Virtual assistance

Virtual assistant skills, and the one that sets your rate

Four skills make this work: writing, order, enough tool fluency, and discretion. Of the four, the one that sets your rate is writing, because it is the only one a client sees before handing you any work.

  • Lesson 2 of 6
  • Beginner
  • Free, no signup

The four skills this work stands on

Their order is not accidental. The first is what a client sees before anything else, which is why it sets the rate.

An assistant who gets hired againstands on four columns

  • Writing

    A message that makes the next question unnecessary. Three things in every update: what is finished, where it is stuck, what it needs from you.

  • Order

    One list rather than four, every task with a date, and defined checking windows instead of being permanently open.

  • Enough tool fluency

    Shared documents, a task list, a calendar. And one rule bigger than the tools: the client tool wins.

  • Discretion

    Rules rather than intentions: no passwords in chat, no identities leaving the room, no client data in a tool without permission, access kept temporary.

These four hold the work up but they do not bring clients. Finding clients and pricing are a different kind of skill and they are lessons five and six of this path.

Last checked: Facts and tool names in this lesson are re-checked against their sources on this date.

Why does writing set your rate?

Before anyone has seen your work, they have seen your writing. The short profile, the first message, the answer you gave to a question about price: that is everything the client holds while guessing how much of their time handing you work will cost. And it is exactly that guess which sets the rate, not typing speed and not the number of tools you can name.

The practical test is one sentence long: a message you send should not force the client to ask the next question. "Done" is an incomplete message. "The supplier list is ready, seven of them with prices, two never replied; the file is in yesterday's folder. I would like your go-ahead to chase those two" is a complete one. The difference is not length, it is that the second leaves no question open.

Three things belong in every update and most people drop one of them: what is finished, where it is stuck, and what it needs from you. The one usually dropped is the third, and that is the one that turns a message into an extra round trip.

One clarification is needed: writing well here does not mean writing elegantly. Long sentences and long courtesies spend the reader's time, and time is exactly what you are selling in this job. If you want to work on the writing itself, the lesson on where content production starts is a good entry point, as long as you remember that here the audience is one person rather than a market.

Four pillars holding a glass roof: a tall white pen, a red clock, a green tool and a blue padlock

An update that creates no question, against one that adds to the client work

One status updateBefore sending

An update that leaves no question open

  • Says what is finished and where it is
  • Says what is stuck and why
  • Says exactly what it needs from the client
  • If a decision is needed, it lays the options side by side

An update that creates an extra round trip

  • Says only that it is done
  • States a problem and proposes nothing
  • Mixes three unrelated subjects into one message
  • Opens with courtesies and buries the news at the end

This test is for a status update, not for every message. Where the news is bad or the subject is sensitive, a short complete message is not enough and a call is better.

What does order mean when nobody is watching you?

In an office the building imposes order on you: an arrival time, a desk, somebody walking past to ask what happened to that task. Remote work has none of it, and the thing that looks like freedom is the first place people fall over. What replaces the building is three habits, not a piece of software.

First habit: one list, not four. A task mentioned on WhatsApp, a task that arrived by email and a task you thought of yourself must land in the same single place that same minute. People whose tasks are scattered across three places do not spend their time doing the work, they spend it remembering the work.

Second habit: every task carries a date. A task with no date is not a task, it is a wish. And set the date yourself even when the client did not; if you do not set it, the client has set one in their head and you do not know what it is.

Third habit: a fixed window for checking. Opening the inbox every twenty minutes is not order, it is regular distraction. Two or three defined windows a day, stated in the agreement, is better for you and for the client, because it makes the other side's expectation real.

And a trap plenty of people fall into: building the system can itself become a new form of putting things off. Someone who spends three days colour-coding a task board delivered nothing in those three days. The simplest system you will actually keep beats the best one you will abandon.

How fluent do I need to be with the tools?

Less than you think. Three groups of tools are genuinely needed and the rest arrive by themselves later: somewhere for shared documents, somewhere for the task list, and a calendar. If you are comfortable with those three, you have cleared the first week.

But there is a rule that matters more than the tools themselves: the client's tool wins. If the client's team works on Telegram, you work on Telegram, even when your own favourite is better. An assistant who makes the client migrate to their system has increased that person's workload in week one, which is the exact opposite of what they were supposed to be selling.

The skill nobody puts in a job ad and which makes the practical difference is keeping naming clean: files and folders named so that they still make sense six months later, and every task having one definite place on the board. It is the kind of thing whose absence nobody can point at when everything has slowed down. Listing and comparing the tools is the job of lesson three in this path, and we are not repeating it here.

