daily nerdingGive ChatGPT a job, not a question
Give ChatGPT a job, not a question
ChatGPT Work takes an assignment and carries it to a finished file, spreadsheet, deck or site, even with your laptop shut. Nate's 68-minute guide, boiled down to the one question before every task, the words that keep it on track, and the checks before you trust it. No coding needed.
Start with where the job lives, or jump to phrases to steal and prompts to copy.
fakenerd.ai/nate-chatgpt-work · Video: youtu.be/Risal7sjYms · Timestamps are minutes into the video.
The video
68 minutes, posted Sept 25, 2026. Section titles on this page follow the video's own title cards. Chapters are rebuilt from the captions, because the ones in the video description run early. Every timestamp opens the video at that moment.
- 1:08 Chat, Work or Codex
- 2:45 Cloud or local
- 5:13 Voice, dictation, screen context
- 8:27 Editing a doc without a rewrite
- 11:27 Cleaning a messy spreadsheet
- 14:51 Photos and receipts into a table
- 17:56 Research that keeps disagreement
- 21:23 One task: report, workbook, deck
- 24:14 The browser
- 27:21 Connected apps
- 30:10 Projects and where files live
- 33:11 Computer history
- 36:32 Instructions and skills
- 39:41 Which model, how much thinking
- 42:40 Multiple agents
- 45:27 Sites
- 48:13 Goals and /plan
- 51:08 Scheduling
- 53:59 Quiet monitors and triggers
- 57:05 Handing work to someone
- 1:00:02 Usage and tokens
- 1:03:32 Stopping a bad run
- 1:06:43 Where to start
Eleven signals
The whole video in eleven boxes. Each links to the part of this page that explains it.
- 011:08
Give it a job, not a question
Chat gives you an answer. Work takes an assignment and carries it to a file, an analysis or an update inside an app.
How - 024:17
First ask where the material lives
A cloud task does not inherit your local files. A better model won't fix that: it's an environment problem, not a model problem.
How - 0311:17
Say how much to change
Review, small edit or complete rewrite are three different assignments. Pick one out loud, and work on a copy the first time.
How - 0411:48
Define clean before it cleans
Write down what a duplicate is, remove exact duplicates first, standardize after, and make the row counts add up. Formulas stay formulas.
How - 0516:32
Blank beats zero
If the source doesn't say, leave it blank and mark it for review. Keep the source file name on every row so you can check it.
How - 0619:32
Keep the disagreement
Ask for an evidence table before the report, and for disagreement to be shown, not smoothed over. A long source list isn't conviction.
How - 0724:14
Its browser, your final click
Searching, planning and buying are different permissions. Have it bring back options with direct links; the final choice stays yours.
How - 0836:32
Save the method that worked
Project instructions and skills hold the decisions you don't want to repeat. Test a skill on a different input before you trust it.
How - 0942:07
More thinking won't fix the wrong file
The model and the reasoning effort are two separate dials. Raise effort for hard judgment, not for missing information.
How - 1051:26
Test by hand, then schedule
Run it once, inspect, adjust, then repeat on a timer. A schedule repeats your mistakes too. Monitors should stay quiet until there's something to do.
How - 111:00:02
Watch the tokens, know how to stop
Work and Codex share one allowance. When a run looks stuck, ask what's done, what's left and what needs you before you retry.
How
Words on this page
Twelve terms, in plain words. Skip this if you already hand real jobs to ChatGPT or Codex.
- Token
- The unit your usage is counted in. The model, the reasoning, the material it reads, the tools it uses and how long it keeps going all spend tokens. Work and Codex draw on one allowance.
- Agent
- A separate worker the main task sends after one part of the job, like one per retailer in the sofa hunt. The main task then combines what they bring back. The Ultra effort setting sends sub-agents after separate parts the same way.
- Slash command
- A short command you type starting with a slash, such as /goal, /plan or /usage. It does one specific thing instead of being read as a normal prompt.
- Library
- Where you put a file you will reuse across tasks, instead of attaching it to one task. For material that keeps changing, connect the live source instead.
