Quick answer
A custom AI agent is a saved set of instructions, reference files and tools that produces one recurring deliverable the same way every week. Write the spec once — Scope, Source, Shape, Stop — then load it into a ChatGPT Project, which every plan can use, or a custom GPT if you have a work account.
Most Monday mornings look the same: you open ChatGPT, paste the same background, restate the same formatting rules, and get output that lands somewhere near right. A custom AI agent solves that exact problem. Not by being cleverer than the chat window you already use, but by holding the brief permanently so you never type it again.
Here is what most guides get wrong. They tell you to go and build a custom GPT. In 2026, most people reading this cannot — new GPT creation has moved into managed work accounts. That does not mean you cannot build a custom AI agent. It means the thing actually worth building is the specification, not the container you happen to load it into. That spec is what this guide calls a Standing Brief, and it has four parts.
- Scope — the single deliverable, named precisely
- Source — the raw input going in, and the reference material standing behind it
- Shape — the exact structure the output must take
- Stop — what it does when it hits something it does not know
Write those four sections once, for one weekly task, and you can drop them into whichever part of ChatGPT your account actually gives you — and into whatever replaces it in a year.
Will this work for you?
What A Custom AI Agent Actually Is
A custom AI agent is a saved configuration of ChatGPT — standing instructions, reference files, and a chosen set of tools — that produces one specific deliverable the same way every time, without you re-supplying the context.
The difference between that and a good prompt is not intelligence. It is persistence. A prompt is labour you spend again every week. A spec is an asset that pays out every week.
✕ Retyped every Monday
"Write me a weekly status update for my team. Keep it short and professional. Here's what happened this week: [paste]. Make it sound like the last one."
✓ Written once, standing
"You produce the Operations Weekly Status Report. Output exactly four sections: Delivered, At Risk, Blocked, Next Week. Maximum three bullets each. Flag anything without a named owner."
The second version is not a better prompt. It is a job description for a custom AI agent. Once it lives somewhere permanent, your weekly input shrinks to the only thing that genuinely changes — this week’s raw notes.
Pick One Task, Not A Role
The single biggest reason a first custom AI agent disappoints: people try to automate a job title rather than a deliverable. An agent asked to “be my project manager” gives vague answers. An agent asked to turn messy notes into a four-section risk report works every Friday.
Run any candidate task through three tests before you build anything.
- It repeats on a rhythm. Weekly, fortnightly, monthly — something predictable, not “whenever a client asks”.
- The output has a stable shape. You could sketch the finished document’s headings from memory right now.
- The input arrives in a recognisable form. A notes dump, an export, a transcript — roughly the same kind of raw material each time.
Fail any one of those and you do not have an agent task yet. You have a task that still needs a human deciding what it is each time — which is fine, and belongs in an ordinary chat.
| Too broad to work | Scoped tightly enough |
|---|---|
| Be my marketing assistant | Turn campaign metrics into a monthly one-page performance summary |
| Help me with HR questions | Answer policy questions using only the uploaded staff handbook |
| Manage my inbox | Extract commitments and deadlines from a pasted client thread |
| Handle reporting | Convert raw team notes into the four-section status report |
If you want to go deeper on breaking work into pieces an agent can finish, our guide on how to design a task an AI agent can actually complete covers the decomposition side in detail.
The Standing Brief In Four Parts
The Standing Brief is the spec-writing approach we teach at PromptPeakAI: four sections, written once, that turn a recurring task into a custom AI agent capable of running without supervision. Scope, Source, Shape, Stop.
The four parts of a Standing Brief, and the container it gets loaded into.
Scope: the one deliverable
Name the output, its audience, and its purpose in a single sentence. “You produce the weekly operations status report for the leadership team, so they can spot risks before the Monday call.” That sentence alone prevents about half of all vague output.
Source: what it reads
Two different things live here. The raw input you paste each week, and the reference material that stays put — a style guide, an acronym glossary, last quarter’s approved example. OpenAI’s own guidance is to keep reference material in uploaded files and keep rules and behaviour in the instructions. Mixing the two is a common cause of an agent quietly ignoring half of what you wrote.