Discretion is a skill, not a feeling

Being trustworthy is a trait, and a trait cannot be practised. What can be practised is a handful of definite rules, and those are what make an assistant safe to hire, not their good intentions.

First rule: passwords are not exchanged in chat. If the client insists, offer the alternative; almost every modern service lets them create a separate account with limited access instead of handing over a password. If they insist anyway and the password arrives, the least you do is ask them to change it once the work is done, and say so in writing. The lesson on what cybersecurity is opens up the background.

Second rule: anything that identifies a person does not leave the room. A customer's name and number, a purchase history, the text of somebody's complaint. You do not put these in a portfolio, you do not retell them in a group of friends, and you do not pour them into any tool. If you want a work sample, strip the numbers and names and show the structure of the work instead.

Third rule, and the one broken most often today: client data does not enter any AI chat without the client's permission. Whether conversations are used to train a model depends on the plan and the settings rather than on your instinct; the detail, with its source, is in the block further down this page. Until the client has explicitly said which tool is allowed, the default answer is no.

Fourth rule: access is temporary. When the work ends the access should end, and it is better that you raise it rather than waiting for the client to remember. An assistant who says close my access now, unprompted, gets access more easily next time.

Four discretion rules you can actually practise

Being trustworthy is a trait and cannot be practised. These four are rules, and they can.

  • No password is exchanged in chat

    Offer the alternative: a separate account with limited access. If a password arrives anyway, ask for it to be changed afterwards and say so in writing.

  • Identities do not leave

    Customer names and numbers, purchase histories, the text of a complaint. Not in a portfolio, not in a group of friends. Show a sample with the names and numbers removed.

  • Client data enters no tool without permission

    Whether a conversation is used to train a model depends on the plan and the settings. Until the client has said which tool is allowed, the default answer is no.

  • Access is temporary

    When the work ends the access ends, and it is better that you raise it. That single move makes getting access easier next time.

These rules do not replace a contract and they are not legal advice. A confidentiality undertaking, where sensitive data is involved, is something to put in writing, and its place is lesson six of this path.

Where can I practise these four cheaply?

None of the four needs a class, and three of them can be practised on your own work before you have a single client.

Practise writing on your own messages. Take the last ten you sent this week and rewrite each one so that it leaves no question open. If you want a harder drill, the recipe further down this page was built for exactly that.

Practise order on a real small business, even one belonging to somebody you know and with no money involved. Running a shop's calendar and message queue for a month teaches more than any course, because it is the only place you find out what it feels like when three things run late at once.

Discretion cannot be practised on other people's work, but it can be decided before the first client: write down what you will never put anywhere and which tools you will not open without permission. Half a page you wrote yourself is the thing that stops a rushed decision when the first client pushes.

And the tools are the one item that genuinely gets learned later. When you get to them, the virtual assistance path has the next lesson. And if what you want first is to see what the job itself looks like, the lesson on what a virtual assistant is is the place to start.

The fast path, with AI

The usual advice about writing is to write clearly, which teaches nothing because no measure comes with it. The fastest route we know is a forced comparison: write one real message three ways yourself, then ask the model to analyse only, not to rewrite. The model is a mirror here, not an author, and that constraint is the difference between learning something and being handed a text.

  1. Pick a real message you sent this week and strip out names, numbers and amounts. Do not put another client message anywhere for the sake of your own practice.
  2. Write three versions yourself: one very short, one complete with every detail, and the one you actually sent. Writing all three is half the drill by itself.
  3. Give the recipe below along with all three and let it analyse. A frontier model does better here because the pass is a judgement call rather than a mechanical one; our current pick stays up to date in the RGB AI reference.
  4. After reading the analysis, write a fourth version yourself and keep that one. What you learn is that fourth version, not anything the model said.

Copy-ready recipe

I am a virtual assistant practising how I write work messages. I am sending three versions of one real message. Names and numbers have already been removed.

Context: the reader is a {client role} and the subject is {subject}. What I want to happen is {desired outcome}.

Version A: {text}
Version B: {text}
Version C: {text}

Rule: do not rewrite any of them and do not produce a fourth version. Analyse only.

1) For each version, write what question the reader still has after reading it. If there is none, say there is none.
2) For each version, say what it signals about its writer: confident, hesitant, rushed, or vague, with one piece of evidence from the text itself.
3) Say which sentences could be deleted without losing anything.
4) Say whether each version has or lacks the three things: what is finished, where it is stuck, what it needs from the reader.
5) Finish with one sentence: if I were to change only one thing in the next version, what is that one thing.

