Skip to the guide
Guide — Getting Started with ChatGPT

Getting Started
with ChatGPT

For project managers whose company just turned ChatGPT on. Take one weekly job, like the owner update, as far as it needs to go, with a person reviewing it the whole way.
BY Naram Alhasani
October 10, 2026
October 13, 2026
Construction and operations teams
Before you start

Start with Thursday night

Reviewed
October 10, 2026
Reading time
About 30 minutes
Written for
Someone new to ChatGPT at work
Assumes
No technical background

If you're a project manager and your company just bought ChatGPT, this guide is for you. It starts with a job you already do every week.

Picture a project manager, Dana. It's Thursday night, the owner update is due in the morning, and she has a week of daily logs from her superintendent, Luis, open next to last week's update. Most of the hour goes into re-reading Luis's notes to find what changed since Monday, and that kind of careful reading is something ChatGPT does well.

One sentence in the request makes it safe to hand over. It tells ChatGPT that if a date, figure, name or owner isn't in the notes, it should mark that line “needs confirmation” and leave it blank. Without it, a missing date tends to come back as a plausible guess, and plausible guesses are exactly what end up in front of an owner.

Can I paste this?
On a Business or Enterprise workspace, OpenAI doesn't train its models on your content by default, so a week of daily logs is usually fine. Your chats are still visible to the company, because workspace admins can view, export and delete them. For pricing, contract terms and anyone's personal information, follow your company's rule, and if nobody has written one yet, ask whoever owns the rollout before you paste the first bid tab.

How this guide is laid out

The guide takes one job up a ladder with six rungs: Ask, Deliver, Remember, Reach, Repeat and Run on its own. Each rung takes a little more setup and pays you back by making you explain the job less often.

Most people should stop at the third rung. Ask, Deliver and Remember cover the owner update, the OAC meeting summary and the bid comparison, and stopping there is the right call for most jobs. The higher rungs earn their setup once a second person runs the job, or once it genuinely needs to start without you.

Three things to do first
  1. Pick one job you already reviewSomething weekly that you already know a good version of, like the owner update. Last week's approved copy lets you judge a draft in a couple of minutes.
  2. Find the switchOn desktop it's a toggle at the top of the window between Chat and Work mode, and on your phone it's the dropdown at the top. If you only see Chat, ask whoever runs the rollout to turn Work on.
  3. Name the reviewerThe one person who reads the output before it goes anywhere, often you. A confident draft and a correct one look identical on screen, and the reviewer is how you tell them apart.
Contents

Chapter 01

The ladder

Everything ChatGPT does fits on one ladder. Each rung takes a bit more setup, and in return you explain the job a bit less often.

Figure 01 — The ladder A staircase of six columns rising from left to right, each one a rung. Rung one, Chat: ask, check, follow up. Rung two, Work: get the file, not a reply. Rung three, Projects: remember, set it up once. Rung four, apps and web: reach the web and your systems. Rung five, plugins: repeat, package the method. Rung six, schedules: run on its own, it starts without you. A thin arrow climbs above the stair tops. Underneath every column runs one unbroken blue band labeled the review: a named person checks the output at every rung. The footer says to climb only as far as the job needs, and that the review never comes off. FIG 01 · THE LADDER — SIX RUNGS, ONE REVIEW CLIMB 01 · CHAT Ask Ask, check, follow up. 02 · WORK Deliver Get the file, not a reply. 03 · PROJECTS Remember Set it up once. 04 · APPS + WEB Reach The web and your systems. 05 · PLUGINS Repeat Package the method. 06 · SCHEDULES Run on its own It starts without you. THE REVIEW — A NAMED PERSON CHECKS THE OUTPUT AT EVERY RUNG Climb only as far as the job needs. The review never comes off.
Figure 01 — The ladder A staircase of six columns rising from left to right, each one a rung. Rung one, Chat: ask, check, follow up. Rung two, Work: get the file, not a reply. Rung three, Projects: remember, set it up once. Rung four, apps and web: reach the web and your systems. Rung five, plugins: repeat, package the method. Rung six, schedules: run on its own, it starts without you. A thin arrow climbs above the stair tops. Underneath every column runs one unbroken blue band labeled the review: a named person checks the output at every rung. The footer says to climb only as far as the job needs, and that the review never comes off. FIG 01 · THE LADDER 01 · CHAT Ask Ask, check, follow up. 02 · WORK Deliver Get the file, not a reply. 03 · PROJECTS Remember Set it up once. 04 · APPS + WEB Reach The web and your systems. 05 · PLUGINS Repeat Package the method. 06 · SCHEDULES Run on its own It starts without you. THE REVIEW A named person checks the output at every rung. Climb only as far as the job needs. The review never comes off.
Figure 01 — The ladder. Each rung up asks for more set-up in exchange for less typing per run, and the person named as reviewer stays the same at every height.

Here's what each rung adds to Dana's owner update.

  • Ask. She pastes Luis's notes into Chat, asks what changed and checks the answer.
  • Deliver. Work mode hands her a finished Word document, so she skips the copy-and-reformat step.
  • Remember. Her template and rules live in a project, so she stops pasting them every Thursday.
  • Reach. ChatGPT reads the web or the company's own tools as well as what she pasted.
  • Repeat. Her method gets packaged, so the next PM runs the update her way.
  • Run on its own. The draft starts at a set time, or when something happens in a connected app.

The band along the bottom of the ladder diagram is the part that stays put. On every rung a named person reads the output before anyone relies on it, and that check gets easier to skip the higher you climb, because the drafts start to look finished well before they're right.

What it's good at

Think of ChatGPT as a sharp assistant PM who has read every spec book ever printed and has never once set foot on your job. Hand it Luis's notes and it'll turn them into a clean draft. Hand it a dense spec section and it'll explain it in plain English or tell you where two revisions of a submittal disagree.

Where it struggles is any answer that depends on a fact nobody gave it. A wrong date arrives in the same confident sentence as a right one, which is why the review sits under every rung.

Check your account
What you see depends on your plan, your company's workspace settings and how far OpenAI's rollout has reached. When a colleague's screen shows something yours doesn't, ask whoever runs the rollout before anyone files a ticket.

Try thisWrite your job at the top of a page and the reviewer's name underneath it. Every chapter comes back to both.

Chapter 02

Ask

This is the rung everyone already uses. It gets far more dependable once you know what an answer was built from and you set how hard it thinks.

