Developer

How Vox Used 105 Opus 5.5 Agents to Turn 1,500 Posts into 66 Skills

Built by Vox · · Source: X

Quick answer

Vox had Opus 5.5 coordinate 105 agents across five dynamic workflows (mining, planning, scaffold, build and QA) to turn 1,500+ of their posts, articles and drafts into a repo of 66 agent skills.

Result: According to Vox, the run produced a repo of 66 skills in just over 7 hours of combined run time, and none of the 24 build and QA agents threw an error. Vox says they are still checking how well the skills work.

What they built

Vox used Opus 5.5 to turn a large archive of their own writing into reusable agent skills. According to Vox, 105 agents across five dynamic workflows read more than 1,500 posts, articles and drafts and produced a repo of 66 skills sorted into 14 categories.

How it works

Vox says the work was split into five stages:

  • Mining (71 agents): read the material in batches, picked out methods that could become skills and sorted them into categories.
  • Planning (9 agents): four designers each drafted a version, a chief architect wrote the blueprint, and three critics flagged 50 problems that were fixed in the final plan.
  • Scaffold (1 agent): set up the repo skeleton and validation scripts before any other work started.
  • Build (21 agents): three lanes ran in parallel with one agent at a time in each, one writing and one reviewing.
  • QA (3 agents): each checked one area and never edited another agent's files.

Mining ran 8 agents at a time. Vox says one pass alone used about 12M tokens and that they hit their usage limit twice. Because the weekly limit was going fast, they switched build and QA to groups of three. Vox now sets each agent's effort to match its job: low for mining, medium for writing, high for review and xhigh for the final decision. They also shared a prompt template that adds done markers so reruns skip finished work. In a quoted post, they suggest having a separate subagent build an HTML progress dashboard before long tasks.

The result

According to Vox, the five workflows took just over 7 hours of combined run time, and none of the 24 build and QA agents threw an error. They say they are still checking whether the skills work well and may publish a browsable site later.

How to build something similar

  • Pick a body of material you've built up over time and decide what the output should be before you start any agents.
  • Split the job into stages, and set up the repo structure and validation checks before any agents write content.
  • Limit how many agents run at once and add checkpoint files so a run interrupted by usage limits can pick up where it stopped.
  • Use cheaper settings for mechanical work and save higher effort for review, then spot-check the output by hand before you rely on it.

The original post

I had Opus 5.5 send out 𝟭𝟬𝟱 𝗮𝗴𝗲𝗻𝘁𝘀 across dynamic workflows and turn 1,500+ of my posts into a repo of 66 skills. For step one, mining the material, I ran 71 agents, 8 at a time. One pass alone burned about 12M tokens. I hit my usage limit twice and was stuck for hours. Then I saw I was burning through my weekly limit too fast, so I switched build and QA to 𝗴𝗿𝗼𝘂𝗽𝘀 𝗼𝗳 𝟯. Not one of those 24 agents threw an error. That actually surprised me. How the work was split (5 workflows, just over 7 hours of run time combined): → Mining (71): read posts, articles and drafts in batches, picked out the methods that could become skills, and sorted them into 14 categories → Planning (9): 4 designers each drafted a version, a chief architect wrote the blueprint, 3 critics flagged 50 problems, and the fixes went into the final plan → Scaffold (1): set up the repo skeleton and validation scripts before anyone else started → Build (21): 3 lanes in parallel, one agent at a time in each lane, one writing and one reviewing → QA (3): each checked one area and never edited another's files Now I match every agent's effort to its job: low for mechanical work like mining, medium for writing, high for review, xhigh for the final decision. That said, I'm still checking whether these skills actually work well. Running this on my own account is just one example. The same approach can be applied to any kind of work that builds up over time. Once I've confirmed the material is genuinely useful, I'll probably put the site up so everyone can browse it. Send this prompt to Claude Code. Keep the word ultracode in it and it'll run as a dynamic workflow 👇 "ultracode: Turn [your material] into [what you want to make]. Work in 5 stages: mine the material, plan, scaffold, build, QA. - Run at most 3 agents at once in each group. If there's more work, queue it and start the next agent only when one finishes. - Each agent writes a done marker when it finishes, so a rerun skips work that's already done. - Effort: low for mining, medium for writing, high for review, xhigh for the final decision. - For the plan, have several agents each draft a version, then send in agents whose only job is to find problems. Show me the workflow script and who does what first. Don't run anything until I confirm." [Quoted post from @Voxyz_ai]: Every time you let Opus 5.5 run a long task on its own, have it vibe-code a quick 𝗛𝗧𝗠𝗟 𝗱𝗮𝘀𝗵𝗯𝗼𝗮𝗿𝗱 like this first. Then give it 𝗮 𝘀𝘂𝗯𝗮𝗴𝗲𝗻𝘁 𝘁𝗵𝗮𝘁 𝗼𝗻𝗹𝘆 𝗯𝘂𝗶𝗹𝗱𝘀 𝗱𝗮𝘀𝗵𝗯𝗼𝗮𝗿𝗱𝘀: → Name it dashboard-builder, set its effort to medium, and preload a design skill → The first time, it asks what style you like and saves it to memory. After that, every dashboard fits your taste and the task at hand → Call it before every long task. It builds the dashboard in the background in about two minutes, and the main session keeps coding on high without stopping → The dashboard shows four things: task progress, what's stuck, questions waiting on you, and what it'll do by default if you don't answer Send this prompt to Claude Code 👇 "Set up a subagent that only builds progress dashboards: 1. Create dashboard-builder in ~/.claude/agents: model: opus, effort: medium, memory: user. Preload the design skills I have installed (for example impeccable). It can only read and write the .dashboard/ folder and its own memory. 2. The first time, it asks me what style I like: dark or light, dense or airy, and one accent color. It saves my answer to memory and follows it every time. Pick the panels for the current task; don't use a template. 3. The dashboard is one HTML file showing tasks and their status, questions waiting for me with the default action, the latest deliverables, and anything stuck. Use the real clock for every time. It opens with a double-click and refreshes itself every 10 seconds. 4. Add a rule to ~/.claude/CLAUDE.md: for any task with more than 5 steps or that should take longer than 30 minutes, ha
Vox (@Voxyz_ai) · September 27, 2026 · View on X

Want to build something like this?

Join the Shipping Skool community free. Paid members get live bootcamp calls every week.

Join Shipping Skool — Free

More use cases

← All AI agent use cases