Before you trust the output: Take two limits of this drill seriously. First, if you let the model rewrite, you learn nothing and you get a text that reads like everyone else's; a client who works with you every week notices that sooner than you would expect. Second, the model does not know your client's tone or what was agreed verbally, so its verdict of vagueness is sometimes about something the other side already knew. And never put a client's real message anywhere without stripping names and numbers first; that single rule is the difference between practice and a leak.

AI in this kind of work

In this lesson the model has one role and does not have another. The role it has is being a mirror: telling you what a sentence signals, what question it leaves open, and where it is padded. The role it does not have is writing in your place. An assistant whose messages are written by a model has delegated the one thing that was supposed to raise their rate.

Tools that actually help

  • Claude In this drill it holds the do not rewrite constraint well, and that constraint is the whole value of the exercise. Iran is on neither of Anthropic two supported country lists; we read that on Anthropic own page rather than measuring it.
  • Gemini It understands Persian itself, so three Persian versions need no translation first and the tone does not get lost on the way. Google own page says the Gemini web app runs in over 230 countries and territories, and Iran is not on that list.
  • ChatGPT It works for this analysis too and most people are already comfortable with it. We make no claim about access from Iran: the OpenAI supported countries page returns 403 to this server, and we do not write a claim with nothing behind it.

Where it backfires

The first risk in this lesson is the very thing that destroys the drill: putting the model in your place. If a model writes your messages, your own voice never forms, and after a few weeks the client notices that every message has the same shape. This is not a claim about automatic detection; it is a simpler point: your client knows you through your text, and if the text is not yours there is nothing left to know. The second risk is more serious and belongs to the whole job: a client message is client data. Whether your conversations are used to train a model depends on the plan and the settings; on its own privacy page Anthropic writes that in its consumer products it uses chats to improve models when you have chosen to allow it, when a conversation is flagged for safety review, or when you have explicitly opted in. So the practical rule is simple and does not hang on any setting: strip names, numbers and amounts before sending, and if the data belongs to a client, do not put it anywhere without their permission. That is why asking about tools belongs to the first week rather than to whenever it comes up.

Sources: Anthropic: supported countries and regions Anthropic privacy: is my data used for model training Google: where you can use the Gemini web app

Where this advice stops

These four skills hold the work up, but they do not bring clients. Someone who has all four and no client does not have a problem this lesson solves; lesson five does. Pricing and contracts are not here either, and their place is lesson six. One more boundary, stated plainly: the discretion rules written here are working habits, not legal advice. If what you do touches sensitive data or a client's financial information, a confidentiality undertaking has to be written down, and its text should be drafted by somebody whose job that is rather than by a teaching page. And last: the drills in this lesson are designed for your own work and for a business you already know; none of them replaces real experience with a real client, and we do not claim otherwise.

From our own work

We had to enforce the rule of writing for the reader rather than from your own position inside our own code, because saying it was not enough. On this site freelance platform, the text of a notification used to be built in whatever language the current request was in, which in practice meant the language of whoever triggered it. The result was that a Persian-speaking client could receive a notification in another language purely because the freelancer had opened the site in that language. Now rgbf_send_notification builds the text inside rgbf_with_user_lang: the language is chosen from the recipient and then restored, and the explanation sits above that function in rgb-freelance/includes/notifications.php. The same rule holds in its human form for a virtual assistant: write the message in the language and at the level of knowledge of the person who will read it, not of the person who did the work.

Real follow-up questions

Does touch typing matter for this work?

It helps and it decides nothing. What sets the rate is that your message needs no follow-up question, not how fast it was typed. Someone who writes one complete message in five minutes has spent less of the client's time than someone who writes in one minute a message that takes three round trips.

The client wants to send me their password on WhatsApp. What do I do?

Offer the alternative rather than only saying no: almost every modern service lets them create a separate account for you with limited access, and that account can be closed later. If they send it anyway, ask them to change the password once the work is finished and put that in writing. It is not rude; in practice it makes the client take you more seriously.

Can I use AI to write my client messages?

For drafting and for analysis yes, for sending without reading no. Two practical conditions come with it: names, numbers and amounts are stripped first, and the client knows which tools you use. If you do not rewrite the message yourself before sending, after a few weeks you will have no consistent voice, and a consistent voice is exactly what a long-term client leans on.

Keep learning