What an answer is built from

Say Dana pastes Monday's and Tuesday's logs and asks what's slipped. The answer is built from whatever ChatGPT has in front of it, which comes from five places.

  1. Your prompt. What you typed this turn.
  2. The conversation so far. Earlier turns in the same chat.
  3. Instructions. Your own custom instructions, or the ones saved in a project.
  4. Files you gave it. Uploads in the chat, or reference files in a project.
  5. What it looked up. A search result, or something from a connected app.

It also draws on general knowledge from its training, which is how it can explain a curing requirement you never mentioned, and with memory on it can use things you've told it before. Both help with tone and explanation. Neither is a source for a figure, a date or an owner in your business.

Figure 02 — What supplies context in an answer A four-line answer with each line traced back to where its information came from. Line one, the period the update covers, came from your prompt, typed this turn. Line two, the definition of practical completion, came from a file stored in the project. Line three, the milestone date in the shared plan, was retrieved from a connected app, with permission, and only because your account may already see it. Line four is drawn with a dashed border: the commissioning owner was absent from every source, so it is marked needs confirmation rather than filled in. A black band states that asking for more confidence does not supply the fourth fact, but asking a person does. FIG 02 · WHAT SUPPLIES CONTEXT IN AN ANSWER WHERE IT CAME FROM THE SENTENCE IN THE ANSWER Your prompt Typed this turn. “This update covers September 29 to October 5.” A project file Stored in the project. “Practical completion means the signed handover note.” A connected app Retrieved, with permission. “The milestone date in the shared plan is October 14.” Nobody supplied it Absent from every source. “Commissioning owner — needs confirmation.” Asking for more confidence does not supply the fourth fact. Asking a person does. An answer is only as grounded as the material behind it. Fluent wording is not evidence that the material was there.
Figure 02 — What supplies context in an answer A four-line answer with each line traced back to where its information came from. Line one, the period the update covers, came from your prompt, typed this turn. Line two, the definition of practical completion, came from a file stored in the project. Line three, the milestone date in the shared plan, was retrieved from a connected app, with permission, and only because your account may already see it. Line four is drawn with a dashed border: the commissioning owner was absent from every source, so it is marked needs confirmation rather than filled in. A black band states that asking for more confidence does not supply the fourth fact, but asking a person does. FIG 02 · WHAT SUPPLIES CONTEXT FROM YOUR PROMPT “This update covers September 29 to October 5.” Typed this turn. FROM A PROJECT FILE “Practical completion means the signed handover note.” Stored in the project. RETRIEVED FROM AN APP “The milestone date in the shared plan is October 14.” Only what you may already see. NOBODY SUPPLIED IT “Commissioning owner — needs confirmation.” Absent from every source. Asking for more confidence does not supply fact four. Fluent wording is not evidence.
Figure 02 — What supplies context in an answer. Two of these four sources you control directly, one depends on access someone else granted, and the fourth is where an instruction to flag gaps earns its keep.

The answer breakdown takes a four-line reply apart. Three lines trace back to something real. The fourth traces back to nothing, and that's the one to plan for, because when a fact is missing a fluent model will often fill the gap with something that sounds right. So add this sentence to any request where accuracy matters.

If a figure, date, name or owner is not in the material I gave you, mark it “needs confirmation” and leave it blank.

Most people skip this because the drafts look fine without it. Don't. It turns every gap into a line your reviewer finds in seconds, before the date in question has passed.

When an answer is wrong
Before you rewrite the prompt, ask “What did you use to answer that?” The reply usually points to the problem (the file was never read, the wrong revision was, or the fact simply wasn't there), and each has a different fix. Treat the reply as a lead rather than proof, because naming a source and having read it are two different things. Open the passage it cites and check it yourself.

How hard it thinks

In Chat, a slider called Thinking sets how long ChatGPT reasons before it answers. It runs from Instant, for quick everyday answers, through Medium, High and Extra High, up to the Pro setting for the hardest questions.

Figure 03 — The Thinking slider A horizontal slider track with five stops, from quicker on the left to deeper on the right: Instant, fast and everyday; Medium, standard; High, extended; Extra High, the most thinking; and Pro, for the hardest and longest questions. A filled blue thumb sits on Instant, marked you are here. Below, two boxes compare ways to get more reasoning. On the left, a dashed box labeled asking: typing think harder about this. On Plus and Pro, Instant stays Instant, and a small copy of the slider shows the thumb blocked by a stop bar. On the right, a blue box labeled moving the slider: Instant to High, more reasoning and a longer wait, with the small slider's thumb moved to the third stop. The footer says effort is a control, not a request, to be matched to the stakes of the question. FIG 03 · THE THINKING SLIDER YOU ARE HERE QUICKER DEEPER Instant Fast, everyday Medium Standard High Extended Extra High Most thinking Pro Hardest, longest ASKING “Think harder about this.” On Plus and Pro, Instant stays Instant. MOVING THE SLIDER Instant → High More reasoning, a longer wait. Effort is a control, not a request. Match it to the stakes of the question.
Figure 03 — The Thinking slider A horizontal slider track with five stops, from quicker on the left to deeper on the right: Instant, fast and everyday; Medium, standard; High, extended; Extra High, the most thinking; and Pro, for the hardest and longest questions. A filled blue thumb sits on Instant, marked you are here. Below, two boxes compare ways to get more reasoning. On the left, a dashed box labeled asking: typing think harder about this. On Plus and Pro, Instant stays Instant, and a small copy of the slider shows the thumb blocked by a stop bar. On the right, a blue box labeled moving the slider: Instant to High, more reasoning and a longer wait, with the small slider's thumb moved to the third stop. The footer says effort is a control, not a request, to be matched to the stakes of the question. FIG 03 · THE THINKING SLIDER QUICKER Instant YOU ARE HERE Fast, everyday Medium Standard High Extended Extra High Most thinking Pro Hardest, longest DEEPER ASKING “Think harder about this.” On Plus and Pro, Instant stays Instant. MOVING THE SLIDER Instant → High More reasoning, a longer wait. Effort is a control, not a request. Match it to the stakes of the question.
Figure 03 — The Thinking slider. Choose the setting before you send; typed requests to think harder change only the wording, and the higher stops earn their wait when a wrong answer would be costly.