- AGENTS.md
- A guidance file that can sit in a local project folder. The assistant reads it for that project. Only the primary folder's instruction files apply automatically.
- Markdown file
- A plain text file with light formatting marks for headings and lists. A skill is really just one of these, written like a recipe.
- Static number
- A value typed into a spreadsheet cell instead of calculated by a formula. It looks finished but does not change when an input changes.
- Batch write
- Adding or changing many records in another app in one go, like events in a calendar. Check what is already there first, and read it again afterwards.
- API key
- A key for signing in through OpenAI's developer API rather than your ChatGPT plan. If you sign in with one, billing follows the API, not your plan.
- Deployment
- Putting a site online so it can be opened. Preview, private deployment and public sharing are separate steps.
- Version control
- The tool that keeps earlier versions of code so you can go back to one. For documents, revision history or a saved copy does the same job.
- Slop
- Careless AI output that looks finished but isn't worth someone's time. Check a file before you hand it over.
Chat, Work or Codex
Same account, three doors. Choose the one that can reach your material and finish the job. 1:08
Chat
A question, an explanation, a quick back and forth.
An answer. Finding the file, pasting it in and building the slides is still your job.
Work
An assignment carried through to something you can use.
A file, an analysis, or an update inside an app. It works where the material lives.
Codex
Software first, but it also does research, documents and work across apps.
The same core abilities as Work. OpenAI's own advice: already using Codex for non-coding work? Stay there.
Astra is a model you pick inside Work and Codex, not another app. Plus includes it as it rolls out. Don't see it? Update the desktop app and check the model picker before deciding it's missing. Nate's written guide starts you on the desktop app, the Plus plan and the Ask for approval permission setting. 2:10
Cloud work needs reachable files
Before every task: where is the material this job needs? That picks the part of Work. Whether you'll walk away picks cloud or local. 2:45
Where is the material?
While it runs, you'll be…
Use
Give Work that folder
Runs
Local, so you can't walk away
This material only exists on your machine. Keep the computer on, the app open and prevent-sleep turned on. Or move the files somewhere connected and run it in the cloud.
Say: Name the folder, what may change, and what must not. “Move files, but don't edit what's inside them.”
Check: Ask for every file's old and new location, and anything else it changed.
Cloud
Keeps going when you close the app, turn the computer off, walk away for the night.
The catch: It does not inherit your local files. Attach them, or put them somewhere connected, and confirm it can reach them before you leave.
Local
Reaches a folder, an app or anything else only this machine has. Runs in the background and can be scheduled.
The catch: The computer must stay on with the app running. Turn on prevent-sleep for a long job.
The easy mistake: you pick cloud so you can leave, then point it at a file on your desktop. And a fancier model won't fix it. “Changing the model does not solve an environment problem.” 4:15
Just talk through what you need
Nate has had 20-minute, one-sided conversations with AI, just telling it what he wants. Spoken context carries more than you realize. 6:40
Dictation
Speech becomes text you can read before you send. Quickest way to get a long prompt down.
Voice chat
A live conversation. Start work, ask how it's going, change direction, all without the keyboard. Fewer tools than typed tasks.
Screen context (Mac)
Lets it see the window you mean when you say “this part right here”. It can include text beyond the visible scroll area, so pick the window on purpose.
What to say out loud: who it's for, which part they care about, “keep the recommendation short but leave the evidence available”. For a complicated brief, ask for a short written read-back first, so a wrong audience or a mistaken permission gets caught before it shapes the whole job. Set a voice hotkey so starting is effortless. 5:18
Choose the level of change
Work can edit a Google Doc or Word file where it already lives. The useful instruction is smaller and more precise than “rewrite this”. 8:27
- 01
Review
Read it and leave notes. Nothing in the file changes.
- 02
Small edit
Change only the parts you name. Everything else stays exactly as it is.
- 03
Complete rewrite
A new document from the same material. Only when you mean it.
Say what changes and what stays
“Keep the headings, preserve the links, change the explanation in this section, add comments below the relevant paragraphs.”
