AfterDesk
EN
Apply

The toolkit

AI tools, used responsibly

AI can genuinely speed up drafting, summarising, reformatting and first-pass formulas, and it can also invent facts, leak client data and cost you a client in a single delivery. This course teaches where the line sits, how to verify what a model produces, what never goes into a third-party tool, and how to prompt, check and disclose like a professional. Everything you deliver stays yours.

8
lessons
~58
minutes
12
exam questions

Free · No paid tier · No certificate fee

After this course

Everything, and what is in it.

Where AI helps and where it fails

~8 min

The honest split

AI writing tools are now part of ordinary virtual assistant work. Pretending otherwise helps nobody. What matters is knowing precisely which parts of a task a language model can carry and which parts it will quietly ruin. The split is not about difficulty. A model will happily produce a polished two hundred word summary of a document it read correctly, and with the same confidence produce a company registration number that does not exist. The dividing line is whether the answer depends on a fact in the world, or on shaping text you already have. Shaping text is where it is strong. Supplying facts is where it is weak. Hold that line in your head while you work and most of the danger disappears.

What it genuinely does well

The strong uses share a shape: you supply the material, the model reshapes it. Turning your rough notes into a clean paragraph. Compressing a document you already have into a summary. Rewriting a blunt sentence so it reads politely. Producing eight subject line options so you can pick one. Converting a block of text into a table, or a table into sentences. Writing a first-pass spreadsheet formula from a description of your columns. Explaining what an unfamiliar formula does, piece by piece, so you can decide whether to trust it. Suggesting a structure for a report before you write it. In every one of these, the truth of the content comes from you or from the source in front of you. The model is doing carpentry on material you already own.

Where it reliably fails

The weak uses share the opposite shape: you ask the model to supply a fact. Prices, dates, phone numbers, email addresses, job titles, company registration details, statistics, study results, quotations, links and citations. Anything about a real person or a real business. Anything that changed recently. It will also miscount and mis-add across a long list, because it is not calculating, it is writing what a calculation tends to look like. None of this arrives with a warning. You get a clean, formatted answer in the same voice it uses when it is right. Treat every fact that originates inside the model, rather than inside your source material, as unverified until you have checked it against something real.

Why the failures look convincing

A language model produces the next piece of text that fits. Fitting and being true are different properties, and they come apart most often exactly where you need precision: a specific number, a specific name, a specific web address. The model has no separate sense of certainty it can show you. It cannot tell you that it invented the citation, because from the inside, generating an invented citation and generating a real one are the same operation. This is why fluency is a trap. We are all trained to read confident, well-structured writing as informed writing. With these tools, the polish tells you nothing about the accuracy. Read the content, not the surface.

What the failure costs

Here is how it usually goes. A list building job, forty companies, six contacts that could not be found. The model offers six plausible addresses. They look right. Three of them bounce when the client runs the campaign. The client does not conclude that a tool made a mistake. The client concludes that the work cannot be trusted, and starts checking everything, which is more expensive than doing it themselves. On AfterDesk the operator reviews every delivery before the client sees it, so an invented row usually comes back to you as a revision rather than reaching the client. That protection is real, and it is not something to lean on. Deliveries that keep coming back are noticed.

Remember

  • AI reshapes material you supply. It does not reliably supply facts.
  • Fluent output and accurate output look identical. Polish proves nothing.
  • Treat any fact that originated inside the model as unverified.
  • Long-list arithmetic and counting are weaknesses, not strengths.
  • One invented row costs more trust than the whole file earned.

Hallucination and the verification habit

~8 min

What a hallucination is

The word is unfortunate, because it suggests a malfunction. Producing text that is fluent, specific and false is not a bug in these systems. It is the same process that produces text that is fluent, specific and true. When you ask for a source, the model writes something with the shape of a source: an author, a plausible year, a title made of words that belong together. Sometimes that shape lands on a real paper. Sometimes it does not. The same applies to a statistic, a contract clause, a product specification, a person's job title. Once you see this as normal output rather than an error state, you stop being surprised by it and start building the habit that handles it.

The verification rule

Here is the rule, and it is short. Every factual claim that came from the model, rather than from your source material, gets checked against something outside the model before it goes into a deliverable. Names, numbers, dates, prices, quotations, links, anything with a source attached. There is no category of fact small enough to skip. The invented ones are usually the small ones, because small facts are the ones nobody feels the need to check. If you remember one sentence from this course, remember that a model is an excellent drafting tool and a research assistant who cannot be trusted about the research.

