Working
with AI.

Pick one real task from this week that you know well enough to check. A follow-up after one of your own calls is a good first one. A task is one job you give the AI, from your request to the finished result. Set aside about ten minutes.

Below is how this job looks at Sidekick, where we run it after every client call: the call notes become a recap, a follow-up email and a list of open questions. The details in our example are changed. After it, you do the same with your own call notes.

Our request

This is the request we give the AI. The call notes go underneath it:

Turn these notes from a client check-in call into:
1. A short recap of what we agreed.
2. A follow-up email for me to check.
3. A list of open questions.

Use only these notes. Don't turn anything that was only discussed
into an agreement. Mark anything missing as unknown.
Don't send anything or change files.

The notes underneath (our example, details changed):

Notes:
- We agreed to move the client's monthly report from the 1st to the 15th.
- K. (client side) will send the updated contact list by Friday.
- The client asked about adding a second team. We said we'd send options.
- A start date of March 3 for the second team came up. Nobody agreed to it.
- The client mentioned a discount for a longer contract. We didn't agree to one.

What came back

A good answer looks like this. The wording can differ. The facts should match.

Recap: The monthly report moves to the 15th. K. will send the updated contact list by Friday. We will send options for adding a second team.

Follow-up email: Thanks for the call today. As agreed, we'll move your monthly report to the 15th, and we'll look out for the updated contact list from K. by Friday. We'll send you options for adding a second team, including possible start dates.

Open questions: Is March 3 the start date for the second team? Do we offer a discount for a longer contract?

If the email had agreed to the March 3 start date or the discount, the answer would be wrong even if the layout was lovely. A tidy answer can still have a wrong fact, so check the facts first.

Now do it with your own call

Open the AI app you already use (or the one your company has chosen): ChatGPT, Claude or Copilot. Start a new chat, paste the request above, and replace our notes with the notes from one of your own calls this week. Then read the answer against your notes. Anything that was only discussed should still show as open.

Make one correction

If the facts are right but the open questions are at the bottom, try:

Put the open questions first so I can see what needs an answer.
Keep anything we only discussed marked as open.
Don't change the facts or send anything.

Check that the new answer changed the order and kept the same facts. You have now used three habits on a real job: giving clear instructions, checking the answer, and correcting it. The rest of this guide builds on those three.

Save a note for next time

Give me a five-line progress note I can save for next time:
the task, what we finished, open questions, any writing preference,
and the next step.

Save that note somewhere you can find it. Later, open a new chat, paste the note and ask, “What is the next step?” A good answer brings back the open questions. It should not claim the email was sent.

Think of a chat as a conversation at a colleague's desk. You ask a question, talk through options, or fix a paragraph together. A work task is more like handing that colleague a job: you say what you want back and give them what they need to do it.

In a chat, the AI answers one message at a time, and you take the next step. In a work task, the AI works through several steps on its own. It opens your files, plans the job, does each step, and hands back a finished file for you to check.

“What goes in a client proposal?” starts a chat. “Use these call notes, our approved past proposals and our template sections to draft a proposal, mark anything not agreed, and give me the draft” hands over a job. OpenAI's Chat and Work comparison explains the difference for ChatGPT.

What happens during a work task

The model is the AI that reads and writes. The app is the screen around it and may offer several models. During a work task, the model reads your request, plans the next step, uses a tool (one action, such as opening a file, searching, or creating a document), looks at the result, and then carries on or asks you a question.

The AI can be wrong and still sound sure, so you check. It can misread a file, miss a condition, or present a guess as a fact.

Chapters 1 to 6 work in ordinary ChatGPT, Claude or Copilot chat. Later chapters use the “Work” mode in paid plans; Chapter 13 shows where to find it in each app.

Choose the right mode for the job

Start with the simplest mode that can finish the job. A two-line reply only needs a chat.

Use it on a question you asked AI this week: rewrite it to name the document or result you want back. Replace “Help with our report” with “Write the report from these files, and check these points.”

See what you could make

Each example below is based on a job Sidekick or one of its clients actually did. Names and details are changed.