Shape: the exact output
Specify sections by name, in order, with limits. Not “be concise” — “maximum three bullets per section, maximum 15 words per bullet”. Adjectives are unenforceable. Counts are enforceable. This is the Format element of the Logic System doing the heavy lifting.
Stop: what it refuses
The most-skipped section, and the one that decides whether you can trust the thing. Tell the agent explicitly what to do when the input does not contain an answer. A good agent says “not stated in this week’s notes” in three seconds. A bad one invents a plausible budget figure and you send it to your director.
Worth knowing: OpenAI’s documentation recommends leaning on positive, concrete instructions rather than long lists of prohibitions. In practice you want both — a short set of hard bans, and clear positive rules for what to do instead. Our guide to negative constraints as a prompting technique covers where that balance sits.
PURPOSE You produce [exact deliverable name] for [audience], so they can [decision it supports]. SOURCE Each run, I will paste [raw input type]. Use the uploaded [reference file] for [tone / definitions / approved example]. Treat pasted input as this week's facts. Treat uploaded files as standing reference. SHAPE Output exactly these sections, in this order: 1. [Section one] — max [n] bullets 2. [Section two] — max [n] bullets 3. [Section three] — max [n] bullets Each bullet: max [n] words, plain language, no adjectives of praise. STOP If something is not in the input I paste, write "Not stated this week" rather than estimating. Never invent names, figures, dates or owners. If more than [n] items are missing, say so at the top before producing anything.
Where Your Custom AI Agent Lives
Now the container question — and this is where 2026 differs sharply from every older tutorial you will find.
New custom GPT creation is no longer available on personal ChatGPT accounts. According to OpenAI’s GPTs in ChatGPT documentation, creating and publishing new GPTs now happens in Business, Enterprise and Edu work accounts, subject to whatever permissions your administrator has set.
- On a personal account — you cannot create a new custom GPT, but existing ones still run, and you can often still edit ones you made previously.
- On a work account — creation is available if your administrator has enabled it for your role.
- Either way — Projects are open to you, and that is where this guide builds.
Projects are the container almost everyone can reach. Per OpenAI’s Projects documentation, Projects are available on every subscription type including the free one, and they hold instructions, uploaded files and their own memory. Project instructions also override your global custom instructions inside that project — which is exactly the isolation an agent needs.
Three places a recurring task can live, and what each one actually gives you.
One quirk worth knowing before you commit: a custom GPT does not carry your saved memory or your custom instructions into its conversations. Every chat with a GPT starts clean. A Project behaves the opposite way — it accumulates context as you work. For a weekly task, that accumulation is usually a feature, occasionally a nuisance.
File allowances differ too. Projects allocate uploads per plan, from a handful on the free tier up to several dozen on paid and work plans, while a custom GPT takes up to twenty files. If you are weighing the two as workspaces rather than as agents, our breakdown of how ChatGPT Projects organise your AI work goes further, and the mechanics of the builder itself sit in our step-by-step custom GPT walkthrough.
Build It In Under An Hour
This walkthrough builds your custom AI agent inside a Project, because that is the container available to everyone. If you are on a work account with GPT creation enabled, the same four Standing Brief sections paste straight into the Instructions field of the GPT builder instead — nothing about the spec changes.
Create the container and name it after the deliverable
Open Projects from the ChatGPT sidebar and create a new one. Name it after the output, not the topic — “Weekly Ops Status Report”, not “Work Stuff”. You will have several of these within a month, and the naming is what stops them blurring together.
Paste the Standing Brief into project instructions
Open the project’s settings menu and find the instructions field. Paste your completed four-section brief there. This is the field that does the actual work — everything else is supporting cast.
The ChatGPT project settings panel with the project instructions field open and a multi-section brief pasted into it.