In his recorded run: a copy of an article with six short red review notes. The original stayed untouched; the copy kept its formatting, links and paywall markers. 9:16
Before you trust it
- First important attempt? Work on a copy. “Don't modify till you know what you're doing.”
- Asked to paste a doc you already connected? Ask if it can read the connected one.
- New doc: give a real example of the format, who reads it, and what they do next.
- Open the actual file: what changed, how pages render, tables running off the page.
A formula should still work
Before it cleans anything, tell it what clean means. Changing the look, changing which records exist, and changing how formulas work are three different jobs. 11:27
Order of operations, from the supplier workbook
- 01
Define a duplicate
Here: exact agreement across every original field. Write the rule down first.
- 02
Remove exact duplicates
Before touching anything else. 236 of 1,516 rows went.
- 03
Then standardize
Fix supplier names only now. Done first, two different records can suddenly look the same.
- 04
Reconcile
Start total, minus removed, equals final: 1,516 − 236 = 1,280. All 7 real grant records survived.
Make it a workbook you can work in
- Filters and frozen headers.
- Dates stored as dates. One that only looks like a date sorts wrong.
- A currency symbol isn't proof the value was converted. Check.
- Formulas stay formulas. Test: change an input and watch the result move. Plausible static numbers look finished and aren't. 13:41
- Have it reopen the exported file and check that, not its working draft.
Photos into rows: 30 receipts, 30 traceable rows
Name the fields before it starts (merchant, date, total, currency, category, needs review). Add one line: “keep a way to get back to the source”, and every row gets its file name. 15:00
Two receipts had no tax amount. A blank and a zero are different answers: “If the source doesn't say, leave it blank and mark it for review.”
Check completeness before formatting
- Did every image produce a row?
- Did any image appear twice?
- Did it skip files that don't belong to the job?
- Is there a source file name on every row?
- Are unknown values blank and flagged, not zero?
Test a few photos first. A subtotal beside a total, an issue date beside a due date or a handwritten correction can hide ambiguity the model smooths over. Each correction you make becomes the rule for next time.
Connect evidence to the claim
Will it tell you when the research disagrees with you, or just flatter you? Ask it to preserve disagreement. A confident summary hides the uncertainty you need to decide. 17:56
Ask for a compact evidence table before any report
- 01
Claim
One sentence the report will rely on.
- 02
Source
The kind that fits the question: product docs, customers, or market data you understand.
- 03
Dates
When it was published and when the event happened. A fresh page can retell an old story.
- 04
Uncertainty
What's shaky, and whether two sources are using different definitions.
Then push on it
“What would change it? Which missing piece of evidence matters most? What's the cheapest useful next test?”
A recommendation is still the goal. In his market run, the customer segment, price and product direction were labeled hypotheses, with the next step being real customers. “Don't confuse a long source list with conviction.” 20:14
One task, three outputs
His meeting-notetaker research came back as a workbook, a written report and a 14-slide deck from the same sources. Give each output its own job: 21:45
- Workbook: Holds the detail.
- Report: Develops the reasoning.
- Deck: Shares the decision. Point-stating titles, sources in the speaker notes.
The slide-title test
Read only the titles. If they're topic labels, ask for titles that state each slide's point and tell a story. Then check each slide for crowding.
Charts you can edit
A picture of a chart is useless the day someone asks you to update the numbers. Give it a real template if one must be followed.
Corrections everywhere
Changed an assumption? Say: update the workbook, the report and the deck, then confirm they agree. One task helps; it won't happen on its own.
Save it somewhere real
Don't leave your only usable copy in a temporary cloud workspace.
Its browser, your final click
Work's browser is its own session. Being signed in on your laptop doesn't sign it in. Use the takeover step to enter a password or code yourself. Never paste credentials into the chat. 24:14
The Japan trip: cheapest isn't the answer
- Connecting round trip~$770
- Non-stop Tokyo round trip$848
- Into Tokyo, home from Osaka$855
The third costs $7 more and skips roughly $90 of backtracking. Which is best depends on the trip, not the fare. Prices were from that one search. 25:26
How to brief the browser
- Constraints that decide usability: dates, travelers, location, occupancy.