A comparison that helps you choose. A firm that runs on Microsoft 365 had to pick one AI tool to roll out: Microsoft Copilot, or Claude or ChatGPT. Its IT provider would approve a tool only after four questions: who owns the account, is every permission needed, what business process it supports, and is it for one user or many. A good result puts the deciding fact up front:

For your own research, give the AI current sources or ask it to find them, then check the facts that decide the choice. Chapter 3 shows how to write the instructions.

A pile of notes, turned into a list you can use. The proposals lead at a small HR consulting firm came back from a conference with a stack of handwritten contact cards. She had the AI turn them into an organized follow-up list: name, organization, what they talked about, and the next step, ready for a follow-up email. A few entries needed correcting by hand. A row looked like this (details made up):

Follow-up list: R. M., operations lead at a regional clinic. Talked about onboarding checklists for new staff. Next step: send the checklist sample.

The AI does the sorting and typing. You check the names and details, because handwriting is easy to misread. Chapter 9 covers checking the work.

A weekly list that puts decisions first. Every Sunday at Sidekick, the AI checks the open items and turns them into a short list of yes-or-no questions for the owner. Reviewing a draft moved the decisions to the top:

You can see at once what needs your attention. Chapter 12 has both weeks' notes, the prompts and the checks, so you can try it yourself.

Pick a job you know well enough to judge. You will spot a mistake in familiar work much faster than in work that is new to you.

Good instructions give the AI everything it needs to do the job without guessing.

Say what you want back, who will read it, which files to use, what to avoid, and how you will know it is done. Give it the files that matter for this job: the current template and the right notes. Your whole company drive is usually too much.

“Write a proposal” leaves the AI guessing what you are offering, to whom, and on what terms. The example here is based on a real client job at a small HR consulting firm; details changed. There, the proposals lead drafts each proposal from the discovery-call notes, the firm's approved past proposals and its template sections. The AI builds most of the draft, the team reviews every one, and fonts and spacing are fixed by hand afterwards. The short version of the instructions:

Draft a proposal from these call notes, our past proposals and template sections.
Mark anything not agreed. Don't send it.

And the full version:

Draft a client proposal for me to review.

Use the attached discovery-call notes, our template sections, and two
approved past proposals for layout and wording.
The reader is the client's owner, who needs to decide whether to go ahead.
Use plain language and make the scope and next step easy to find.

The scope in the notes is agreed. A start date and a phased payment option
came up on the call, but neither was agreed. Mark them clearly.
Don't make up terms. Use the past proposals for layout and wording only:
don't copy their client names, scope or fees.

It's done when it follows the template, every fact matches the call notes,
open items are easy to see, and the next decision is clear.
Give me the draft. Don't send it.

Ask me only for missing information that would change the result.

If you can't describe the result yet, ask the AI to help:

Help me work out this job before you start. Ask me one question
at a time about what I want back, who it's for, what to use,
and what a good result must include. Then write the instructions out for me.

Tell it what matters most

Sometimes your needs pull in different directions. A proposal can't explain every detail and still be one page. Say which matters more: “Make the scope and next step easy to find. Put the detailed method in an appendix.”

If you have a good example, share it and say what to copy from it. “Use the layout of this proposal” should never turn into “reuse its old client's name, scope and fees.” An old proposal used for layout can still carry another client's details, so check for them.

You can also talk instead of typing. Describe the job out loud, then ask the AI to write the instructions back to you. Check names, numbers and dates afterwards, because dictation gets them wrong.

You're ready for the next chapter when the AI can repeat back, in a few sentences, what you want and what it must not do. If it can't, fix your instructions before asking for more.

Most AI apps use a few words for the parts of your setup. A project is a space in the app for one area of work, such as one client. A folder is where your files are stored. A task is one job inside that project. One project can hold many tasks.

What the AI is holding for this task is its context: your request, your saved instructions, the files it has opened, and anything it looked up. The AI can only keep so much in view at once. In many apps, when a long task fills that space, older parts get summarized or dropped. If it starts forgetting, start a new task and paste your progress note (Chapter 5).

Keep clients separate

Telling the AI “only use this client's files” doesn't stop it seeing other files, if those files are within the app's reach. A cloud project only sees what you uploaded. Keep clients in separate projects or accounts. An assistant who does admin work for two or three clients at once, for example, keeps each client in its own project. Use the permission settings in your accounts, which control who can open which files. When you link an account such as email or Google Drive, it has its own settings for which files the AI can open. The ChatGPT Work overview explains this for ChatGPT.