This is where people get caught. Say the slider is on Instant and you paste in two spec sections that contradict each other. You'd expect ChatGPT to notice the question is hard and slow down. Don't count on it. On Plus and Pro it stays at Instant, and typing “think harder” won't move it, so whatever plan you're on, the reliable way to get more thinking is to move the slider yourself, as the slider diagram shows. Leave it on Instant for a quick rewrite, and slide it up when a wrong answer would cost you something, like a spec conflict or a contract clause.

Model names change several times a year and differ by plan, so this guide leaves them out and your own picker stays the accurate list. In a team procedure, write “use the highest effort your account offers for this step,” and the procedure survives the next rename.

Try thisAsk a question you already know the answer to, once on Instant and once on High. If the extra thinking only made it longer, that tells you something about your kind of work.

Chapter 03

Deliver

The second rung changes what you get back. Chat hands you a reply to read. Work mode hands you a finished document, spreadsheet, presentation or report, along with the steps it took to make it.

Three modes behind one switch

The switch moves you between three modes, and you'll spend almost all your time in the first two.

  • Chat is the back-and-forth you already know.
  • Work is for longer jobs that end in a file, like the owner update as a formatted Word document.
  • Codex, in the top-left menu on desktop, is for writing and changing software. Most PMs can leave it alone.
Figure 04 — One request, three experiences The same request, prepare this week's client update from the approved site notes, held against three experiences. In Chat, what comes back is a draft inside the conversation, and you review every fact against the notes you supplied; choose it when you want to think out loud. In Work, what comes back is a built document or spreadsheet plus the steps it took, and you review those steps, the sources it used and anything it asked permission to do; choose it when you want the deliverable itself. The third column is deliberately a negative answer: Codex returns nothing for this request, because there is no software in it. Codex would only apply to a separate, bounded job, such as changing the script that produces the milestone export, where the output is a code difference and test evidence. The footer notes that the request did not change, only what came back, and that the honest third column is the one that says this is not a software task. FIG 04 · ONE REQUEST, THREE EXPERIENCES THE REQUEST — HELD AGAINST ALL THREE COLUMNS “Prepare this week's client update from the approved site notes.” Chat CONVERSATION COMES BACK A draft inside the chat. Text you read and paste. YOU REVIEW Every fact, against the notes you supplied. CHOOSE IT WHEN You want to think out loud. Work DELIVERABLE COMES BACK A built document or sheet, plus the steps it took. YOU REVIEW The steps, the sources, and anything it asked to do. CHOOSE IT WHEN You want the deliverable itself. Codex NOT THIS JOB COMES BACK Nothing for this request. There is no software in it. YOU REVIEW A code difference and tests — but on its own job. CHOOSE IT WHEN The export script changes. The request did not change. What came back did — and so did the thing you are now accountable for checking. The third column is the honest one: Codex is the wrong tool until a real software change is the work.
Figure 04 — One request, three experiences The same request, prepare this week's client update from the approved site notes, held against three experiences. In Chat, what comes back is a draft inside the conversation, and you review every fact against the notes you supplied; choose it when you want to think out loud. In Work, what comes back is a built document or spreadsheet plus the steps it took, and you review those steps, the sources it used and anything it asked permission to do; choose it when you want the deliverable itself. The third column is deliberately a negative answer: Codex returns nothing for this request, because there is no software in it. Codex would only apply to a separate, bounded job, such as changing the script that produces the milestone export, where the output is a code difference and test evidence. The footer notes that the request did not change, only what came back, and that the honest third column is the one that says this is not a software task. FIG 04 · ONE REQUEST, THREE WAYS THE REQUEST “Prepare this week's client update from the site notes.” Chat CONVERSATION COMES BACK A draft inside the chat. YOU REVIEW Every fact, against the notes. Work DELIVERABLE COMES BACK A built document, and steps. YOU REVIEW Steps, sources, permissions. Codex NOT THIS JOB COMES BACK Nothing — no software here. ONLY IF The export script changes. Same sentence. Different output — and a different thing to check. Codex only once there is real software to change.
Figure 04 — One request, three experiences. The typed sentence stays the same across all three columns; what changes is the thing you hold at the end, and so the thing you must check.

The three-mode comparison holds one request up against all three, and what differs is the review each one leaves you. A Chat draft asks you to check the facts. A Work deliverable also shows the steps it took, the sources it used and anything it asked permission to do.

Work is something you supervise. While it runs you can watch, answer its questions, change direction and approve the important steps. It replaced what older articles call “ChatGPT agent.”

Where it runs

On the web and on your phone, Work runs in the cloud. Dana can start the draft at her desk and open the finished document from the truck an hour later. The desktop app can also open folders on your own computer, one at a time.

Running locally changes where the work happens, though not where everything is stored. Messages and task context can still be stored in the cloud, so if you're choosing the desktop app for a data-handling reason, confirm it with whoever owns that rule.

Check your account
Opening local folders from the desktop app depends on your plan and your company's settings. A missing option means it hasn't been turned on for you.

Check the file before you summarize it

Give ChatGPT a spreadsheet or an export and it can run code to clean, total and chart the data. Ask what's in the file before you ask what it means, because a summary built on a broken export reads just as confidently as one built on a clean one.

This is our RFI log exported from Procore. Before you summarize anything, check it for RFIs with no ball-in-court, duplicate RFI numbers and anything open past its due date, and list each one by RFI number. Then summarize what's still open. Use only what's in the file.

Sample — made-up data
An eight-row log, exported on October 10:
rfi_no,subject,ball_in_court,due,status
014,Curtain wall anchor spacing,Architect,10/02,Open
015,Roof drain overflow size,MEP engineer,10/15,Open
016,Elevator pit depth,,10/16,Open
017,Stair 2 guardrail height,Architect,10/20,Open
018,Level 3 slab penetration,,10/17,Open
019,Door hardware set 4,Owner,10/24,Open
020,Lobby ceiling height,Architect,10/06,Closed
021,Fire damper access panels,MEP engineer,10/28,Open

A correct answer names RFIs 016 and 018 as having no ball-in-court, flags 014 as overdue since October 2, and reports no duplicate numbers, before it says a word about what's open. Your check takes a minute. Scan the ball-in-court column for blanks, read down the due dates and count the rows, and if ChatGPT's list disagrees with yours, find out why before you trust the summary.

Charts deserve the same suspicion, because a chart inherits every problem in the data underneath and hides most of them. Name the period and recheck one number by hand.

Change one paragraph at a time

Canvas opens a document beside the conversation. You point at the paragraph you want changed and everything around it stays exactly where it was, which matters most on a document someone has already approved.

