Detailed content
🌊 Dynamic Workflows
Dynamic Workflows is when Claude writes a script which runs — in extreme cases — for up to 1,000 agents to handle enormous jobs. Instead of processing everything in one queue, it splits the work up in parallel and brings it all together at the end. It’s the foundation that shows up again in other features in this module (deep research, ultrareview).
🧩 How it divides the work
The secret isn’t “one giant agent,” but many small agents coordinated by a script that Claude writes on the fly.
- •Divisible tasks (scanning thousands of files, processing batches) are a perfect fit.
- •Handles parallel processing very well—the total time drops dramatically.
- •It’s new: it’s still finding its place in your day-to-day work.
💡 Practical tip
How to use it many tokens, save dynamic workflows for truly large, divisible tasks. For something small, a single agent or a few subagents are enough.
Claude writes it on the spot.
Parallel agents.
Use for large tasks.
Still finding its place.
🌿 Git Worktrees
When several agents work in the same repository at the same time, the risk is obvious: collision. Git Worktrees solves this by running each agent in a isolated copy of the repo, so they never step on each other’s work, and merges everything at the end. It's the natural complement to dynamic workflows.
main ├─ worktree/feature-a agente 1 (isolado) ├─ worktree/feature-b agente 2 (isolado) └─ worktree/fix-c agente 3 (isolado) # cada um trabalha sem colidir → merge de volta na main
✓ When it’s especially worthwhile
- ✓Working as a team in the same repo.
- ✓Multiple features being worked on at the same time.
- ✓Parallel agents that can't get in each other's way.
✗ What to avoid
- ✗Multiple agents editing the SAME branch live.
- ✗Expecting parallelism without isolation leads to conflicts.
- ✗Forgetting to merge at the end and losing the work.
One worktree per agent.
No one steps on anyone else.
Everything comes back together.
Multiple features at once.
🛡️ Ultrareview
Ultrareview is a slash command which brings up a cloud agent fleet (in the style of a code review) to review your code. The difference: each agent checks every bug before reporting it, which reduces the noise from false positives that plagues ordinary automated reviews.
You trigger the command
Runs the slash command on the diff/project; the work goes to the cloud instead of blocking the terminal.
The fleet checks in parallel
Multiple agents analyze at the same time, and each suspected bug is checked before it becomes a report.
You get what matters
Instead of a noisy list, you get verified findings—more signal, fewer false positives.
💡 Practical tip
Because it runs on the cloud, you launch ultrareview and keep working in the terminal. Think of it for the heavy review before an important merge.
One command kicks everything off.
Doesn't block the terminal.
Every bug is checked.
More signal per finding.
🔬 /deep-research & Advanced Research
You ask for research and it automatically invokes a dynamic workflow, fans out web searches (and in your connected apps) and returns A cited report. It's the parallelism from topic 1 applied to research: dozens of searches at once, synthesized into a response with sources.
📚 What happens under the hood
Instead of one search at a time, it orchestrates many in parallel and only then drafts—with references.
- •Invokes a dynamic workflow without you needing to ask explicitly.
- •Searches the web and your apps/connectors too.
- •Delivers a single report cited, not 30 tabs for you to read.
💡 Practical tip
Treat it as a real research: ask for scope, guiding questions, and preferred sources. The clearer the question, the better the cited report you get back.
Invokes a dynamic workflow.
Searches wherever you connect.
Cited, not 30 tabs.
Same engine, more depth.
🖱️ Claude in Chrome & Computer Use
Claude in Chrome is a side panel that clicks, fills out forms, and acts across tabs from the browser. The Computer Use is the generic point-and-click option: it uses any app on screen when there’s no connector. Together, they extend Claude’s reach to what lives outside the terminal.
✓ Claude in Chrome
- ✓Click and navigate between tabs as you would.
- ✓Fills out forms in the side panel.
- ✓Great when the entire workflow lives in the browser.
✓ Computer Use
- ✓Point and click in any app.
- ✓Universal fallback when there’s no connector.
- ✓Uses the screen as a human would.
💡 Practical tip
Prefer connector when available (faster and more reliable); use point-and-click as a last resort for apps without an integration.
Works inside Chrome.
Click and fill it in.
Point and click anywhere.
Universal fallback.
🧱 The B ecosystem
Several Tier B features don’t shine on their own, but shape your day-to-day work: reuse context, help you pick up where you left off, and give Claude new ways to act. They’re all worth a quick look.
Projects workspaces: chats + arquivos + instruções juntos Memory carrega contexto de chats antigos /recap mostra o que a sessão fazia quando você volta Artifacts renderiza num painel editável (+ Interactive Charts) # e ainda: Browser Recording · Claude for M365 · Mobile Push # Cross-Device Convo · Connectors Directory
✓ What they’re for
- ✓Projects: everything for a project in one workspace.
- ✓Memory + /recap: resumes context without re-explaining.
- ✓Artifacts & Charts: editable panel and charts in the chat.
✗ Common misunderstandings
- ✗Thinking that /recap "remember everything" — it summarizes the session state.
- ✗Confusing Memory with manually saving files.
- ✗Skip the Connectors Directory and reinvent integrations.
💡 Practical tip
O /recap is especially great with many tabs open: when you return to a session, it shows you what was in progress before you continue.
Everything-in-one workspace.
Context from old chats.
Picks up where you left off.
Editable dashboard + charts.
🏁 Module summary
Next: the honorable mentions
Module 2.2 covers Tier A — the features that almost made the Top 12, from Google Workspace in the terminal to the visibility and cost-saving layer.