Start with a small set of files

For regular proposal work, this would do:

Client proposals/
  Progress note.docx
  Files to use/
    Template sections.docx
    Approved past proposals/
    Discovery call notes.docx
  Drafts/
    Proposal draft 1.docx

This is only an example. If your company already has a sensible place for these files, use it.

Keep one agreed current version and rename old drafts clearly, for example with a date. When three files are all called “Final”, neither you nor the AI can tell which one is current.

Start a new task for a new result, such as writing the proposal and, separately, checking the presentation. Stay in the same task for changes to the same result. Keep work in the same project when it uses the same files and the same people can see it.

How each app handles projects

In ChatGPT, a cloud project keeps project files and background across related chats. A local project in the desktop app gives chats access to folders on your computer. A cloud project can't reach the files on your computer. See OpenAI's projects guide.

Claude Cowork projects can hold folders, instructions, links and things the project remembers. Claude's chat projects and Cowork projects work differently, so check the Cowork projects guide before moving work between them.

In Microsoft Copilot Cowork, use the file and task views in your account. Its setup differs from Claude's folders. Microsoft's getting-started guide shows how files go in and come out.

What the AI can see has limits

Your folders may hold years of material, but the AI only works with what it has opened for this task. A search may bring back only part of a file. A new task may need to read your saved notes before it can carry on correctly.

For important work, ask: “Which files did you use, and what couldn't you open?” If two files disagree, ask the AI to show you where. Then tell it which one to trust, so it doesn't blend two versions together.

Use it on your next document job: give the AI your current template and one old example. Ask which one it will follow for layout and which for facts. Sort out any confusion before it writes anything.

The AI doesn't reliably remember yesterday's conversation. A short progress note fixes that. It says where the work stands: what is done, what is still open, which file is the latest, and what comes next. You save it at the end and give it to the AI at the start, so you don't spend ten minutes explaining everything again.

Start from your progress note

Read the attached progress note and the files it names.
Tell me the last finished step, what is still open, and the next step.
Tell me which points come from the note and which are your guesses.
Tell me if a file is missing. Don't change files or send anything yet.

Once you agree on where things stand, give it the next job. You don't need a new plan after every small step. If it starts to drift, steer it while it works: “Keep the agreed recap. Only change the email's last paragraph.”

Finish with a progress note

Wrap up this task. Write a short progress note: what we finished,
what we decided, what is still open, where the latest file is saved,
and the next step. Include only corrections that matter next time.

If you are allowed to update our progress note, update it and show me
what changed. Otherwise give me the note to save myself. Don't say it
is saved unless you can see the saved file. Don't send anything.

The note can be short:

Client check-in follow-up. Recap and email drafted. The monthly report moves to the 15th (agreed). The second team's start date and the longer-contract discount are still open. Put open questions first in follow-up emails. Next: confirm the open items, then send the email. Nothing has been sent.

Save a short note on where things stand. You don't need to save the whole back-and-forth. The note only helps if the AI reads it next time, so give it to the AI or tell it where to find it.

Test it once

Open a new task in the same project and ask the AI to tell you the next step from your saved note. If it can't, check where the note is saved and how you asked it to read the note. You don't need to rebuild the whole setup.

When you move work from one AI app to another, bring the same short note and the files that matter. The new app can't see your old conversation. Let only one task edit a given file at a time, so two AI apps don't overwrite each other's changes.

This chapter has worked when you can stop, come back another day, and carry on from your note without retelling the whole story.

“Make it better” gives the AI very little to work with. Helpful feedback says what missed, why it matters, and what you want instead.

An executive coach we work with has the AI draft her session recaps and follow-up emails in her own voice, and she reviews each one before it goes out. Compare “This sounds wrong” with the kind of feedback she gives: “This recap sounds stiff for a client I coach every week. Keep the action points, drop the formal phrases, and match the tone of the recap I wrote last month.” (Based on a real client job; details changed.)

You don't need to work out why the AI went wrong. Point to the part that missed. The AI can help turn that into a clear correction.

Does the AI remember what I tell it?

The AI doesn't learn from your chats. A memory feature saves some of your text and shows it to the AI again next time. Some plans can use your chats to train their AI unless you turn that off in the app's settings.