Keep the approved headings, prices and dates exactly as they are. Rewrite only the scope section in plain English, at a similar length. List every change you made.

Sample — made-up document
Before: “Scope: Contractor shall furnish all labor, materials and equipment necessary for the execution and completion of the works described in Schedule A, inclusive of ancillary items reasonably inferable therefrom.”

After: “Scope: We supply the labor, materials and equipment needed to complete the work in Schedule A, including the smaller items that work obviously requires.”

The change list: “Changed: the scope paragraph only. Unchanged: all headings, the Schedule A reference, prices and dates.”

Check that last line yourself against the approved version.

Photos and voice

ChatGPT can read a photo of a whiteboard, a scanned table or a screenshot, and it can make images from a description. Treat its reading of a drawing as an interpretation to check, since a misread dimension looks just as tidy as a correct one, and where a dimension or a safety note matters the drawing is the authority.

Voice lets you talk instead of type, in Chat and in Work, which suits a superintendent walking the site. On the web and on your phone, approvals still happen with the on-screen button, and saying yes out loud doesn't count.

Try thisRun something you asked in Chat this week in Work mode instead, and judge what you had to do afterward. Did you copy and reformat, or open and review?

Chapter 04

Remember

By the third Thursday, Dana has pasted the same template and explained the same definitions three times. This rung gives them a home. A project keeps the chats, files and instructions for one piece of ongoing work together, so every new chat inside it already knows the job.

Keep your owner-update project private. That one decision lets Work mode draft from the project and keeps its file list from turning into a distribution list. Make one project per recurring deliverable, such as the owner update for one job or the monthly WIP report, because a project for a whole department turns into a junk drawer.

Figure 05 — What goes where The material behind a weekly client update, sorted into the four places it can live, drawn as four columns. Project instructions, outlined in blue, hold what complete means, the sections the output must have, the rule for missing information and who reviews it; they last for every chat in the project. Project files, drawn as documents, hold the update template, the last approved update and the milestone list; they also last for every project chat. Memory, drawn with a dashed outline, holds that you like short bullets and your role and team, and lasts across your chats if memory is on; a grayed dashed card inside it says company definitions do not belong there. This chat, in gray, holds this week's site notes, today's question and a one-off attachment, and lasts for this conversation only. FIG 05 · WHAT GOES WHERE — MATERIAL FOR ONE WEEKLY UPDATE PROJECT INSTRUCTIONS What “complete” means The output's sections The missing-info rule Who reviews it LASTS Every project chat PROJECT FILES The update template Last approved update The milestone list LASTS Every project chat MEMORY You like short bullets Your role and team Not company definitions LASTS Your chats, if on THIS CHAT This week's site notes Today's question A one-off attachment LASTS This conversation Put each fact where it needs to last. Definitions go in instructions, never in memory.
Figure 05 — What goes where The material behind a weekly client update, sorted into the four places it can live, drawn as four columns. Project instructions, outlined in blue, hold what complete means, the sections the output must have, the rule for missing information and who reviews it; they last for every chat in the project. Project files, drawn as documents, hold the update template, the last approved update and the milestone list; they also last for every project chat. Memory, drawn with a dashed outline, holds that you like short bullets and your role and team, and lasts across your chats if memory is on; a grayed dashed card inside it says company definitions do not belong there. This chat, in gray, holds this week's site notes, today's question and a one-off attachment, and lasts for this conversation only. FIG 05 · WHAT GOES WHERE PROJECT INSTRUCTIONS What “complete” means The output's sections The missing-info rule Who reviews it LASTS Every project chat PROJECT FILES The update template Last approved update The milestone list LASTS Every project chat MEMORY You like short bullets Your role and team Not company definitions LASTS Your chats, if on THIS CHAT This week's site notes Today's question A one-off attachment LASTS This conversation Put each fact where it needs to last. Definitions go in instructions, never in memory.
Figure 05 — What goes where. Sorting the material once, before the first run, is what lets next week's update start from the same rules without anyone pasting them in again.

The four-places diagram sorts the material for a weekly update by how long each piece needs to last. Your definition of “complete” outlasts every chat, so it goes in the instructions. Luis's notes only need to last until the update is approved, so they go into this week's chat.

What goes in the instructions

Project instructions apply to every chat in the project and take priority over your personal custom instructions, so the rules of the job live here.

  • Definitions. What “complete” means on this job, written once so two updates can't drift into two meanings.
  • The shape of the output. Sections, order, length, and anything that must stay out, such as pricing.
  • The missing-information rule. The “needs confirmation” sentence, stated once.
  • The reviewer. Named, so everyone who opens the project sees who signs off.

Use the report structure in this project. Keep our definitions of percent complete and open items exactly as written. Summarize this week's changes, flag every missing input, and list the decisions the project manager needs to make. Do not infer a figure that is not in the supplied material.

Memory is about you

With memory on, ChatGPT keeps useful things you've told it, like your role and how you like answers laid out, and you can view, edit or delete them whenever you like. Because it's personal, a definition your team relies on belongs in the project instructions, where everyone in the project can read and fix it.

Why private matters

Each project runs on either default memory or project-only memory, and sharing a project switches it to project-only memory for good. Project-only memory turns Work off. So the day Dana shares her owner-update project with Luis is the day Work mode stops drafting from it, which is the first reason to keep it private.

The second reason is the file list. Everyone in a shared project can download every file in it, so a shared project behaves like a distribution list and needs reviewing whenever someone joins or leaves the job. If Luis needs the template, send him the template. And if you ever delete a project, its chats, instructions and any files stored only there go with it, while files saved separately to your library stay put.

A project doesn't refresh itself
Someone has to add this week's notes and notice when they're missing. Decide who before a Monday meeting depends on it.

Try thisCreate the project for your job and leave it private. Add your definitions, the missing-information rule and last week's approved version.

Chapter 05

Reach

So far ChatGPT has only seen what you handed it. This rung lets it go and get information, from the web or from the systems your company already runs on.

Search or deep research

Search looks something up and comes back with a short answer and its link. Use it for a quick fact, like a manufacturer's listed lead time on a rooftop unit.

Deep research is the slow, thorough version. It plans an investigation, reads across many sources and returns a report with citations, which is worth the wait when the answer drives a decision, like choosing between four equipment rental vendors.