How to check quickly

Verification does not have to be slow. Go to the primary source rather than a page quoting it: the company's own site for a company fact, the official statistics agency for a statistic, the report itself for a study. Open the link with your own eyes rather than trusting that a link exists. If two independent sources disagree, you have found a real problem worth flagging, not a coin to flip. Check the date on everything, because a true fact from four years ago can be false today. For a person's role or a company's address, one authoritative page beats five blog posts that copied each other.

When you cannot verify

Sometimes the fact is simply not findable in the time you have. That is not a failure, it is information, and the professional move is to say so. Leave the cell blank rather than filling it. Mark the claim clearly in your draft as unverified. On AfterDesk, write it in the note to the operator with your delivery: which item, what you tried, what you could not confirm. An honest gap is a small cost the operator can absorb. A confident invention is a cost that lands on the client, with your delivery attached to it.

Where the risk is highest

Some work is more dangerous than other work. Research summaries, list building, anything with citations, and anything touching money, tax, law or health. In those last areas, be strict with yourself. Do not let a model produce rates, thresholds, filing deadlines or dosages that end up in a client document. If a task needs that kind of content, take it from the official body that publishes it, note the date you read it, and say plainly in your delivery that the client should confirm it with a qualified professional. We are not qualified to give that advice, and neither is the tool.

Remember

  • Hallucination is normal output, not a malfunction. Expect it every session.
  • Every model-sourced fact is checked outside the model before delivery.
  • Open the link yourself. A cited source is not a real source.
  • An honest gap costs far less than a confident invention.
  • Never let a model produce rates, deadlines or dosages.

The confidentiality line

~7 min

On AfterDesk, the default is never

Start with the part that has no nuance. On a AfterDesk task, client files, client text, client names and client data do not go into a third-party AI tool. Not a chat assistant, not an online transcription service, not a document summariser, not a writing tool that uploads the whole file, not an image tool. The default is no. The only exception is a brief that explicitly says otherwise, in writing, in the task itself. Silence is not permission. A brief that does not mention AI is a brief that does not allow it. This is not a suspicion about you. It is the promise the client was given when they handed over their data, and everyone who touches the task keeps it.

Pasting is a transfer

It helps to stop thinking of the paste as a keystroke and start thinking of it as sending. When you paste a client's customer list into a website, you have moved that data onto a company's servers in another country. It may be stored. It may be reviewed by people. It may be used to improve a product. Free tiers in particular often keep and learn from what you send. Whether or not anything visibly bad happens, you have made a decision the client never agreed to, using data that was not yours to move. That is the breach, not the outcome. Your intent does not enter into it, and neither does how careful the tool's marketing sounds.

What counts as client data

More than you would guess. Obviously names, email addresses, phone numbers, home addresses, invoices, bank details, salaries, customer lists and internal documents. Also the less obvious things: the client's business name paired with a problem they have, an unreleased product name, pricing they have not published, a screenshot of their dashboard, a file name that reveals a deal, the raw text of an email even with the sender removed. Deleting names is not anonymising when the remaining detail identifies the business. Ask one question before you paste anything. If the client read a transcript of what I just sent to this website, would they be comfortable? If you have to think about it, the answer is no.

How to get help without sending data

Almost everything you want from the tool can be had without the data. You need the method, not the material. Rebuild the shape of the problem with invented content: the same column headers renamed, four fake rows, a made-up company called Acme with made-up figures. Ask about the technique in the abstract. How do I match two lists on partial names. How do I structure a comparison table for a supplier review. What is a clean way to phrase a bad-news paragraph. Then apply the answer to the real file yourself. This costs an extra three minutes and removes the risk completely. Build the habit while the stakes are low, so that under deadline pressure your hands do the safe thing automatically.

After the task, and other clients

Two more edges. When a task is delivered, delete your working copies. Downloads folder, desktop, any temporary spreadsheet, anything you sent to yourself. You keep no copy after delivery, and you do not reuse client work as a portfolio sample, on AfterDesk or anywhere else. Build samples from invented data instead. Outside AfterDesk, some of your own clients will actively want you to use AI. Good. Get it in writing and get it specific: which tools, which data, whether output can be reused, whether disclosure is expected. A verbal yes is worth writing down in an email and having them confirm. Permission that is not written is permission you cannot prove.