- Ask for direct links, and what's included, conditional, verified, or still unconfirmed.
- Steer mid-run when early results show a tradeoff. Don't wait for it to finish.
- Searching, planning and buying are different permissions. “I want the final choice presented to me.”
Connect an app
Open Plugins, pick the service, read what it provides, install, connect the account. Then start a new task so the ability is there.
Name the account
“My calendar” is ambiguous when several are visible. Ask whether it can do the exact read or write you need.
Before a batch write
Check for existing records, then read the destination again afterwards. Catches duplicates and fake “success”.
Say what you're not asking
School calendar image to Google Calendar: timed events approved, a date-only closure kept without an invented time. Adding an event isn't sending invites. Say the time zone.
Use the right source for the job
A project keeps related tasks, sources and instructions together, so the next assignment doesn't start with you rebuilding the background. 30:12
- 01
This task only
Attach the file.
- 02
Reuse later
Put it in the Library.
- 03
Keeps changing
Connect the live source. Last month's saved copy won't update itself.
- 04
Ongoing work
Make it a project. Instructions carry audience, terms and expectations to every task.
Local project folders
A local project can attach several folders, one marked primary. New tasks start there and project guidance is found there. A second folder is more like a book the AI might pull off the shelf: its instruction files don't automatically apply. Attach only what the work needs. 31:20
Same task for corrections that need the history. New task for independent work. And no files named final, final 2 and final really.
Computer history (optional)
Finds “the doc I had open before lunch” without the file name. Mac only, Pro, Business and Enterprise, off by default, needs memories on. Start with an include-only list of work apps. 33:11
Per current docs: no screenshots, no mic or system audio, no private browsing. Typed and on-screen text can still be sensitive. Some processing happens on OpenAI's servers even though the history is stored locally. Recording other people's messages without consent: turn it off.
Save the repeatable workflow
Save the decisions you don't want to repeat. The model can't infer standing rules from today's prompt. 36:32
- 01
Project instructions
Standing rules for one area of work: who you write for, which source wins, where results go.
- 02
Skill
A reusable method for one type of job. Really just a markdown file, like a recipe.
- 03
Plugin
Skills and connections packaged together.
- 04
Connection
An ability or access the task didn't have before.
The skill trim: 10,000 words → 1,200
One of Nate's writing workflows had grown to about 10,000 words of required reading, loaded every conversation. He cut it to 1,200 and kept every standard, source rule and authority boundary. What went: repeats, and steps most tasks didn't need. Real token savings. 37:45
A local project may have an AGENTS.md for guidance. Have the assistant help write it, specific to this work. Don't dump every prompting tip you've ever read into it.
Making a skill that works
- Start from a job that already worked: method, example format, the corrections, the checks.
- Try it on a different input, or you saved one good file, not a method.
- Only for work you really repeat, several people need, or where your judgment keeps getting lost.
- Someone else's skill: inspect it before installing. It can contain code, not just instructions. 39:18
Which model, and how much thinking
Two separate dials under the prompt box. The model is who does the work. Reasoning effort is how long it may spend on it. Options vary by account. 39:41
Astra
Demanding work across several steps and tools. Careful analysis, visual judgment.
Soul
Still useful for a lot of substantial work.
Terra
Practical for familiar everyday assignments.
Luna
Clear, repeatable tasks, or work other agents already structured. Very cost-effective.
Nate's starting points as of Sept 2026. New to Work? His written guide says pick Default.
Raise effort for
Conflicting evidence, a difficult plan, a complicated reconciliation. Not for reformatting a clean paragraph.
Max isn't Ultra
Max gives more room to reason. Ultra sends sub-agents after separate parts, which only helps when the parts really are separate.
Cheaper can cost more
If a routine gets noticeably worse on a lighter option, the savings disappear into your fixing time.