Figure 06 — Search versus deep research Two paired examples. On the left, search: you ask whether a standard has been revised this year, you get a short answer with the link it came from, and you open the link and carry on with the conversation. One source, one answer. On the right, deep research, drawn as a four-step chain: step one is a proposed plan you read and edit before anything runs; step two is the sources, which can be the open web, sites you name, your uploads and apps you allow; step three is the report, structured, cited and downloadable as Markdown, Word or PDF; step four, drawn with a dashed border, is what it could not settle, left visible rather than filled in. The footer warns that a citation existing is not the same as a citation supporting the sentence attached to it. FIG 06 · SEARCH VERSUS DEEP RESEARCH Search A FACT YOU ASK “Has this standard been revised this year?” YOU GET A short answer, with the link it came from. YOU DO Open the link. Carry on with the conversation. One source. One answer. Deep research A DECISION STEP 01 · THE PLAN A proposed plan you read and edit before anything runs. STEP 02 · THE SOURCES The open web, sites you name, your uploads, apps you allow. STEP 03 · THE REPORT Structured, cited, downloadable as Markdown, Word or PDF. STEP 04 · What it could not settle, left visible. A citation existing is not the same as a citation supporting the sentence attached to it. Open the two that would change your mind.
Figure 06 — Search versus deep research Two paired examples. On the left, search: you ask whether a standard has been revised this year, you get a short answer with the link it came from, and you open the link and carry on with the conversation. One source, one answer. On the right, deep research, drawn as a four-step chain: step one is a proposed plan you read and edit before anything runs; step two is the sources, which can be the open web, sites you name, your uploads and apps you allow; step three is the report, structured, cited and downloadable as Markdown, Word or PDF; step four, drawn with a dashed border, is what it could not settle, left visible rather than filled in. The footer warns that a citation existing is not the same as a citation supporting the sentence attached to it. FIG 06 · SEARCH VS DEEP RESEARCH Search A FACT YOU ASK “Has this standard been revised?” YOU GET A short answer and its link. One source. One answer. Deep research A DECISION STEP 01 · THE PLAN You read and edit it before anything runs. STEP 02 · THE SOURCES Open web, named sites, uploads, apps you allow. STEP 03 · THE REPORT Structured and cited. Markdown, Word or PDF. STEP 04 · STILL OPEN Left visible, not filled in. A citation existing is not a citation supporting the sentence attached to it. Open the two that matter.
Figure 06 — Search versus deep research. Deep research pauses for you to edit its plan, so the cheapest correction you will ever make to the report is the one made before it starts.

The most useful control in deep research is the one most people skip. Before anything runs, it shows you a proposed plan you can edit, and as the research diagram shows, that's the cheapest moment to catch a misread question. You can also point it at specific websites and interrupt to change direction. It reads from connected apps without writing to them, and reports download as Markdown, Word or PDF.

Compare these four equipment rental vendors against our requirements. Use each vendor's own pricing, service and safety pages first. Give a link and the date checked for every factual claim. Keep published facts separate from marketing claims. List the questions you couldn't answer instead of filling them in.

The two-citation check
Open the sources for the two claims that would most change your decision, because a citation can exist without supporting the sentence beside it. If both hold up, read the rest with more confidence. If either fails, check every claim the decision rests on.

Connected apps

Apps connect ChatGPT to the tools your company already uses, so it can find a file in SharePoint, summarize a thread from Outlook or Teams, or look something up in Google Drive, if your workspace has those apps turned on. Some apps can also take actions. Type @ and the app's name, or pick it from the + menu, and it may check with you before reading or acting.

On Business, Enterprise and Edu plans there's also Company Knowledge, which answers company-specific questions from the sources you're already allowed to see. Type @Company Knowledge and ask where the latest site logistics plan is. Answers sometimes arrive without citations, so ask for them and check any answer that lacks one.

Figure 07 — From connection to action A five-stage permission flow that forks at stage three. Stage one, available: the app is enabled in the workspace, which an administrator decides. Stage two, connected: your own account is signed in and authorized with the provider. Stage three, the source: ChatGPT returns only what you may already see, which is set in that service rather than in ChatGPT. From stage three the path forks. The not-permitted branch terminates: a document your account cannot open stays closed, and the branch ends there, with no onward connection to any later stage. The permitted branch continues independently to stage four, the action, where reading is one thing and sending another, and then to stage five, approval, where a sensitive action can pause and wait for your decision. The footer records that an app permission setting changes when ChatGPT asks you before acting, not what the app is allowed to reach, and that installing a plugin clears the first box only. FIG 07 · FROM CONNECTION TO ACTION PERMITTED 01 · AVAILABLE In the workspace An administrator decides this. 02 · CONNECTED Your account Signed in and authorized. 03 · THE SOURCE Only what you may already see Set in that service. 04 · THE ACTION Read, or write Finding is not the same as doing. 05 · APPROVAL It can pause and wait for your decision. NOT PERMITTED THIS BRANCH ENDS HERE A document your account cannot open does not become readable here. The permission setting on an app changes when ChatGPT asks you before acting. It does not change what the app may reach. Installing a plugin clears box 01 only. It asks nothing of boxes 02 to 05.
Figure 07 — From connection to action A five-stage permission flow that forks at stage three. Stage one, available: the app is enabled in the workspace, which an administrator decides. Stage two, connected: your own account is signed in and authorized with the provider. Stage three, the source: ChatGPT returns only what you may already see, which is set in that service rather than in ChatGPT. From stage three the path forks. The not-permitted branch terminates: a document your account cannot open stays closed, and the branch ends there, with no onward connection to any later stage. The permitted branch continues independently to stage four, the action, where reading is one thing and sending another, and then to stage five, approval, where a sensitive action can pause and wait for your decision. The footer records that an app permission setting changes when ChatGPT asks you before acting, not what the app is allowed to reach, and that installing a plugin clears the first box only. FIG 07 · CONNECTION TO ACTION 01 · AVAILABLE In the workspace An admin decides this. 02 · CONNECTED Your account Signed in and authorized. 03 · THE SOURCE Only what you may already see in that service. Set there, not in ChatGPT. NOT PERMITTED A document you cannot open stays closed. The branch ends. 04 · ACTION Reading is one thing. Sending is another. 05 · APPROVAL It can pause and wait for you. An app permission changes when ChatGPT asks before acting — not what it may reach. Installing clears box 01 only.
Figure 07 — From connection to action. Each gate has a different owner, so when a retrieval fails, the first question is which of the five stopped it and who controls that one.