Remember

  • On AfterDesk the default is never. Silence in a brief means no.
  • Pasting into a website is sending data to a company.
  • Ask for the method, not with the material. Rebuild it with fake rows.
  • Delete working copies after delivery. Client work is never a portfolio sample.
  • Outside AfterDesk, get AI permission in writing and specific.

Prompting that actually works

~8 min

Four things every good prompt has

Context, task, constraints, example. Context is who this is for and what it belongs to: a follow-up to a supplier who has missed two deadlines, in a relationship the client wants to keep. Task is the single thing you want produced. Constraints are the boundaries: ninety words maximum, no apology, one clear request, plain sentences. Example is a piece of writing that already sounds the way you want, pasted in with the instruction to match its register. Most bad output comes from prompts that contain only the task. Write a follow-up email gives you the average of everything ever written under that heading, and that average is exactly the bland text everyone now recognises.

Give it the material

A model cannot use what it does not have. If you want a summary that reflects your notes, paste your notes, not a description of them. If you want a rewrite in a client's public voice, paste two paragraphs they already published. If you want a formula, describe every column by header, position and content. This is where the confidentiality line does its work. You supply material that is yours or already public, and you rebuild anything confidential as an equivalent with invented content. Given real material, output improves immediately, because the model is doing what it is good at, working on what is in front of it, instead of what it is bad at, inventing what is missing.

Constraints beat adjectives

Make it shorter is weak. Cut to sixty words is strong. More professional is weak, because a model's idea of professional is corporate padding. Remove every adjective, no sentence over twenty words, never use leverage or seamless is strong. Numbers, banned words, required words, output format, reading level, point of view. The rule of thumb is that any instruction you could not check afterwards is an instruction the model cannot follow reliably. If you cannot tell whether it obeyed, it does not know either. Write constraints you could verify with your own eyes in ten seconds, and you will get output you can verify in ten seconds.

Iterate, do not restart

When the first output is wrong, most people delete everything and rewrite the prompt from scratch. Continuing is usually faster. Say what was wrong and what to keep: the structure is right, the second paragraph is too soft, rewrite only that paragraph and leave the rest alone. Three rounds of correction beat one perfect prompt, and each round teaches you something about how to ask next time. Two habits help. Ask for several options rather than one answer while you are still exploring, then narrow. And when a thread has drifted into confusion, start fresh with the best version so far pasted in as the starting point, rather than dragging the confusion along.

Keep the prompts that worked

The prompts you use most are the ones you will use for years. Keep them in a plain document, one per task type, with the constraints already written in: your invoice reminder prompt, your meeting-notes-to-summary prompt, your formula prompt with your usual column conventions. Add a line each time you learn what to include. This file is one of the few genuinely compounding assets in this work, and building it costs nothing beyond the discipline to paste. Keep it free of any client detail, so the file itself stays safe to keep, safe to reuse across clients, and safe to open in front of anyone.

Remember

  • Context, task, constraints, example. Task-only prompts produce averages.
  • Write constraints you could check yourself: word counts, banned words, format.
  • Correct the draft in the same thread instead of starting over.
  • Keep a personal prompt file, with no client detail in it.

Using AI to check your work

~7 min

The second reader you do not have

Working alone is the hardest part of remote work when it comes to quality. You wrote it, so you read what you meant instead of what is there. A model is a genuinely useful second reader, with one condition: it is checking your writing and your logic, never your facts. Consistency, clarity, structure, tone, whether the conclusion follows from what came before, whether a paragraph contradicts one three pages earlier, whether everything the brief asked for is actually present. That is real value on a long document and it costs two minutes. It does not tell you whether the numbers are right. Nothing except the source tells you that.

Ask for problems, not praise

If you ask whether your report is good, you will be told it is good. That is the shape of answer the question invites, and it is worth nothing. Ask adversarially instead. List every internal inconsistency. Find every sentence that could be read two ways. Name every claim here with no supporting evidence. Where would a sceptical client push back. What did the brief ask for that this document does not deliver. Rewrite the three weakest sentences and say why they were weak. Then judge the answers yourself, because a good share of them will be wrong or irrelevant. You are generating candidates for review, not verdicts.