If a correction matters next time, save it somewhere the AI will read, and check that later work follows it.

OpenAI explains the difference between saved memories and project instructions in its personalization guide. Claude explains project memory in its projects guide. Either way, when the AI “remembers” something, check it against your saved note or file before you rely on it.

Turn one correction into a rule you can reuse

You treated the March 3 start date as agreed. Fix this draft.
Then suggest one short rule for our follow-up instructions that would
prevent this mistake. Apply it only to client follow-ups. Tell me where the rule
should be saved and how we can test it on a new example.
Don't change any saved instructions yet.

A good rule would be:

Keep agreed terms separate from suggestions. Mark a price, date or responsibility as agreed only when the files clearly say it was agreed. Otherwise, mark it “to confirm”.

Once the right person approves the rule, save it with your other instructions. Next time, give the AI notes that include one agreed item and one suggestion, and check that it tells them apart. That is better proof than the AI saying, “I've learned my lesson.”

Keep corrections small and specific

A one-off preference may only belong in this task. A layout you use for a monthly report belongs with that report. A rule for the whole company belongs with whoever owns that rule. Don't pile every correction into one big set of instructions; rules can clash or make simple jobs awkward.

Keep a change only if the next result is better.

Use it on something the AI got wrong this week: pick one thing it keeps getting wrong. Write the correction, decide where to save it, and pick the next ordinary task that will show whether it worked.

In Chapters 1 to 6 you paste text in or attach files. This chapter is for when you want the AI to open your email, calendar or shared drive itself. Use it when a task needs that.

A connector lets the AI open your files. A skill tells it how to do a job. A plugin can bundle both. You still check that it opened the right file. OpenAI's Skills and Plugins page and Microsoft's Cowork customization page explain each app's version.

Add a connector only when a task needs it

If you only need the AI to summarize one document, attach it. If you keep downloading the latest file from a shared drive and uploading it again, a connector saves you that step.

Setting one up can take several steps: add it in the app, sign in to the right account, choose what the AI may do, and check that it can reach the file you need.

Then give it a small test: “Find this document by name and tell me its title and the date it was last changed. Don't change it.” Open the file it found. A confident summary of the wrong file is still the wrong file.

Decide what the AI may change

Reading your calendar is one thing. Inviting ten people to a meeting is another. Finding a proposal is one thing. Replacing it is another. The AI can open whatever your connected account can open, so limit what the account can reach as well as what you ask for. OpenAI's Work access overview explains these settings for ChatGPT.

For your first connector, pick one narrow job with a clear result. Ask for a draft and send it yourself.

Three safety habits

1. Use a setting for anything that matters. The AI can ignore a rule you type. It can't get past a setting in the app, so use a setting for sending, deleting and paying. Example: make the email connector read-only.
2. Watch for hidden instructions. An email or web page can contain hidden text meant as orders for the AI. Point it at material you trust, and if it does something you didn't ask for, stop it and ask what it was doing.
3. Let schedules make drafts only. A scheduled task (Chapter 11) runs without you watching, so let it prepare drafts and leave sending, deleting, paying and changing records to a person.

Use it on a chore you repeated this week: name one step in it that a connector would remove. If you can't name one, leave connectors for later.

When you keep giving the AI the same instructions or the same corrections, save them once and reuse them (the app calls this a skill). Reusable instructions can hold steps, a template, examples and checks. Project instructions apply to everything in a project. A skill is used only when the job matches it.

For a one-off task, a prompt is enough. Save instructions for the jobs you do every week or month.

What to include

For client proposals, reusable instructions would cover:

• When to use them: turning discovery-call notes into a client proposal draft.
• What they need: the call notes, the firm's approved past proposals, its template sections, and the terms you have confirmed.
• Steps: pull out the facts, list what is missing, write the draft, then check it against the files.
• What they produce: a proposal with anything unconfirmed clearly marked.
• Limits: never mark something as agreed without proof, and never send the result.

Keep each client's details in that client's files. Keep the method in the reusable instructions. Otherwise one client's details can slip into the next client's proposal.

Write them out before you save them in the app

Turn the way we just wrote this proposal into a draft skill for future
client proposals. Include when to use it, what it needs, the writing steps,
the checks, and what needs my approval. Use our corrected draft as an
example of the layout, but remove this client's details.