A connected app returns what your account can already see in that system, and nothing more. Its permission settings decide when ChatGPT asks before acting, as the permissions diagram traces, so a PM who can't open the bid leveling folder in SharePoint won't reach it through ChatGPT either.

Your construction systems

Most of a contractor's real data lives in Procore, Autodesk Construction Cloud, CMiC, Viewpoint or Sage. Whether ChatGPT can reach any of them depends on whether that platform offers an app for your plan and whether your workspace has it turned on.

In the meantime, an export works well. A Procore RFI log or a CMiC cost report, uploaded on purpose, gives ChatGPT exactly the material you meant it to have, and that's a habit worth keeping even after an app arrives, because you always know what it read.

Before anyone relies on it

  • Start read-only. Let the first connected workflow find and summarize, and add actions once someone owns the approvals, since an action in the wrong job's folder is harder to undo than a bad summary.
  • Run the negative test. Have someone ask for a document they shouldn't be able to see, like another job's bid leveling. When the lookup comes back empty, the permissions are doing their job.
  • Read the app's terms. Data you share with an app is handled under that app's own terms and privacy policy.
Check your account
What an app can reach depends on your plan, your workspace, your role and the app itself. Ask whoever runs the rollout what's turned on before you design a process around it.

Try thisConnect one app, read-only. Run the negative test on one document before you connect a second.

Chapter 06

Repeat

This rung is for the day someone else runs your job. Maybe Dana's on vacation, or a new PM inherits the owner update on a second job. The question becomes where the method lives. Plenty of people never need this rung, so come back to it when a second person actually has to run the job.

Skills and plugins

Think of the page you'd leave for whoever covers your job while you're out: what to ask for, what to read, what to send and who signs off. Written so ChatGPT can follow it too, that page is a skill.

The box that carries it to your colleague is a plugin. It holds one or more skills, along with the reference files the job needs and any connected apps, so the other PM installs one thing and gets the whole method.

To use one, type @ and its name, or open + and choose More. ChatGPT may pick a relevant skill by itself when its description matches your request, though it won't always, so name the plugin when the job matters. Installing a plugin grants no new access, and app sign-ins, permissions and admin approval stay where they were.

See a real skill library
Formwork's Contractor Toolkit is an open-source set of 38 skills for general contractors and trades, covering estimating, proposals, RFIs, submittals, daily logs and change orders. It's packaged for Claude, but each skill is plain written instructions, so a couple of them show what a good skill contains, from the questions it asks first to where a person signs off. It's built for Claude, so you can't install it in ChatGPT, but the written steps make a good model for your own.

If your team built custom GPTs

OpenAI plans to retire custom GPTs and help the people who built them move them into plugins. If you only use one that someone else made, there's nothing to do yet. If you built one, read the migration notes for whoever runs the rollout at the end of Chapter 9 before the notice arrives.

Try thisWrite your job's method as one skill: what it asks for, what it reads, what it returns and who signs off. Even unpackaged, the page earns its keep the first time someone covers for you.

Chapter 07

Run on its own

The top rung is work that starts without anyone typing. A task can begin three ways. You ask for it, a schedule fires, or something happens in a connected app.

Figure 08 — Three ways a task starts Three entry points converge on one task and one review. Start A is you asking, a request in the moment. Start B is a schedule, a time you set, once or again. Start C is an app event: a new mail, a channel message, or pull request activity. Each start connects to a shared vertical line clear of every box, and that single line leads into the task, which runs in the relevant experience with your permissions and not more, and then into the review, where a person reads the result before it is used or sent anywhere. Two separate notes sit below, connected to nothing. The first is scoped to a task you create inside a project: the Scheduled Tasks documentation says it cannot access that project's uploaded or stored files. The second records that the limit does not run the other way, because a Work chat opened from a private default-memory project can use its context. FIG 08 · THREE WAYS A TASK STARTS START A · YOU ASK A request, in the moment. START B · A SCHEDULE A time you set, once or again. START C · AN APP EVENT A new mail, a channel message, or pull request activity. THE TASK It runs in the relevant experience, with your permissions, not more. THE REVIEW A person reads it before it is used or sent. SCOPE · SCHEDULED TASKS IN A PROJECT It cannot access that project's uploaded or stored files. This is the Scheduled Tasks page's project-file restriction. NOT THE OTHER WAY ROUND A Work chat in a private project can use its default-memory context. A trigger starts work. It does not supply missing context, and it does not supply missing permission.
Figure 08 — Three ways a task starts Three entry points converge on one task and one review. Start A is you asking, a request in the moment. Start B is a schedule, a time you set, once or again. Start C is an app event: a new mail, a channel message, or pull request activity. Each start connects to a shared vertical line clear of every box, and that single line leads into the task, which runs in the relevant experience with your permissions and not more, and then into the review, where a person reads the result before it is used or sent anywhere. Two separate notes sit below, connected to nothing. The first is scoped to a task you create inside a project: the Scheduled Tasks documentation says it cannot access that project's uploaded or stored files. The second records that the limit does not run the other way, because a Work chat opened from a private default-memory project can use its context. FIG 08 · THREE WAYS A TASK STARTS START A · YOU ASK A request, in the moment. START B · A SCHEDULE A time you set. START C · AN APP EVENT New mail, a channel message, pull requests. THE TASK It runs with your permissions, not more. THE REVIEW A person reads it before it is used or sent. SCOPE · SCHEDULED TASKS A scheduled task inside a project cannot access that project's stored files. Not ordinary project chats. NOT THE OTHER WAY ROUND A Work chat in a private default-memory project uses context. A trigger starts work. It supplies no missing context and no permission.
Figure 08 — Three ways a task starts. Whichever start you choose, the person reading the result is the same, and the project-file note applies only to scheduled tasks created inside a project.

All three starts lead to the same task and the same review, as the three-starts diagram shows. That's worth remembering on the morning a scheduled draft lands looking finished.

Schedules

ChatGPT can run a task once or on a repeating schedule, and it can run monitoring tasks that watch for a change and tell you when something relevant happens. You create and manage all of them from Scheduled. Make the first one low-stakes on purpose, like a Friday task that drafts next week's look-ahead questions for your superintendent, so a misfire costs you nothing.

A scheduled task created inside a project can't see the project's files, whether you uploaded them or they're stored there. So write everything the task needs into the task itself. A Work chat you open yourself from a private project still uses the project's context. Only scheduled tasks miss it.

Event triggers

Work mode can also start when something happens. The supported events include a new Gmail message, a new message in a Slack channel and pull request activity in a GitHub repository. If your team runs on Outlook and Teams, check the current list in your account before you plan around a trigger.