Recurring jobs
Spend enough to get a result you trust. Then test whether a lighter setting repeats it. Compare finished work plus your review time.
More thinking doesn't fix missing information. “Asking it to think longer about the wrong file just makes the wrong task much more expensive.” 42:07
Parallel needs independence
Several agents help only when the pieces can move without each other. Separate stores are perfect; one search doesn't need another's results. 42:40
Multiple agents: the sofa hunt
One agent per retailer, the same requirements for all: budget, dimensions, material, configuration, delivery, return terms. Same return format, so the main task can build a clean comparison. 43:00
The main task reconciles and explains. Keep “delivery confirmed” apart from “conditional”. Skip agents that wait on one decision or edit one file.
Sites: things people can use
29 curated book picks from an article became a searchable reading room. Start from what the visitor needs to do, not “make this a website”. 45:55
Try the controls, try an edge case, try your phone. Preview, private deployment and public sharing are separate steps. A site that looks like a dashboard needs a connected source and a schedule to stay current.
Goals: a finish line it can see
Type /goal and describe the result. It keeps working, checks itself and fixes problems before calling it done. Pause, edit or clear it any time. 48:35
Not “work on this research”. Instead: “compare these options, identify unresolved evidence, and deliver the agreed report format.” Undecided? Start with /plan. Steer in the same task; use side chat for status.
A goal doesn't grant new permissions or keep a shut-down computer available. Goals are for one outcome that deserves sustained work. Work that repeats is a schedule. 50:16
Test the job. Then repeat it.
“Do not schedule an unclear prompt and just pray that repetition will improve it.” 51:26
- 01
Run
Do the job once, by hand, in a task.
- 02
Inspect
Is it what you'd want every week? Each item: a source, a reason to care, a next step.
- 03
Adjust
Fix the instructions now, not after twelve bad runs.
- 04
Repeat
Schedule the version you trust. Say the time zone. Open Scheduled and check the next run.
Existing chat or standalone?
A schedule inside a chat returns to that chat and its ongoing work: good for a continuing investigation. A standalone task starts fresh from the saved prompt every time: cleaner for a report that stands on its own. 52:38
Review the first few runs. Too broad? Change it. Nothing changes that often? Run it less. Stopped being useful? Pause it.
Quiet monitors
“Watch my inbox” is vague. “Look for messages with this subject, extract the dates, notify me only when there's something to do” is a job. 53:59
His school-update monitor found a match whose body just said hello. It reported the missing details once, marked it handled, and has stayed quiet since. Quiet is the value.
Time or event?
At a set time
The clock says it's time. Cloud schedules run with your laptop off. Local schedules need it on, with the app running.
When an event happens
Native Gmail (by sender or subject), Slack (channel messages, author, thread replies) and GitHub triggers, on web and mobile for eligible plans. A task can have several event triggers but can't mix them with a time schedule.
Test every trigger with one message that should match and one that shouldn't. Drafting a reply isn't sending one. Give it an end: a monitor for next month's decision shouldn't run forever. 56:30
Watch the tokens. Know how to stop.
Work and Codex share one allowance, so switching between them buys nothing. Check Usage & billing or type /usage before a big project. 1:00:02
Two expensive lessons
Nate once burned about 100 million tokens a day on repeated runs he didn't mean to repeat daily. Another time he went to bed with a $300 budget; it was gone by morning and the task still wasn't done. 1:01:51
A schedule multiplies an inefficient instruction. Set frequency by how often the information really changes.
Where tokens hide
- Model, reasoning, material processed, tools, and how long work continues.
- Voice has its own allowance, but tasks started by voice use Work's.
- Computer history spends tokens summarizing. Background features count.
- Signed in with an API key? Billing follows the API, not your plan.
- Astra has sometimes been cheaper than Soul overall. Measure tokens per task, not per chat.
When a run goes wrong
- 01
Ask for status
A quiet screen isn't proof of anything. What's done, what's left, what needs you?
- 02
Correct precisely
“This total doesn't match the source.” Say what not to touch, so it changes only that part.