Checking without exposing the work

The confidentiality line applies here too, including to your own drafts, because a draft written for a client contains the client's information. On a AfterDesk task, do not paste the deliverable into a tool to have it checked. Run your own checklist, read it aloud, and use the tool only on things containing nothing of the client's: a formula pattern with fake data, a general question about phrasing a comparison, a grammar question about a sentence you rewrote with invented names. When a client outside the platform has approved AI use, the same document can go in with fewer contortions, but the approval has to exist first.

What a machine check cannot do

It cannot verify facts. It cannot know that this client dislikes the word solutions, that last month they asked for dates written a particular way, or that the file has to be named a certain thing. It cannot open your spreadsheet and see that a filter is still applied. It will sometimes flag correct sentences as errors and confidently rewrite something accurate into something wrong. So the sequence matters. Finish the work, run your own checklist, use the tool as a last sweep, then read its suggestions with your judgement switched on. Never accept a change you cannot explain the reason for. A silent improvement you did not understand is how a correct file becomes an incorrect one.

Remember

  • Ask it to find faults. Asking whether it is good yields flattery.
  • It checks writing and logic. It cannot check facts.
  • On AfterDesk, never paste the deliverable in for a check.
  • Accept no change you cannot explain in your own words.

Formulas, code and spreadsheets

~8 min

The best everyday use

Describing a spreadsheet problem in words and getting a working formula back is one of the strongest uses there is. It is fast, the ground is well documented, and you can test the answer completely. Say what you have and what you want. Column A holds full names, column B holds email addresses that may contain the surname, I want a column that flags rows where no email contains the surname, in Google Sheets. Name the tool you are using, because syntax differs between Google Sheets and Excel and the model will otherwise pick one for you. Say what should happen to blanks and errors. Ask for a short explanation of each part alongside the formula.

Test before you trust

A formula that returns numbers is not a formula that returns correct numbers. Build a small test where you already know the answer: twenty rows you have checked by hand, including the awkward ones. A blank. A duplicate. A name with an apostrophe or a suffix. A date in the wrong format. Text where a number should be. Run the formula and compare against your hand-checked result, row by row. If it matches on all twenty, extend it. If it fails on one, you have learned exactly where the logic is weak, which you would never have learned from the nine hundred rows where it happened to look fine.

Explain it or do not ship it

You will sometimes get a formula three layers deep that works and that you do not understand. Do not deliver it. The reason is practical rather than moral. It will be run again next month on slightly different data, and when it breaks, it is your file and you cannot fix it. Ask for a plain-language walkthrough of each function. Ask for a simpler version even if it is longer. Break it into helper columns you can read, then collapse it once you understand it. The standard for delivery is that you could explain what the formula does to someone who asked, without opening the tool again.

Structure without data

You can describe a spreadsheet completely without exposing a single real value. Give the headers, renamed if the headers themselves are sensitive, the data type of each column, the row count, and three or four invented rows carrying the same quirks as the real ones. That fake sample is worth building carefully, because the quirks are the whole problem: the trailing spaces, the mixed date formats, the two rows where the region is spelled differently. Solve it on the fake sample, then run the solution on the real file yourself. Nothing left the task, and you still got the help you needed.

Scripts and the same discipline

The same applies to a script, a formula inside a form, or a small automation. Read it before you run it. Understand what it touches. Never run something that deletes, overwrites or sends on a client file until you have run it on a copy of a fake file first and watched what it does. Keep an untouched original of any client file you received, work on a duplicate, and check the result before you deliver. Most disasters in this line of work are not dramatic. They are a well-meaning script quietly overwriting a column that took someone a week to compile.

Remember

  • Name your tool and your column layout. Spreadsheet syntax differs.
  • Test on twenty rows you checked by hand, including the awkward ones.
  • If you cannot explain the formula, you cannot deliver it.
  • Describe structure with invented rows that carry the real quirks.
  • Run anything destructive on a copy of a fake file first.

Disclosing AI use honestly

~6 min

The brief decides

Disclosure has two separate questions inside it, and people muddle them. The first is whether you were allowed to use the tool at all. That is answered by the brief and by nothing else. The second is whether, having been allowed, you should say where you used it. Handle the first question first. If the brief is silent about AI on a AfterDesk task, you did not have permission to put client data into one, and the disclosure question never arises. Where a brief permits it, disclosure becomes a matter of being useful and being honest, which here are almost always the same thing.