To set one up, connect the account in Settings, open Work, describe the event and what should happen, and read the trigger, condition and prompt before you save. For Slack, add @ChatGPT to every channel the task watches. The task uses only the connected account's existing permissions and pauses for your approval when an action needs one.

Check your account
Event triggers aren't available on Free or Go plans, and on Enterprise and Edu an admin has to switch them on first. You create and edit triggers on the web or in the mobile app. The desktop app shows them but can't change them.

Sharing a task

You can share a scheduled task as a link. Whoever opens it gets a copy of the title, instructions and schedule to run under their own account, while your files, history and connections stay with you. Anyone with the link can read the instructions, so keep passwords, account numbers and health or financial details out of them.

Keep a person between the schedule and the client
A trigger starts work. It can't supply context the task lacks or permission the account lacks. When a message to an owner has to go out on a schedule, send it from the system you already trust and let the task nudge the person who approves it. Tasks can also pause on their own, after inactivity or when their chat is deleted, so glance at Scheduled before you assume something ran.

Try thisSet up one task, with everything it needs written into the task itself, and watch it for two weeks before you add a second. This rung usually fails as a pile of notifications nobody reads.

Chapter 08

One job, all the way up

Here's Dana's owner update, taken up every rung it needs. It's an example, not a measured result, so time your own.

Dana runs a four-story office building for a mid-sized general contractor. Every Friday the owner's rep gets an update built from Luis's daily logs and the milestone schedule, and every Thursday night Dana writes it. The update already has a reader, a reviewer (her project executive) and a run of approved versions to compare against, which makes it a good first job to hand over.

Figure 09 — One week of the weekly update A week drawn as a timeline from Monday to Friday. Monday to Wednesday, the site team writes and approves notes before ChatGPT sees them. Thursday, Work drafts the update from the project's template and rules, which flow in from the project instructions; then the owner, drawn as a black box, checks each fact and every needs-confirmation line. Friday, the approved update is filed in the system of record, not left in the chat, and one line per fix is logged, recording what the draft got wrong. A feedback arrow runs from Friday's log back under the week into the project instructions, labeled repeated corrections: when a correction repeats, the rule is fixed there rather than in the draft. The footer says the review happens every week, but the instructions improve only when someone writes the corrections down. FIG 09 · ONE WEEK OF THE WEEKLY UPDATE MON TUE WED THU FRI MON–WED · THE SITE TEAM Notes are written and approved before ChatGPT sees them. THE PROJECT INSTRUCTIONS When a correction repeats, fix the rule here, not the draft. THU · WORK Drafts the update from the project's template and rules. THU · THE OWNER Checks each fact and every “needs confirmation” line. FRI · FILED System of record Not the chat. FRI · LOG IT One line per fix What it got wrong. Repeated corrections The review happens every week. The instructions improve only when someone writes the corrections down.
Figure 09 — One week of the weekly update A week drawn as a timeline from Monday to Friday. Monday to Wednesday, the site team writes and approves notes before ChatGPT sees them. Thursday, Work drafts the update from the project's template and rules, which flow in from the project instructions; then the owner, drawn as a black box, checks each fact and every needs-confirmation line. Friday, the approved update is filed in the system of record, not left in the chat, and one line per fix is logged, recording what the draft got wrong. A feedback arrow runs from Friday's log back under the week into the project instructions, labeled repeated corrections: when a correction repeats, the rule is fixed there rather than in the draft. The footer says the review happens every week, but the instructions improve only when someone writes the corrections down. FIG 09 · ONE WEEK, ONE UPDATE MON WED THE SITE TEAM Notes are written and approved before ChatGPT sees them. THU WORK Drafts the update from the project's template and rules. THE OWNER Checks each fact and every “needs confirmation” line. FRI FILED System of record Not the chat. LOG IT One line per fix What it got wrong. Repeated corrections THE PROJECT INSTRUCTIONS When a correction repeats, fix the rule here, not the draft. The review happens every week. The instructions improve only when someone writes the corrections down.
Figure 09 — One week of the weekly update. The week's only new habit is the one-line log on Friday; without it, the same fix gets made by hand every Thursday.

The weekly loop lays out one of her weeks. The Thursday review stays fixed, and the loop back into the project instructions makes week six better than week one.

The first Thursday

She starts in Chat, pasting last week's approved update and asking what structure it follows and what an owner would want that's missing. The answer notes that her updates never say who's waiting on whom, which is fair, so she adds a line for it.

Then she builds the project with the template, her definitions, the “needs confirmation” rule and her reviewer's name, and leaves it private on default memory so Work mode can draft from it.

Every Thursday after that

Luis's notes for the week are approved by mid-afternoon. Dana opens the project, picks Work, attaches the notes and says up front what she'll check. A document comes back, with the steps it took.

Use only the approved notes below and the template in this project to prepare this week's update. Separate completed work, open issues and next steps. Include the note date for every factual statement. If a deadline, owner or status is not stated in the notes, mark it “needs confirmation” and do not infer it. Do not estimate percent complete. Before you finish, list anything you couldn't find in the material I gave you.

Sample — made-up notes
The approved notes going in:
“Mon 9/29 — slab pour completed to grid line 4. Wed 10/1 — commissioning walkthrough postponed, new date not set. Fri 10/3 — dumpster haul billed to cost code 03.”

The draft coming back:
Completed. Slab pour completed to grid line 4 (note, 9/29). Dumpster haul billed to cost code 03 (note, 10/3).
Open. Commissioning walkthrough postponed (note, 10/1).
Needs confirmation. New commissioning date — not stated in the notes. Commissioning lead — not stated in the notes.

Those last two lines are the workflow doing its job, and the reviewer reads them first. A draft that quietly supplied a plausible date would read better, and nobody would catch the error until the date passed.

Then she checks every fact against its source note, “needs confirmation” lines first. This is the step that quietly disappears around week four, once the drafts start looking right.

The approved version goes into the system of record, Procore for Dana and maybe SharePoint for you. Every fix gets one line in a corrections log, and when the same correction turns up twice she fixes the project instructions instead of the draft.

If Dana ever shares the project with Luis, Work mode stops drafting from it, which is why the Remember chapter says to keep it private. She sends Luis the template instead.

Where it stops

Her update doesn't need Repeat or Run on its own yet. She'll package it when a second PM runs the same update and schedule it only after a month of clean reviews, and even then the schedule drafts and a person sends.