Keep it short. Reuse instructions we already have instead of repeating them.
Give me the instructions to review. Don't install or share them yet.

Once you approve them, save them using your app's skill setup. Test them in a new task with a different client of the same firm, using that client's own notes: the result should follow the method without carrying over the old client's details.

In ChatGPT you can pick saved instructions yourself or let the app choose; OpenAI's skill guide shows how. Microsoft's customization guide explains how to create, import and test them. For Claude, use the Claude skills help page.

Moving them to another app

The written method usually carries over well. The settings and file names may need changes. Ask the new app to check that the instructions work before you rely on them.

Your saved instructions work when a new task uses them on a different customer and passes the same checks, with none of the old client's details.

Asking the AI to re-check its own work, even as a “customer” or “finance manager”, helps with clarity but often misses the same wrong facts. Check facts against your files yourself. Ask a qualified person when the decision needs one.

For bigger tasks, ask the AI to work in stages, each doing one thing: write the draft, check the facts, check it reads clearly, ask what could go wrong, then fix it. You can ask for these stages in plain words, with nothing installed.

Follow one proposal through the stages

In the HR firm's proposal from Chapter 3, the fact check catches the start date that came up on the call but was never agreed. The reader check notices that the next step is buried at the end. Asking what could go wrong raises whether the past proposal used for layout still carries another client's details. Each check finds a different problem.

Check the work against your instructions

Check this proposal against my instructions, the call notes and the template.
Check that every fact is supported, nothing required is missing, it is clear
for the customer, and the next decision is obvious.

For each serious problem, show where it is, why it matters, and the smallest fix.
Fix the draft, then check the parts you changed again.
Don't change agreed terms or add claims the files don't support.

Ask what could go wrong

What is the most likely reason this customer would say no, or the job
would go wrong? Use the scope we agreed and the files.

Separate problems you can see in the files from risks that might happen.
Tell me what information would change your view. Suggest only changes
that would make a difference. Don't invent objections to fill a list.

Ask for every stage at once

You can ask for every stage in one message. Paste this prompt and change the task:

Take this task from my instructions to a checked draft. Use only the steps it needs:
ask about important gaps, plan, write, check the facts and requirements,
fix serious problems, and give me the result with any open decisions.
Stop before sending, posting or changing anything I haven't approved.

Stop checking when the checks pass and no serious problem remains. Asking for small improvements again and again rarely changes anything that matters.

Use it on a draft from this week: run one fact check and one “what could go wrong” check on it. Notice whether they find different problems. Keep the checks that change the result.

Say what the finished job looks like. “Work on my proposal” is too loose. “Write the proposal from this template and mark anything unconfirmed” is clear.

Write a finish line the AI can check

Write a draft of the client proposal from the attached files.
Include the outcome, the scope, who does what, anything not yet
agreed, and the client's next decision.

Use our template. Don't make up prices, dates or promises.
You're done when every section is there, every fact is backed by the files,
the document opens properly, and the open decisions are listed.
Give me the draft and stop before sending it.

If you aren't sure what finished looks like, work it out with the AI first. “Automate my business” is too vague. “Write a weekly update from these two files, and don't post it” is a good start.

Test it on three different inputs

Run your checks on three different inputs, including one messy one, such as call notes with no start date and a payment option nobody agreed. For the HR firm's proposal, the checks look like this:

These are simple checks for everyday work. They make “good” less of a matter of opinion and catch mistakes that smooth writing can hide.

Match the check to the job: add up a spreadsheet total again, trace a summary back to its notes, look at a design at the size people will see it, or question the evidence behind a recommendation.

A score from the AI proves little

If the AI scores its own work, that tells you little. Ask what it checked, then check it yourself. Perfect spelling can't make up for a missing legal term or the wrong customer's name.

One good result shows the task worked. Several good runs show the method is worth keeping.

You're ready to move on when you can name a few ways the work could go wrong, run those checks, and tell a finished result from one that still has open items.

Once the AI gives you a good result again and again, you can hand over the routine parts around it: starting the task, collecting the files, and saving the draft. Automate one dependable part at a time.

A schedule runs the same job again. If the job has a mistake, you get the mistake every time, so fix the method first. Let a schedule make drafts only (see the safety habits in Chapter 7).