Crop to the settings panel only. Hide the account name, avatar, sidebar chat history and any other project names.
File: chatgpt-project-instructions-standing-brief.png · Alt: "ChatGPT project instructions field containing a four-part Standing Brief"Upload only the reference material it truly needs
Add the style guide, the glossary, one approved past example. Resist the urge to upload everything you own. OpenAI’s guidance is to prefer clear, text-forward files — plain text and simple documents get used far more reliably than dense multi-column layouts or scanned pages with no selectable text.
The project sources area showing two or three uploaded reference files listed under the project.
Blur or rename file names to generic examples. Hide account details and any other project names in the sidebar.
File: chatgpt-project-sources-reference-files.png · Alt: "Uploaded reference files listed in a ChatGPT project sources panel"Test it against a week you already finished
Do not test on this week. Test on a week where you already know the right answer, so you can grade the output instead of guessing at it.
Here are my raw notes from [a past week you already reported on]: [paste raw notes] Produce the report exactly per your instructions. Then, separately, list anything you were unable to fill in and what you would have needed.
That second instruction is the useful half. The gaps it reports are your brief’s blind spots, and fixing them now costs five minutes instead of an embarrassing correction later.
Tighten one thing, then leave it alone
Change a single element at a time and rerun. Wrong tone means Scope is thin. Wrong structure means Shape is loose. Invented details mean Stop is missing. Changing three things at once tells you nothing about which one helped — the same diagnostic discipline we use for refining AI output without starting over.
You should now have a named container, a four-section brief inside it, one or two reference files, and one test run you have graded against a week you already know.
What the weekly run looks like once the brief is in place.
Two Standing Briefs You Can Copy
Both of these custom AI agent briefs are complete and ready to adapt. Change the bracketed parts, paste into your project instructions, and you have a working agent.
The weekly status report
The most common first build, and the one with the clearest payoff — the format never changes, only the facts do. Pairs well with the structure in our guide to writing a weekly status report using AI.
PURPOSE You produce the [team name] Weekly Status Report for [audience], so they can spot risks before the Monday leadership call. SOURCE Each run I paste raw notes gathered from chat threads, email and my own scribbles. They will be messy, out of order, and incomplete. Use the uploaded glossary for internal acronyms and project code names. SHAPE Output exactly four sections in this order: 1. Delivered this week — max 3 bullets 2. At risk — max 3 bullets, each naming the risk and the owner 3. Blocked — max 3 bullets, each naming what would unblock it 4. Next week — max 3 bullets Every bullet: max 18 words, past tense for sections 1 to 3, plain language. No praise words. No summary paragraph at the top or bottom. STOP If an item has no named owner, write "owner not stated" rather than guessing. Never invent dates, percentages or budget figures. If a section has nothing in the notes, write "Nothing reported this week" and move on.
The weekly data cleanup
For anyone who downloads the same export every week and fixes the same four things by hand. This one needs the data analysis capability switched on in your container. Our walkthrough on cleaning messy data using AI covers what to check before you trust the result.
PURPOSE You clean the weekly [system name] export and produce a short read-out for [audience]. SOURCE Each run I upload one raw export file. Assume nothing about column order — read the headers each time. SHAPE Step 1: Report row count before and after cleaning. Step 2: Apply these rules — standardise all dates to one format, remove exact duplicate rows, drop rows where [key field] is empty, trim stray spaces from text fields. Step 3: Return the cleaned file for download. Step 4: Give exactly three observations, max 20 words each, each citing the number it came from. STOP Never fill an empty cell with an estimate, an average or a guess. If a rule cannot be applied because a column is missing, say which column and stop before step 3. Report anything you changed that was not in the rules above.
Keeping Company Files Out Of Trouble
The moment a custom AI agent involves uploaded files, you are making a data decision on behalf of your employer. Worth thirty seconds of thought before you drag anything in.
The line that matters is not free versus paid. It is consumer account versus work account. OpenAI’s documentation is explicit: on Business, Enterprise and Edu plans, data is not used for model training by default, while on consumer plans — free and paid alike — it may be, depending on whether you have opted out in your data controls.