Engineering decisions, safety calls, contractual notices and anything that commits the company stay with the person accountable for them. If a sentence in the draft could be read as a formal notice, the person who signs notices writes it.

Two more jobs to test
Operations: turn an approved procedure into a draft checklist that keeps every requirement and lists unclear steps separately, with a competent person reviewing anything safety-critical. Preconstruction: draft a proposal outline from approved discovery notes, separating agreed scope, assumptions, exclusions and open questions.

Try thisRun the weekly steps on last week's notes, for an update you've already sent, and compare the draft line by line with what actually went out.

Chapter 09

Rolling it out

One PM with one job is a pilot. Most of what it takes to make it a company habit belongs to whoever runs the rollout, and your part fits in a week.

Before Friday

  1. Find out what you can paste. Ask whoever owns the rollout. Until there's a rule, keep pricing, contract terms and personal details to yourself.
  2. Pick one job and name its reviewer. Just one, until it works.
  3. Run it on last week's material. Use an update you've already sent, so you have something true to compare against.
  4. Keep personal use on your personal email. One login can reach both a personal and a company workspace, and moving personal content into a company workspace can't be undone.

What your data settings settle, and what they leave open

  • Training. By default, OpenAI doesn't train its models on content from Business, Enterprise, Edu or Healthcare workspaces. On a personal account it's a setting, and you can switch it off.
  • Visibility. Colleagues can't see your chats by default, but workspace admins can view, export and delete them.
  • Temporary chats. They stay out of your history and memory and aren't used for training, though they can be kept for up to thirty days for safety review.

Pitfalls, in the order people hit them

  • Treating a missing menu item as a bug. It's usually a rollout or a role permission.
  • Calling everything “memory.” Team definitions belong in the project instructions.
  • Scheduling a task inside a project to work on that project's files. The task can't see them.
  • Letting the review lapse. It goes quietly, once the drafts start looking right.

For whoever runs the rollout

The rest of this chapter is for the person setting ChatGPT up for everyone else. If that isn't you, skip ahead to the glossary.

A rollout, in order

  1. Settle the workspace first. Company or personal accounts, who owns it, and what happens to someone's chats when they leave.
  2. Write a one-paragraph rule for what can go in, and put it where people work, such as the team's Teams channel.
  3. Start read-only on any connected system, with the negative test.
  4. Measure the real process before and during: drafting time, review time, rework and errors.
  5. Put a settings review on the calendar, because the product will have changed by then.

Access and accounts

Work mode and Codex access are four separate switches, set per role under Permissions & roles.

SwitchWhat it allows
Work CloudStarting and viewing cloud Work tasks on any supported surface
Work LocalWork in the desktop app on local folders; on its own, no cloud tasks
Codex LocalCodex in the desktop app; unaffected by either Work switch
Codex CloudCodex tasks on OpenAI-managed machines
  • Codex Cloud access is separate from Work Cloud and Codex Local, and managing shared environments takes one more permission.
  • Admins can set a starting model and reasoning level for Work and Codex. Chat has its own separate default, and a default never grants a model a role can't use.
  • The Pro setting on the model picker isn't the same thing as a Premium seat.
  • A standalone Business plan supports single sign-on but doesn't include SCIM provisioning. Test what happens to a ChatGPT account when you disable someone in your identity provider.
  • Members can't export their own Business workspace data; that request goes through the workspace owner.
  • In ChatGPT for Healthcare, event-triggered tasks aren't covered by the Business Associate Agreement and mustn't touch protected health information.

Moving custom GPTs into plugins

OpenAI plans to retire custom GPTs and help creators move them into plugins. Here's what carries over.

Part of your GPTWhat happens to it
InstructionsBecome a skill inside the new plugin
Knowledge filesCopied into the plugin's reference files
Connected appsAdded to the plugin as apps
Custom actionsDon't carry over. Anything that depends on them needs a replacement, which may take technical setup
The selected modelDoesn't carry over
Sharing settingsDon't carry over. A migrated plugin starts private

Migration uses the GPT's latest published version, so unpublished edits stay behind. The original keeps working, read-only, until it's retired, and permission to use someone else's GPT doesn't let you migrate it. Retirement dates vary by workspace, so go by the notice in your own account. Before you migrate, save a few real prompts, including one hard case, and run them against the new plugin. The GPTs with custom actions are where the real work is.

If your company builds software

Codex is the mode for writing, testing and reviewing code. If your team maintains a reporting script or an internal app, a non-developer's job is to manage the process. Ask for one bounded change at a time, such as “add the cost code to the export, with a test,” read the list of what changed, and treat releasing it as its own decision. Codex has its own automations, separate from ChatGPT's scheduled tasks.

Glossary

ChatGPT
OpenAI's AI assistant: the product that holds everything below.
Chat
The conversational mode, for questions and quick help.
Work
The mode for longer tasks that end in a finished deliverable, supervised while it runs.
Codex
The mode for software development.
Model
The engine doing the reasoning. Names change often; your picker shows what you have.
Thinking slider
How much the model reasons before answering in Chat: Instant, Medium, High, Extra High and the Pro setting.
Project
One place for the chats, files and instructions of a piece of ongoing work.
Memory
Personal preferences ChatGPT keeps about you. Team definitions belong in project instructions.
Canvas
A document beside the chat that you edit by pointing at what to change.
Deep research
A planned, multi-source investigation that returns a cited report.
App
A connection to another system, such as SharePoint or Google Drive.
Company Knowledge
Answers company questions from the sources you're already allowed to see.
Skill
Reusable written instructions for how a job is done.
Plugin
A package of skills, apps or both, plus reference files.
Scheduled task
Work that starts at a set time, or when something happens in a connected app.
Workspace
Where your chats, files and settings live.

What this guide leaves out

Prices, seat counts, usage limits, model lists and retirement dates. They change often and differ by account, and a stale number does more harm than none. Reviewed October 10, 2026; your workspace may show something newer. ChatGPT, Work and Codex are products of OpenAI. Nothing here is legal, engineering or safety advice.

Where to go nextGo back to the rung your job is on and do that chapter's exercise. An hour in the product with a real job teaches more than a second read.

Formwork — building beyond today
Follow Formwork Follow Naram
Getting Started with ChatGPT — reviewed October 10, 2026 · updated October 13, 2026
formworkgroup.co · naram@formworkgroup.co
© Formwork LLC. All rights reserved.