Move up one step at a time

Here is how this went for our own after-call follow-up at Sidekick. The AI gathers the notes, earlier work and open promises, then drafts the recap, the follow-up email and the CRM note. Safe record updates happen on their own. Only the judgment calls go to the owner, and nothing is sent without him.

Sending emails, changing records or buying things is harder to undo than writing a draft. Keep a person in charge where the work needs one. Let the AI act alone only when it saves time and its mistakes are small and easy to catch.

Plan a small routine

Pick one regular piece of work you already know how to check. Write down when it starts, which files it uses, where the result goes, what to check, and what should happen when a file is missing or a run fails. Name the person who handles problems and can pause it.

Test it by hand three ways: a normal run, a run with no new notes, and a run where a file is missing. Then use your app's own scheduling settings to set it up, and check the time zone, files, destination and permissions before you turn it on.

Chapter 12 shows the whole routine for our Sunday decision list as a worked example, with our method, checks and schedule plan. You can run it on your own items before connecting anything.

Where scheduled tasks run

A task runs now. A schedule starts it again later.

Where the task runs matters too. Cloud: runs even if your computer is off. Local: your computer must stay on and the app open. Chapter 13 has the details for each app. See also OpenAI's scheduled tasks guide. For Claude, see the Cowork getting-started guide. For Microsoft Copilot Cowork, check the Cowork overview or ask your admin.

Add up the total time

Compare the total time spent gathering, checking, fixing and finishing with how long the job took before. Check your plan's usage limits, and time the job by hand once so you can tell whether it saves time. A fast first draft followed by an hour of fixes may not save anything.

If a scheduled task gives you too much noise, give it fewer files or tell it to report only when something changes. If it keeps missing the same thing, fix the method before running it more often. If you aren't sure whether a step finished, look at the result before retrying, so you don't create duplicates. Pause the routine when the files or the people involved change enough that the old rules no longer fit.

A sensible next step: pick one regular draft you already know how to check. Improve it until each run needs less explaining. Then decide whether a schedule would help.

Every Sunday at Sidekick, the AI goes through the open items and checks each one against email, the calendar, the live website and our records. It hands back a short list of yes-or-no questions for the owner. It takes about 10 minutes. In one week's run, it started from 51 open items: 9 were closed and 21 were parked with return dates. It also caught a draft email that said “attached” with no file attached. Nothing is sent without the owner.

This chapter shows that routine as a worked example, using the earlier chapters. The items below follow the shape of our real list; the details are changed. Then run it on your own open items from last week, in a chat first, with nothing connected or installed.

Our request: the open items and a finish line

This is the request we use, with the items changed. To run it yourself, swap in your own open items from last week:

Turn these open items into my Sunday decision list.
Use only the notes below. Keep it short, with these headings:
questions for me (yes or no), closed, parked. After each fact,
add the label of the note it came from, such as [A1].
Don't send anything or change files.

It's done when I can answer every question with yes or no.
Mark an item closed only if a note shows it is done.
Don't make up names or dates.

The notes underneath (our example, details changed):

Week 1 notes:
[A1] Proposal for a new client: draft ready, waiting for my yes to send.
[A2] Invoice for last month's work: the Sent folder shows it went out Tuesday.
[A3] Website update promised to a partner: no record that it's done.
[A4] Follow-up email to a prospect: the draft says “attached”, but no
file is attached.

The labels in brackets show which note each fact came from: [A1] means note A1. A good list could say:

Sunday list, week 1

Questions for you: Send the new client proposal? [A1] The partner's website update has no record of being done. Do it this week? [A3] The prospect follow-up says “attached” with no file. Attach the file before sending? [A4]

Closed: Last month's invoice went out Tuesday. [A2]

Parked: Nothing yet.

The labels make checking easy: you can trace each line back to a note. Open the note each label points to and check it says that.

Correct the method once

Suppose the AI parked the website update without a date. Tell it:

Every parked item needs a return date. Keep the questions first,
then closed, then parked. Keep the note labels next to the facts.
Rewrite the list in that order.
Then give me a short method I can reuse every Sunday.

