What is ultracode in Claude Code? What it did in my 4 planning documents

What is ultracode in Claude Code? What happened each time I used it

I typed the word ultracode into Claude Code for the first time on June 3. I had read in the docs that it was the keyword for "dynamic workflows", where Claude writes a script that runs many subagents at once. In that first session nothing spectacular happened. Three subagents went off and audited my skill folder. Since then I have used it for four planning documents, one of them built by 17 agents, and I have also learned when not to use it.

I'm not a developer. I run a small home setup of scripts, bots and Claude Code sessions, and I describe what I want in plain sentences. So this post skips how to write a workflow script. It covers what the keyword did for me, with dates.

By L

Contents

  1. What the keyword does
  2. The first run did not use a workflow at all
  3. The four planning documents
  4. The July fix that mattered to me
  5. Measuring it on August 10
  6. When I don't use it
  7. FAQ

What the keyword does

A dynamic workflow is a JavaScript script that Claude writes for the task you describe. A runtime executes it in the background while your session stays usable. The script decides what runs next, holds the intermediate results in variables, and launches subagents, so Claude's own context only receives the final answer.

The difference from a normal subagent is who holds the plan. With subagents, Claude decides the next step every turn. With a workflow, the script decides.

How you turn it on, as I noted from the docs on June 3 and checked again today in my local Claude Code (version 2.1.307):

  • Put the word ultracode in your prompt. That is the opt-in for a single task.
  • Ask for a workflow in your own words. "Run this as a workflow" also counts.
  • Turn it on for the whole session with /effort ultracode.

The tool description in my current version says it should only be used when the user has opted in one of those ways. By default it keeps a run under 10 agents, and there is a "Dynamic workflow size" setting in /config if you want a different scale. /workflows shows a run's progress while it's going.

Two limits I wrote down from the docs on June 3: workflows were a research preview for paid plans on version 2.1.154 or later, with a cap of 16 agents at the same time and 1,000 per run. Those numbers are from June and may have moved since.

The first run did not use a workflow at all

On June 3 I was in the web version of Claude Code, version 2.1.161. The keyword was recognized. The system clearly injected an instruction to use the Workflow tool. But the Workflow runtime wasn't available in that web environment, so Claude fell back to its ordinary subagent tool and launched 3 of them in parallel, one per domain, and merged the results.

The 3 agents audited 61 skills in my setup for overlap and relevance to my remaining work. The report listed about 24 candidates for merging or retiring; usage logs were missing, so it was a structural audit, not a usage count. It also hit 2 server errors (500) before it succeeded, so I would not call it stable.

The keyword is an opt-in, not a guarantee. If the runtime isn't there, you get the normal fan-out.

The four planning documents

From late June I used ultracode when I asked Claude Code to write a plan or a spec. The runs I recorded:

DateDocumentAgentsWhat the script did
June 23Spec for my "AI office" (a front end plus a message relay plus a PC bridge)175 discovery agents, then 3 competing architecture proposals judged and merged, 3 component designs, a draft spec, and 3 adversarial reviewers
June 24Spec for re-centering my always-on agent around its memory9Read-only verification and adversarial review of the draft
July 2Master plan with 5 directions9A survey of my own system and outside references, 228 tool calls in total
July 11Spec for a draft publisher for my Korean blog4Donor pattern search, asset inventory, integration contract, web check, 90 tool calls

The June 23 run is the one I would show someone. The 3 adversarial reviewers at the end found 6 blockers and 13 major issues in the draft, and all of them were folded back into the spec before I approved it. One blocker was that two parts of the system were both set up to receive the same Telegram messages.

The run produced a reviewed document in one sitting. I don't have the background to review an architecture proposal myself.

The July fix that mattered to me

My July 18 notes record this changelog entry for version 2.1.210:

Fixed the ultracode keyword opt-in firing on non-human-originated input such as webhook payloads and relayed PR comments

Before that fix, the keyword could trigger a multi-agent run from text that a human did not type, such as a webhook payload or a relayed pull request comment. My system relays Telegram requests to Claude Code. A relayed message containing that word could have started a large run on its own, and I would have found out from the usage meter. If you run Claude Code from any kind of automation, update past 2.1.210 before you leave the keyword anywhere near it.

Measuring it on August 10

In August I ran a small pilot to see whether a workflow was better than doing the same job by hand in a session. It was a document-scoring pipeline, with a cross-check by Codex, a different company's model, reserved for the highest-tier documents.

Result
Agents4
Wall time8 minutes 39 seconds
Failures0
ThroughputAbout 3 times the manual way
ReproducibilityOne document scored 73 in the workflow and 76 by hand, same grade and same weakest axis

I did not separate token cost in that pilot, so I can't tell you it was cheaper. It was faster, and the scores held.

The pilot also caught a mistake in my own experiment notes. My notes used 89, a score after one round of polishing, while the first-pass score was 67, below the baseline. The earlier manual review had missed the discrepancy. I saved the script as a reusable workflow file so the scoring can be rerun without rebuilding it.

When I don't use it

My August 10 notes set a rule: use a workflow when a job involves at least two of repetition, branching, parallel work, retries, approvals, or an audit trail. A single lookup or a one-line change gets no workflow.

For workflows that write files, I added 4 conditions: the agents work in an isolated git worktree, my Claude Code is version 2.1.222 or later, they never push or merge (I review the diff in my session), and the first run covers 1 to 3 items. Four things stay off limits: my rules and settings, operational data such as spreadsheets and ledgers, secrets, and deletions. The August 11–12 handover records the first write run under those conditions. It updated README files in 3 repositories, and the originals stayed clean.

