Detailed content
🎯 /goal
With the /goal you define one goal with a "done" criterion and Claude doesn't stop — has built-in verification — until the condition is met. The secret is a "done" as goal possible: subjectivity is its weak point. You kick it off, go to sleep or lunch, and come back to a result delivered with high confidence.
/goal pesquise X e não pare até 3 resultados que passem no critério # exemplo real: otimizar um elemento da página /goal reduza o tempo até este elemento carregar a ~instantâneo # rodou ~1h30, tentando e verificando, até bater a meta
✓ A good definition of "done"
- ✓Loads in under X ms, measurably.
- ✓3 results that pass an automated test.
- ✓Green build and all tests passing.
✗ Weak criterion (subjective)
- ✗"Really hard riddles" — what counts as hard?
- ✗"Make it look good" — with no stopping point in sight.
- ✗Any “done” that depends on taste.
💡 Practical tip
You don't need to run it for hours. Start a /goal before heading out to lunch and come back to high-confidence completed work. Spend time writing the stopping criterion — that’s what lets the agent know when it has succeeded.
Measurable criteria.
Keep trying until it works.
Built-in check.
Come back with the result.
🗺️ Ultraplan
O ultraplan does offload planning to the cloud, with multiple planning agents. You type "ultraplan," see the rainbow effect, it asks whether you want to run it in the cloud, and gives you a link. There it clones the repo, computes in the cloud, and you bring the plan back to the terminal — or run it directly on the web. Ideal for tasks large and ambiguous.
Type "ultraplan"
The rainbow effect appears, and it asks if you want to run planning in the cloud.
Clone the repo and compute in the cloud
It gives you a link; several planning agents work in the cloud on a copy of your repository.
Bring the plan back—or run it on the web
You import the plan into your local terminal, or run it directly on the web. Meanwhile, you keep working (background tasks).
💡 Practical tip
Use ultraplan especially when the task is large and poorly defined — that’s where multiple cloud planners pay off. Since it runs in the background, kick off the plan and move on to something else in the terminal while the cloud version thinks.
Offload planning.
Computes on a copy.
Import it or run it there.
Work while you plan.
📈 /insights
O /insights generates a HTML report for the last 30 days of your Claude Code usage: what you did a lot, what’s not working, features to use more, and skills worth building. It shows messages and sessions for the period, quick wins, and ambitious workflows, all with a navigable index.
Reads your last 30 days
Scans messages and sessions from the period to understand how you actually use Claude Code.
Combines what works and what gets in the way
Points out what's good, what's bad, underused features, and skills that would make a difference.
Delivers an HTML file with a navigable index
Short-term quick wins and ambitious long-term workflows, organized for you to explore.
✓ What you discover
- ✓What you do a lot (and could turn into a skill).
- ✓Features you’re not using properly yet.
- ✓Quick wins and ambitious workflows.
✗ Without /insights
- ✗Keep repeating the same friction without realizing it.
- ✗Leave valuable features sitting idle.
- ✗Never see your own usage progress.
💡 Practical tip
Run the /insights once in a while—not every day. It’s great as a periodic checkup to see your usage evolve and identify the next areas for improvement.
Browsable index.
Messages and sessions.
And ambitious workflows.
See usage evolve.
🧠 Auto Memory
Previously called "autodream," Auto Memory is Claude automatically improving memory without you asking. At intervals between sessions, conversations, or periods of time, it reviews what was done and updates itself — “dreams” and gets smarter, per repository.
🌙 How the “dream” works
Memory stops being something you maintain by hand: the agent automatically updates what it learns about your project.
- •Runs at intervals — from sessions, conversations, or time.
- •É per repository: each project builds its own memory.
- •The more you use it, the smarter it gets in that context.
/memory # veja se a auto memory está ligada; ative se quiser. # há docs sobre como o processo funciona.
💡 Practical tip
Check in /memory whether auto memory is enabled in your main repository. When enabled, it works in the background: each time you come back, the agent already knows a little more about the project.
Without you asking.
Sessions, conversations, time.
Project memory.
Check and enable.
👥 Agent Teams
One of the favorite features — and token-intensive. Agent Teams is a team named of agents running in parallel and talk to each other: you create personas (beginner, CEO, etc.); they analyze, brainstorm, and debate in rounds until you agree. It’s the best way to hear different perspectives on the same idea.
✓ Good use of Agent Teams
- ✓Brainstorm with personas (beginner, CEO, skeptic).
- ✓Debate the pros and cons until you reach a consensus.
- ✓Hear different perspectives on a decision.
✗ Don’t confuse
- ✗They're not subagents — these talk with each other.
- ✗It's not cheap: it uses a lot of tokens.
- ✗It’s experimental—you need to enable it in settings.
# habilite o recurso experimental no settings (.claude) # depois, invoque em linguagem natural: monte um time de agentes — um iniciante, um CEO e um cético — e debatam essa ideia em rodadas até concordarem
💡 Practical tip
Because Agent Teams are token-intensive, save them for decisions that truly benefit from debate. Give them personas with very different points of view—the friction between them is exactly what leads to the best perspectives.
Named personas.
Rounds until they agree.
Use selectively.
Enable it in settings.
⏪ /rewind
O /rewind is the time machine: it brings back the code AND the conversation to an earlier checkpoint, and you choose the exact prompt or point. Since everything is cached and reloaded, rewind when something went wrong, it’s better than saying “no, do it this way” — it avoids piling up messy context and preserves quality.
✓ Rewind (better)
- ✓Go back to the exact point before the error.
- ✓Clears the messy context from a bad attempt.
- ✓Code and conversation return together, consistent.
✗ “No, do it this way” (worse)
- ✗Stacks one error on top of another.
- ✗Carries polluted context forward.
- ✗Quality drops with every correction.
/rewind # escolha o prompt / checkpoint exato para onde voltar. # código e conversa retornam juntos àquele estado.
💡 Practical tip
When a response went in the wrong direction, the right instinct is to /rewind — don't patch over it. Returning to the checkpoint keeps the window clean and often produces a better result than trying to fix things in a polluted context.
They come back together.
Choose the prompt.
Without piling up clutter.
Better than fixing it.
🏁 Module summary
Next module: 3.2 — The podium (#6 → #1)
We’ve seen the first half of the Top 12. Now step onto the podium: from Subagents (#6) to the #1 feature of all, Skills.