- 03
Check outside first
An external action with an unclear result may already have happened. Look before you retry.
- 04
Pick the right stop
Stop the run, pause the goal, or disconnect the app. Disconnecting doesn't undo past changes.
- 05
Recover with its own tool
Doc revision history, a saved copy, or version control for code. Ask for that earlier state only.
Evidence, not permission
A web page or attachment can contain text that reads like instructions. It must not override your request or authorize a different action. That's how prompt injection works, so be careful what you tell it to read. 1:05:55
Handing work to someone: give the smallest useful thing
- 01
The answer
When that's enough. Give them the file somewhere they can open, and make sure it isn't slop.
- 02
The evidence
When they need to look for themselves. Share the site or snapshot and check access from their side.
- 03
The method
When they have to do the job again. Inputs, output format and the decisions it depends on, plus a small test run on their own files.
A shared Codex task is a read-only snapshot. It doesn't update as you keep working, and on a personal account anyone with the link can open it. 57:45
Phrases to steal
Things Nate says to type or say, from the captions. Each timestamp opens the video there.
“Keep the headings, preserve the links, change the explanation in this section, add comments below the relevant paragraphs.”
Says what stays and what changes. “Improve the doc” hands the model your editorial decisions.
9:00“A review, a small edit, or a complete rewrite.”
Pick one out loud. They are three different assignments.
11:17“Tell it what clean means.”
Changing the look, changing which rows exist and changing how formulas work are different jobs.
11:48“Keep a way to get back to the source.”
One sentence adds a source-file column to every extracted row.
15:24“If the source doesn't say, leave it blank and mark it for review.”
Blank and zero are different answers, especially for tax.
16:32“If sources disagree, show me the disagreement.”
Then: explain whether they use different definitions, and state your assumptions.
19:32“What would change it? Which missing piece of evidence matters most? What's the cheapest useful next test?”
For a conclusion that feels too broad. Turns more research into conviction.
20:52“Split out the independent research. Give every part the same criteria. Bring me one combined answer.”
The whole multi-agent brief in three sentences.
44:25“Notify me only when there's something to do.”
Turns “watch my inbox” into a monitor that stays quiet and saves tokens.
54:21“Burn through my usage on this problem until it reaches 25% and then stop.”
New with Astra: it can see your remaining allowance.
1:00:55“What is complete? What is left? What needs you?”
The first thing to ask when a task looks stuck.
1:03:56
Rules worth keeping
“The cloud task does not inherit your local files.”
4:17“Changing the model does not solve an environment problem.”
5:05“Don't modify till you know what you're doing.”
9:52“Don't confuse a long source list with conviction.”
21:05“Asking it to think longer about the wrong file just makes the wrong task much more expensive.”
42:07“A site does not automatically become a live dashboard just because it looks like one.”
47:53“Do not schedule an unclear prompt and just pray that repetition will improve it.”
51:26“It's really about tokens per task, not tokens per conversation.”
1:02:47“Material that ChatGPT reads supplies evidence but not permission.”
1:05:58
Prompts to copy
FakeNerd's prompts, built from Nate's rules above. Fill in the [brackets]. Nate's own prompts for ten assignments are in his written guide (the first two are free).
Edit a doc without rewriting it
Make a copy of [document name or link] and work only on the copy. Level of change: small edit. Not a rewrite. Change: [the section or passage] so that [what it should now say or do]. Keep exactly as they are: the headings, the links, the formatting and every other section. Add short review comments below any paragraph you think needs my decision, instead of changing it. When you're done, list every change you made and confirm the original is untouched.
Level of change, what stays, and a copy first.
Clean a messy spreadsheet
Make a cleaned copy of [workbook name]. Don't change the original. What clean means here: 1. A duplicate is a row that matches another on every original field. Remove exact duplicates first. 2. Only after that, standardize [supplier names / other field]. 3. Dates must be real dates and amounts real numbers in [currency], not text that looks like them. 4. Formulas stay formulas. Don't replace them with typed values. 5. Add filters and freeze the header row. Then reopen the saved file and check it. Report: starting rows, rows removed and the rule that removed them, final rows.