The facts should stay the same. A short method is enough:

Sunday list method

Read this week's notes and the last progress note. Check each open item against the records it names. Write questions for me first, each answerable with yes or no, then closed items, then parked items with a return date. Label each fact with the note it came from. Mark an item closed only if a note shows it is done. Flag any draft that mentions an attachment it doesn't have. Say if any notes are missing. Give me the list to review.

For now, save or copy that method. Later it can become a skill (Chapter 8). You can test whether it helps before saving it in the app.

Save where things stand, separately from the method

Ask for a progress note, as in Chapter 5. Compare it with this example and save a correct version for week 2:

Sunday list, end of week 1. Invoice closed. Waiting on my answers: send the new client proposal, do the partner's website update, attach the file to the prospect follow-up. Use the Sunday list method, questions first. Next: read week 2 notes and update these items. Nothing sent.

The method says how to do the job. The progress note says where this week's items stand. Keeping them apart lets you reuse the method every Sunday without carrying last week's details into the new list.

The next week: start fresh and test the method

The next Sunday, we open a new chat and paste the method and the saved progress note, then this request. Do the same with your own second week:

Continue using the Sunday list method and the week 1
progress note above. Read the week 2 notes below. Write the new list and
show what changed. Then list any check that failed or any information
still missing. Where a new note settles an item, update it.
Don't carry a closed item forward. Don't send anything or change files.

The notes underneath (our example, details changed):

Week 2 notes:
[B1] I said yes to the new client proposal. The Sent folder shows it
went out Monday.
[B2] The partner's website update is done. The live page shows the change.
[B3] The prospect follow-up still says “attached” with no file.
[B4] A partner asked for a call. The notes don't say when.

A good result could look like this:

Sunday list, week 2

Questions for you: The prospect follow-up still says “attached” with no file. Attach it before sending? [B3] A partner asked for a call. Offer times this week? [B4]

Closed: The new client proposal went out Monday. The partner's website update is live. [B1, B2]

Parked: Nothing.

Check two things. Did the facts update correctly? Did the new chat put questions first, as the method says? If both are right, this test passed. Future runs may still drift, so keep checking.

Before you finish, replace the week 1 progress note with the week 2 position: proposal sent, website update live, prospect follow-up still missing its file, partner call not yet offered. Keep the old note as history if you like, but mark the week 2 note as current.

Four checks for the weekly list

You now have a routine in steps: read → check against records → draft → check → fix → save a progress note. You can run all these steps in one task. Save a step as its own skill only if you want to reuse it elsewhere.

Asking what could go wrong may raise “What if the prospect follow-up goes out without the file?” It should not invent evidence that it already has. You still decide; the AI makes the open item easier to see.

Before you trust the routine, run two checks on your own list. Each one tests whether the AI invents facts when a record is missing:

1. Run your method once with one item's record left out, such as the email that shows something was sent. That item should come back as unconfirmed, because nothing shows it was done.
2. Run it once with no new notes at all. The result should say the notes are missing. It should not report that nothing changed.

These two checks test the mistakes most likely to mislead you.

Connect the records when copying them in becomes the chore

Once the routine works by hand, you may want the AI to check email, the calendar and your records itself. Keep the same method and checks. Only the way the notes arrive changes.

See Chapter 7 to test access to one named file first. With your own records, ask the AI to tell you the file name and date it found, and open that file once to confirm it is the right one. In your own version, replace the example labels like [B1] with links to the real records.

Let a schedule start the tested routine

When the connected version works well week after week, set up a schedule with your app's own settings (Chapter 11). Set up the schedule in the app's settings; a chat message can't do it. A plan for this example:

In the app, check the saved settings and the first run, including what happens when something fails. If your app offers one, set a limit on runs or spending.

You have gone from writing one list to setting up a weekly routine. The possible gain is less time chasing loose ends. Weigh that against the time spent checking and fixing the result over several weeks. If the routine adds more work than it saves, simplify it before you schedule it.

Use the same pattern for other jobs

For marketing, the small HR consulting firm from Chapter 3 builds its marketing pieces from its own templates and writing style, and a person approves each one before anything is posted. Swap the list checks for checks that the facts match the source, the tone fits the firm, and the format is right. Save a correction such as “use the figures from the source; don't make the results sound bigger than they are.” (Based on a real client job; details changed.)