When disclosure matters

It matters most when the client's decision would change if they knew. A client paying for original research, an opinion piece published under their own name, anything they will sign, anything touching law, money or health. Those people are buying your judgement, and they are entitled to know what produced the words. It matters less for the mechanics. Nobody needs to be told you used a tool to convert a date format, to draft a formula you then tested, or to check your own grammar. The test is not how much the tool did. The test is whether a reasonable client would want to know.

How to say it

Say it plainly and specifically, in a sentence or two, without apology and without over-explaining. The structure that works is what you used it for, what you did yourself, and how you checked. For example: the first draft of sections two and three was AI-assisted, I rewrote and restructured both, and every figure and citation was verified against the original reports. On AfterDesk, where the brief allowed AI use, that goes in the note to the operator with your delivery. The person reviewing your work benefits from knowing which parts to look at hardest, and telling them costs you nothing but a sentence.

Never claim it did not happen

There is one absolute here. If someone asks you directly whether AI was involved, you tell the truth. Not a technically-true answer, not that it is your own process, not that you only used it for research. A direct question gets a direct answer, even when the honest answer is awkward, even when you suspect it will cost you the work. Getting caught in that particular lie ends a working relationship permanently and deservedly. The reason is not that detection tools exist, because they are unreliable and they are not the point. The reason is that a client who cannot trust your answer about this cannot trust your answer about anything.

Remember

  • Permission comes from the brief. Disclosure is a separate, later question.
  • Disclose where a reasonable client's decision would change if they knew.
  • Say what you used it for, what you did, how you checked.
  • A direct question about AI gets a direct answer. Always.

The output is yours

~6 min

The tool is not an excuse

When a wrong figure reaches a client, there is no version of that conversation where the AI wrote it improves your position. You chose the tool. You wrote the prompt. You read the output and decided it was good enough. You put it in the file and you delivered it. The responsibility is complete and it does not divide. This is not harshness. It is the same standard that applies to a calculator, a template, or a colleague you asked for help. Everything that leaves under your name is yours. Once you accept that fully, everything else in this course stops being a rule imposed on you and becomes obviously self-interested.

The signature test

Before any delivery that involved a tool, ask one question. Would I defend every sentence of this if the client called me about it right now. Not is it probably fine. Would you stand behind this specific number, this specific claim, this specific sentence, and explain where it came from. Anything that fails the test gets verified, rewritten, or removed. In practice this catches the two failure modes that actually matter: the fact you did not check, and the paragraph you did not really read. It takes a minute on most deliveries and it is the single most useful habit in this course.

Keep your own skill

There is a slower risk worth naming. If the tool writes every email, you stop getting better at writing emails. If it builds every formula, your spreadsheet skill stops growing, and the ceiling of what you can check becomes the ceiling of what you can safely deliver. That matters because your judgement is the thing clients are paying for, and judgement is what tells you an output is subtly wrong. Protect it deliberately. Write the first draft yourself sometimes. Work out the formula yourself when there is no deadline pressure. Read the explanation of anything you accepted without understanding. Use the tool to go faster at what you can already do, and to learn what you cannot.

A policy in one paragraph

Here is the whole course in something you could say out loud. I use AI tools for drafting, reformatting and first-pass ideas. No client data goes into any third-party tool without written permission, and on AfterDesk that permission is not given, so it does not happen. Every fact, figure, name and citation is verified against a real source before delivery. I disclose AI use where it would matter to you, and I answer directly if you ask. Whatever I deliver, I can explain, and I stand behind it. Say that, mean it, and you are ahead of most people working today, with tools or without them.

Remember

  • The tool is never an excuse. Everything under your name is yours.
  • Signature test: would you defend every sentence if called right now.
  • Use AI to go faster at what you can already do.
  • Your judgement is the product. Do not let it atrophy.

Sit the exam

The courses are free. The work is real.

Twelve scenario questions. Pass at ten. Three attempts a day. The bar is the point.

No paid tier. No certificate fee. No upsell. Not now, not later.

Create a free account

You can start the first course tonight.

The curriculum

Everything, and what is in it.