Define clean before it starts, and make it reconcile.
Photos into a table
Read the photos in [folder] and build one workbook with a row per [receipt / item]. Columns: source file name, [merchant, date, total, currency, category], needs review. If a photo doesn't show a value, leave the cell blank and mark the row for review. Don't guess, and don't use zero unless the photo says zero. Skip files that aren't [receipts] and list them separately. Start with the first 3 photos and show me the rows before doing the rest. At the end, confirm: every photo produced a row, no photo appears twice.
Name the fields, keep the source, blank beats zero.
Research that keeps disagreement
Research [question]. Use these kinds of sources: [current product documentation / customer reviews / market data]. Before writing any report, give me an evidence table: claim | source | published date | date of the event | how certain If sources disagree, show me the disagreement. Say whether they use different definitions, and state your assumptions. Then make a recommendation, labeled as a hypothesis where it is one. Tell me what would change it, which missing evidence matters most, and the cheapest useful next test. I'd rather have a few sources tied to claims than a long list of links.
Evidence table first, recommendation second.
Compare with parallel agents
Split this into independent searches, one agent per [store / vendor]: [list]. Every agent uses the same criteria: [budget, dimensions, material, delivery, return terms]. Every agent returns the same format: name, price, direct link, each criterion met or not, and what it could not confirm. Then bring me one combined answer: drop options that miss the brief, keep confirmed and conditional apart (for example “delivery confirmed” vs “delivery unclear”), and explain your top pick. Don't buy or reserve anything. The final choice is mine.
Independent parts, same criteria, one answer.
A quiet monitor
Check [inbox / channel] every [interval] for messages [from sender / with subject containing “…”]. When one matches, extract [dates and actions] only if the message states them. If the message is missing that information, tell me once that it's missing. Don't invent it. Remember which messages you've handled so you never report one twice. Notify me only when there's something to do. Otherwise stay quiet. Draft replies if useful, but never send anything. Stop after [date or event].
What counts, what to do, when to stop.
Where to start
Pick a job you know well enough to have a strong opinion about. Put it in the right environment, give it sources that make sense, and ask for something you'd be delighted to use. Correct it, save the method that worked, and add features one at a time, starting with the one that takes work off your plate today. 1:06:43
FakeNerd's take
Copy first, then say how much to change
Work on a copy and pick review, small edit or rewrite before it starts. It costs nothing, works in any AI tool, and prevents most of the damage beginners fear.
The habits travel, the buttons don't
Blank beats zero, an evidence table before the report, test before you schedule: these work in Claude or anything else. Astra, Soul, the plan tiers and the menus are OpenAI's and will move. His own prompts sit in a partly paywalled guide; the six here are ours.
Leave schedules and agents for later
This is a beginner's guide, and the first half is the beginner part. A schedule repeats your mistakes on a timer: Nate's own stories are 100 million tokens a day and a $300 budget gone overnight. Get one job right by hand first.
One clean run each
The workbook, the 30 receipts and the Japan fares are single recorded runs, and the fares were from one search. They show it can work. They don't show how often it does on messier files. Check the output the way he tells you to.
Related on this site: 28 things Codex can do, Riley Brown's tour of the same tools and the same Astra model. Theo on token-maxing, the opposite view on spending tokens on purpose. Nicolas Cole's digital brain, saving what works once so the AI repeats it, like Nate's skills.
Attribution
- Nate B. Jones · How to use ChatGPT Work: the complete beginner's guide (2026) (AI News & Strategy Daily, Sept 25, 2026). Section titles follow the video's title cards; quotes and figures from its captions.
- Beginner's guide to ChatGPT Work, Nate's written version with ten assignments and copyable prompts (partly paywalled). Setup details on this page (desktop app, Default model, Ask for approval, /usage) come from it.
- Tokens, context windows and rate limits, his explainer on why long conversations cost more.
- Nate B. Jones: X · YouTube
- The “where does the job live” picker and the six prompts are FakeNerd's, built from the video.