For a spreadsheet, an executive coach we work with has her 56-question team questionnaire scored automatically in a linked sheet. Before trusting it, test the scoring against one questionnaire scored by hand, and check what happens with a blank answer. A blank should not count as a zero. (Based on a real client job; details changed.)

The steps carry over to other jobs. What counts as a good result depends on the job.

Each app puts these controls in a different place. Follow the steps for the app and account you have. While you set things up, you can still use instructions and checks on real tasks in an ordinary chat.

Which one do you have?

Not sure which to use? Start with the app you already pay for. These details were checked on October 20, 2026. Details change often. If something here doesn't match your app, its help page has the current answer.

Claude Cowork

Open Cowork and choose where the task should work: start with a small approved folder or a project. Anthropic's short setup tutorial walks through the first steps. You can leave connectors until a task needs them.

Your files and your project are separate things. A project helps the AI remember background between tasks; check that it is using the latest file. For devices, plans and settings, see Claude's Cowork help.

Microsoft Copilot Cowork

Attach the files you need or pick them from your cloud storage. Start with Microsoft's setup instructions.

For reusable instructions, the Customize area has Plugins and Skills. Use the customization guide, and test a changed skill in a new conversation. What you can use depends on your account. A folder setup that works in Claude tells you nothing about Microsoft's permissions.

ChatGPT Work

In ChatGPT, switch from Chat to Work when you want a bigger piece of work done. Give it the files and say what result you need. Local work uses files on your computer (see Chapter 11 for what that needs). Cloud work uses files stored online. The Work introduction, with video, shows the first steps.

If you use the ChatGPT desktop app, a /goal command can hold your finish line on long jobs. Use projects for related work, and use the Plugins area to add extra features. On the web, a project uses the files you upload to it and can't read your computer. OpenAI's Chat and Work comparison explains the differences.

If the app doesn't match this guide

Don't keep clicking around to match an old screenshot. Tell the AI which app, which device, and the labels you can see, and ask it to use the app's current help pages. Ask it whether the feature is missing or you lack permission. Whoever owns the account decides which tools and information you can use: you, or your IT admin in a larger company. A tutorial can't give you access.

Find what happened in the left column, then try the fix on the right.

When you can't tell what went wrong, use:

Stop and help me work out what went wrong. What did you try, what happened,
and how do you know? Tell me whether the problem is missing information,
missing permission, a tool that failed, or an unclear instruction.
Suggest the smallest safe next step. Don't retry anything that sends,
posts or changes something, and don't change any access settings.

You don't have to become the IT department. When you ask someone for help, include what you were trying to do, the step that failed, the exact error message, and what is still unfinished. Leave passwords and private information out.

Find your problem below. Use the short prompt now, or follow the chapter link for the full explanation and example. For an error or an unexpected change, see Chapter 14.

You don't have to read in order. Move on when your work needs it. You are making progress when good work takes less effort and fewer corrections.

Pick the resource for the question you have now. You don't need to finish a course before trying the first task. Some official learning sites ask you to sign in; the examples in this guide work without them.

Start with your app

Learn one app at a time. A few links per app is plenty to begin with.

• ChatGPT: the Work introduction, with video, to see Work mode in action; ChatGPT Projects, to keep one client's files and instructions together; the Meet ChatGPT Work video (1 min 40 s), a short official demo.
• Claude: the short Cowork setup tutorial, to set it up step by step; Claude skills explained, to save instructions for a job you repeat; the Claude Cowork, coming to mobile and web video (1 min 21 s), a short official demo.
• Microsoft Copilot: Microsoft's setup instructions, to open Cowork for the first time; the customization guide, to add skills and plugins.

Watch an introductory video with one question in mind: “What do I do to give it my next job?” Pause and try that step. The video pages linked here were checked; we have not reviewed every video from start to finish.

What to read next, by stage

Words used in this guide

More from Sidekick

• AI Opportunity Map: a free assessment that shows where AI could help your business, how to handle your data, and what to do first.
• The hours test: pick one office task and see how many hours AI could free up.
• AI policy for Canadian businesses: what staff can and cannot put into AI tools.
• All Sidekick guides: more free guides and articles for business owners.

Next time you sit down with the AI, bring one task you know well enough to check. Say what you want back, give it the files, and use your first correction to improve the next attempt.