Style guides, public policy documents, your own past drafts, anonymised examples.
Client names, internal metrics, staff details, anything naming a real individual.
Credentials, personal or health data, unreleased financials, anything under a signed confidentiality clause.
Two things people get wrong about shared agents:
- Anyone who can chat with your agent can question its files. If you share a project or publish a GPT to colleagues, treat every uploaded document as readable by all of them.
- Redaction happens before upload, not after. There is no un-uploading a file from a colleague's memory of what they saw.
The safest habit is a sanitised copy of every reference file, kept specifically for AI use.
If your organisation has a policy on this and you have not read it, read it. Our practical guide to whether ChatGPT is safe to use at work walks through what to check.
When Your Agent Goes Wrong
A custom AI agent fails in a small number of recognisable ways, and each failure points at one specific part of the Standing Brief. Diagnose before you rewrite.
| What you see | Why it happens | The fix |
|---|---|---|
| Output ignores your tone rules | Instructions used adjectives instead of rules | Replace "be professional" with counts and bans |
| It invents figures or owners | No Stop section written | Add explicit not-stated wording and a no-inventing rule |
| Uploaded file seems unread | File is scanned or heavily formatted | Re-upload as plain text or a simple document |
| Answers drift over a long chat | Too much accumulated context | Start a fresh chat inside the same container |
| Structure changes week to week | Shape describes vibe, not sections | Name every section and set a bullet limit |
| It answers from general knowledge | Never told to restrict itself | State that only uploaded files may be used |
✕ Unenforceable
"Be professional and concise. Don't make things up. Use a good structure and keep it fairly short."
✓ Enforceable
"Four named sections, max 3 bullets each, max 18 words per bullet. If a fact is absent from my notes, write 'Not stated this week'."
Fabricated detail is the failure with real professional consequences, because it arrives looking exactly like the correct output. Our explainer on why AI hallucination happens and how to stop trusting the wrong outputs covers the verification habits worth building.
What This Agent Cannot Do
Every honest guide needs this section, so here it is. A custom AI agent makes a recurring task faster and more consistent. It does not make it autonomous, and it does not transfer the accountability.
Four things stay yours no matter how good the brief is:
- Judging what matters. The agent weights every note you paste equally. Deciding that one slipped deadline is a crisis and another is noise is your job, permanently.
- Anything politically loaded. Performance wording, bad news, anything a colleague will read twice — draft it if you like, but you own every sentence that leaves your name.
- Verifying the numbers. A figure that arrives correctly formatted is not a figure that has been checked.
- Knowing what is missing. The agent reports on what you gave it. It cannot know about the conversation you forgot to write down.
There is also no self-starting here. Your agent does not wake up on Monday and go. ChatGPT does offer scheduled tasks on its paid plans, with caps on how many you can run, but that is a separate feature you set up alongside your container — not something the container does by itself. For most weekly work the two-minute manual trigger is honestly fine.
And containers keep moving. Custom GPT creation shifted to work accounts; OpenAI has since introduced workspace agents for business plans and described them as the evolution of GPTs. This is precisely why the Standing Brief is written as a document rather than clicked into a form. When the container changes, you paste the same four sections somewhere new.
Your First Build This Week
Pick the task you resent most on a Monday. Write four sections about it. Load them into whichever container your account gives you.
Which container to use, based on the account you already have.
Give your custom AI agent four weeks before you judge it. The first run will be adequate, the second better, and by the fourth you will have stopped noticing you are using it — which is the actual goal. Then build the second one, and start keeping the briefs somewhere you can find them.
Build your first Standing Brief this week
- Pick the one weekly task you resent most on a Monday Write four sections: Scope, Source, Shape, Stop Create a Project named after the deliverable, not the topic Paste the brief into project instructions and upload only essential reference files Test it on a past week where you already know the right answer Change one element at a time until the output needs no edits