Lesson 3 · 15 min
A finished script can become a summary, clips, and posts. Asking for everything in one conversation fills it up and mixes the parts together. With helpers, the conversation splits up the work: each helper handles one part and returns only the result. In this lesson, Carla does this with lesson 1 of the spreadsheet course, and you try it with a text of your own.
Start the lesson
Goal: Split a big job among helpers (subagents) and check the result before using it.
Tap or click each step.
Scroll down: each step below lights up its part of the diagram.
You hand over the whole job at once, already saying what the parts are. Carla asks for a summary, 5 short clips, and 3 posts based on the lesson 1 script.
A session is a Claude conversation with its own history. It receives the request, separates the parts, and assigns each one to a helper. You don’t open anything else: the helpers work inside that same session.
Helper is the name this course uses for a subagent, its official name. Each one handles only its own part, in its own context (the memory for that task), and returns a summary to the conversation. That way, the main conversation doesn’t fill up with everyone’s drafts. Many people call this “nested sessions,” but that isn’t the official name.
The conversation compares the three parts with the script: timestamps, names, numbers, and tone. It corrects and notes anything that doesn’t match. Without this step, a helper’s mistake goes straight to you.
You get the three files in the folder and a short list of what was corrected. Now it’s your turn to check: open a file and verify one detail against the script.
You ask for the whole job, already split into parts. The conversation gives one part to each helper, combines the responses, checks them against the original, and returns the files with a list of what it corrected.
Carla’s example: lesson 1 script → summary, clips, and posts. Helpers are available in the app (Code tab) and in the terminal.
Four things to keep in mind. Click each one.
Helpers work well when each part can move forward on its own. The summary, clips, and posts all come from the same script, but none has to wait for the others.
If part B depends on the result of part A, do A first and ask for B afterward.
A summary, clips, and posts from the same lesson. Or: proofread 3 different lessons.
Say what each helper will deliver, in which file, and how long it should be. “Make some posts” has no endpoint; “3 posts of up to 280 characters in posts.md” does.
With a clear deliverable, the review knows what to look for.
Helper 2: 5 short clips, each with a start, end, and hook sentence, in cortes.md.
Ask the conversation to check the parts against the original and list what it corrected. If there’s nothing to correct, it has to say how it checked.
The list is your shortcut: look first at what was corrected.
At the end, check the three files against the script and give me a list of what you corrected.
Each helper has its own context, which counts toward your plan limit. Start with 2 or 3 helpers, not 20.
In the app, the usage ring is a circle next to where you choose the model. It shows how much of your limit you’ve used. In the terminal (optional), /usage shows your plan usage and /context shows how much of the conversation is already in use.
Coordinating several sessions at the same time is another topic: see the course Claude as an agent manager.
The first is Carla’s request, as a list. The second is the template for you: the same request, with one field per line. It passes the checker below and works when pasted straight into the conversation. Replace what’s between < > with your own text, and replace the curso-planilhas/… path with the path to your file (.md, Word, or PDF; see lesson 2).
Read the script for lesson 1 in curso-planilhas/aula-1/roteiro.md. Use one subagent for each item below. Each one does only its part and saves the result: 1. A summary of the lesson, up to 150 words, in curso-planilhas/aula-1/resumo.md 2. 5 short clips for the channel, each with a start, end, and hook sentence, in curso-planilhas/aula-1/cortes.md 3. 3 short social media posts, each up to 280 characters, in curso-planilhas/aula-1/posts.md At the end, check the three files against the script (timestamps, function names, numbers). Correct anything that doesn’t match and give me a list of what you corrected. Don’t publish anything or change the script.
Task: <what to do and where the original text is, e.g., folder/text.md> Helper 1: one subagent does <deliverable 1, with length> and saves it in <file1.md> Helper 2: one subagent does <deliverable 2, with length> and saves it in <file2.md> Helper 3: one subagent does <deliverable 3, with length> and saves it in <file3.md> Review: at the end, check the three files against the original text and give me a list of what you corrected. Limits: don’t publish anything or change the original text.
Done. The three subagents finished, and I checked everything against the script.
Files
What I corrected
Nothing was published. The script wasn’t changed.
/usage in the terminal) to see how much the work used.Write your request in this format (one field per line) and paste it here. The page checks whether each helper is a subagent with a delivery file, whether the Review compares with the original and asks for a list of corrections, and whether the Limits say what not to do. This is a format check: you decide whether the parts make sense.
Use a text you already have: a script, lesson, article, or meeting notes. In the Desktop app (Code tab) or in the terminal; on Linux, terminal only.
Choose the text and the three deliverables (e.g., summary, clips, and posts; or summary, quiz questions, and an email to students). Fill in the pedido-ajudantes.txt template. If your text has no timestamps, ask for “the excerpt” instead of a start and end time.
Done when: the checker above responds, “OK: request has the 6 fields, one subagent per part, and review with a list.”
Open a session in the text’s folder (in the app: Code tab, choose the folder when opening the session, as in lesson 2; in the terminal, optionally: go to the folder and type claude). Paste the pedido-ajudantes.txt request that passed the checker, exactly as it is, and wait for the final response.
Done when: the 3 files you requested are in the folder, and the final response includes a list of what was corrected (or says how it checked if nothing needed correcting).
Pick one item from the corrections list and check it against the original text. Then open one of the files and verify one complete piece of information (a clip, a date, a name). Finally, look at the usage ring in the app (the circle next to the model selector) or type /usage.
Done when: you checked 2 pieces of information against the original and looked at the usage ring or /usage to see how much of the limit the work used.
The explanation appears after your answer.
Question 1 of 5
What is a helper (subagent) in Claude Code?
Question 2 of 5
Which of these tasks is a better fit for helpers?
Question 3 of 5
Why ask for a review at the end?
Question 4 of 5
In the request, which Review line does the page accept?
Question 5 of 5
Carla wants to use 20 helpers at once. What’s the best advice?
Choose an option. The explanation appears after your attempt.