and workflows all feel similar. They're built for very different jobs.
- A Skill is a reusable recipe. You can run it yourself, trigger it as an automation, or let a subagent or a workflow run it.
- A Subagent is a parallel worker with its own context window, so it keeps your main session clean. Subagents can't talk to each other. They only report back to you.
- An Agent Team is a small crew that does talk to each other. Shared task list, individual roles, individual tools. Think war rooms and councils where agents debate toward one goal. More power, more cost.
- A Workflow is Claude writing a script that runs many agents, sometimes hundreds. They work alone like subagents, then all the results merge and come back to the main session.
The real variable across all four is how many agents are in the loop and whether they can talk to each other.
Picture a ladder. The higher you climb, the more complexity and functionality you get, and the more risk and money come with it.
1️⃣ Bottom rung: just talk to Claude Code. Ask, reason, maybe a web fetch or an API call.
2️⃣ A process you repeat: turn it into a Skill.
3️⃣ Work that should run parallel to your main session: hand it to a Subagent.
4️⃣ Subagents that need to talk to each other: Agent Team.
5️⃣ A giant parallel job: dynamic workflow.
The good news is you won't trip into a workflow by accident. Claude confirms before running one, and it takes real intent to fire it off.
Depth vs Width: /goal vs Workflow
This is where I kept getting confused, because I've been using /goal a lot.
The cleanest way to separate them is depth versus width.
/goal is depth. It's a loop. One agent (it can spin up subagents too) takes multiple passes, checking whether done = true, until the criteria is met. You could let it run 24 hours if you wanted.
A workflow is width. A bunch of subagents doing different things at the same time, synthesized at the end. You could have 50+ agents running horizontally. They're not looping against a finish line. They execute a plan set at the start, then hand back the result.
The wild combo would be a workflow nested inside a /goal. Very powerful, and a very fast way to burn money, so be careful with that level of autonomy.
The Money Part
Here's what stood out.
Some people are saying this feature only exists to make you burn more tokens. I don't buy that. There's real value here. The honest question is whether you actually need it, and a lot of the time it's overkill.
The bill depends on how big a job you point it at. Each agent is a full Claude call with its own window that has to read its own context. Spin up a lot of them and it adds up fast.
Most of the cost lands on input tokens, like my 5 million. Input is cheaper than output, but volume still stings.
My ugly example: I asked it to search my entire desktop. It crawled every local file and every repo looking for an analysis. That's the run that ate half my plan.
So before you fire one off:
- Bound the scope.
- Name the deliverable.
- Put all the worker agents on Haiku.
Same discipline applies to /goal. Give it a vague request and it'll loop forever because it never knows when done = true.
My filter is simple. If the job breaks into many pieces that can run on their own at the same time, a workflow fits. Reviewing every file in a codebase. A 400 file migration. High-risk work where you want max compute on each piece. For single edits, quick questions, or general knowledge work, skip it.
Running One Yourself
Here's the prompt I used:
"Run a workflow that audits all of my Claude Code skills in two places, my user level global skills and this project's skills. Grade on clarity, frontmatter pass/fail, and trigger quality, and give me the single highest value fix. One final agent ranks every skill worst to best."
I also told it to keep all the workers on Haiku.
When I sent it, Claude asked permission to run a dynamic workflow. I had to explicitly say yes. I could also view the raw script first, since under the hood it's just writing and then executing a script.
While it ran, /workflows showed me everything live. All 41 agents, what skill each was on, that they were all on Haiku, tokens used, tools, and how long each had been running. You can stop them from there too.
One gotcha. The JavaScript file it creates doesn't always land in your project. By default it saved somewhere more global in my Claude Code working directory. I had to tell it to drop the file into this project's .claude/workflows folder so it lived inside my own operating system, not off in a global directory.
Once it's saved, reuse is easy. Point at the workflow and say run it again. I've already got a second one saved that ranks Claude Code features.
ultracode And /deep-research
Two pieces worth knowing.
Run /effort to set your Opus reasoning level. The options are low, medium, high, xhigh, max, and then ultracode.
ultracode is the smartest and the most expensive. It runs xhigh reasoning and turns basically every prompt into a workflow. It bypasses a lot of permissions and just starts orchestrating, so the spend climbs quick.
Quick tip on invoking workflows. Type "workflow" in a turn and it lights up in rainbow, since the word shows up naturally in normal speech. That highlight alone doesn't run anything. The most reliable trigger is being explicit: "set me up a dynamic workflow to do this."
Then there's /deep-research. It automatically kicks off a workflow, spins up agents to research in parallel, has them vote on each claim, and hands back a cited research report. If you do a lot of research or you're building your own research skills, that one is worth testing.
The Whole Thing In One Slide
- Quick one-off → just ask Claude Code.
- Something you repeat → make it a Skill.
- Messy side task → Subagent.
- A small crew that needs to talk → Agent Team.
- Keep going until a goal is met → /goal.
- A giant parallel job → dynamic workflow. Just be careful.
And MCPs, CLIs, and API endpoints plug into any of these, so you can wire your own data and systems into the whole stack.
One last thing. Most of my Claude Code work is automations and knowledge work, not shipping software. So honestly, I probably won't reach for workflows much, and that's fine.
Knowing what a feature does and when it fits is the win. That's different from needing to use it every day to feel like you're keeping up.
I walk through the full build, the live run, and the ugly token bill in the video. Link in the first reply.