One correction to my June notes. I had written that a running workflow could not take user input, so sign-offs would have to be split into separate runs. By August I found that was wrong. You can watch a run with /workflows, skip a step, stop it, or resume it later. So an approval step can sit inside the script instead of between two scripts.

My usual split is still Claude Code for the plan and Codex for the review (why I give them different jobs).

FAQ

What does ultracode mean in Claude Code?
It's the keyword that opts a prompt into multi-agent orchestration. Claude writes a workflow script, and a runtime executes it in the background with many subagents. Without the keyword, a session or a workflow request in your own words, Claude isn't supposed to use that tool.

Does typing it always start a workflow?
Not in my experience. On June 3, in the web version, the keyword was recognized but the runtime wasn't available, and Claude used 3 normal subagents instead.

How many agents does it use?
Whatever the script needs. My runs used 17, 9, 9 and 4. My current version's default guideline keeps a run under 10 unless the prompt calls for more, and /config has a setting for the size.

Is it cheaper than doing the work in one session?
I don't know. The docs say it uses more tokens, and my August pilot didn't separate token cost. It was about 3 times faster for the scoring job.

Can I stop a run once it's started?
Yes. /workflows shows progress, and you can skip a step, stop the run, or resume it later. I had this wrong in June.

Comments

Post a Comment

Popular posts from this blog

Python script works manually but not in Task Scheduler: the Microsoft Store python stub (and a daily alive check)

Image
My Python script ran perfectly when I started it by hand. My weekly publish job had produced nothing for 3 weeks, and I got zero error messages. On July 7 I found one problem: the scheduled task called plain python , which in the scheduler's environment resolved to the Microsoft Store stub. That fake python.exe did no work and exited. I replaced python with the full path to the real interpreter. Then I added a daily "I'm alive" message so the next silent failure can't hide for weeks. Correction (October 17, 2026): The first version of this post said the Saturday task's result code, -196608, came from the Microsoft Store python stub. My later repair notes show that code came from escaped quotes in the task's settings. The stub problem below was real, and I fixed it, but it wasn't behind that code. Details are in the "A second cause" section. If you only need the fix, jump to "The one-line fix" below. The rest is how I found it ...
Read more »

Schedule a weekly AI summary with Windows Task Scheduler (with a human OK step)

Image
My AI assistant had one big weakness: it does nothing until I open it. Every Sunday night I was supposed to ask it to summarize my week, and I forgot more often than I remembered. Windows Task Scheduler fixed the forgetting part. At 10 pm every Sunday, Windows runs a small PowerShell script that drops a signal file and notifies me. The AI does not run at 10 pm. The next time I open a session, the AI sees the signal file and tells me a weekly job is waiting. I press OK, and only then does it do the work. During my 14-day trial, my first approach failed and I ran into two bugs in the first week. I also added three guardrails. I call the whole setup "Jarvis." Everyone names their assistant, and I'm no exception. By L Contents Attempt 1: telling the AI "run at 10 pm on Sundays" (failed) What worked: queue the job from the OS Why a human still presses OK Week 1, bug 1: the PC was off Week 1, bug 2: it fired twice A minimal version of the setup The test on my real...
Read more »

Codex vs Claude Code: which jobs I give each (3 months of work logs)

Image
I use Codex and Claude Code for different jobs. Between July 17 and October 16, 2026, my AI work log got 442 lines. Claude Code's models appear on 314 of them, mostly building and writing. Codex appears on 90, and 80 of those were in the reviewer role. Scripts with no AI took 18 lines, and 20 lines have no model name (10 of them are no-model steps marked "none"). By default, a model from the other company reviews the work. If you only want the setup, jump to "My routing table." The rest is the reasoning and what the log says. By L Contents Why a second company What the 442 lines say My routing table When one of them runs out If you're starting with one tool FAQ Why a second company I'm not a developer. I describe what I want, and the AI writes the code, runs the commands and writes the documents. That means I can't always tell when a result is wrong. A broken screen is easy to spot. A wrong number in a correct-looking table is not. My reason...
Read more »

What not to automate with AI: my approval rule for money and irreversible actions

Image
I let AI handle all of my reading, summarizing and analysis automatically. I let it create and edit files too. But anything that spends money or publishes something to the outside world still waits for me to press approve. I decide where a task goes with two questions: if this goes wrong, does money leave my account? And can I undo the result? If a mistake could cost money, or the result can't be undone, the task needs my approval before it runs. That line hasn't moved since I drew it. What did change was how I approve things. I built a one-tap approve button for my phone, and I took it down myself 26 days later. That story gets its own post. This one is about the line itself. By L Contents The worry that grew as the AI got better Why "just send me a notification" wasn't enough What my approval step looked like The line you can draw today Common questions The worry that grew as the AI got better I had already split my AI helper into three roles: planning, d...
Read more »

Telegram bot receives /start automatically: how it triggered my kill switch (one-line fix)

Image
I opened my Telegram bot's chat, tapped Start , typed nothing else, and my bot executed "release" on its emergency stop. I found that Telegram sends /start to the bot when you first open its chat and tap Start . One of the words I had mapped to my "release" command was from the start family, so that automatic /start was read as "release." I fixed it in one line by removing every start-family word from commands that change state and keeping /start for a usage message only. By L Contents The setup: a 3-second emergency stop on my phone The symptom: "release executed" in the log The cause: Telegram's automatic /start Reproduce it in 5 minutes The one-line fix 3 kinds of commands that should never share a word with /start Checklist: safe bot command design What I took from it FAQ The setup: a 3-second emergency stop on my phone I hand money-related chores to a few AI assistants. One assigns work, one processes it, and one checks the r...
Read more »