How to use AI effectively at work: brief it like a new hire

The people who get great work out of ChatGPT and the people who get mush are using the same model. What differs is how much the model knows about the job.

I run hundreds of AI agents across my own companies and client companies. Running them taught me one lesson that carries straight over to a person at a keyboard. It fixes most of what goes wrong when a business owner or anyone on their staff sits down at ChatGPT, Microsoft Copilot, Gemini or Claude and types one line.

Treat AI like a smart new hire who doesn’t know your business. That’s most of how to use AI effectively at work. Tell it who you are, who the client is and how the job is done here. Give it the files. Say what it may do on its own and what it brings back to you. Check its work until it has earned your trust.

Why does AI give generic answers at work?

AI gives generic answers at work because it starts every conversation knowing nothing about your company. It doesn’t know your customers, your house style, your terms or what you promised someone last week, so it fills the gaps with the average of everything it has read. A brilliant new hire on their first morning has the same problem.

The companies that make these tools say so in their own manuals. Anthropic’s prompting guide for Claude puts it this way: “Think of Claude as a brilliant but new employee who lacks context on your norms and workflows. The more precisely you explain what you want, the better the result.” Its “golden rule” is a test anyone can run: “Show your prompt to a colleague with minimal context on the task and ask them to follow it. If they’d be confused, Claude will be too.”

OpenAI measured the cost of skipping context. In GDPval, a test of models on real tasks from 44 occupations graded by industry experts, OpenAI researchers led by Tejal Patwardhan cut the prompts to 42% of their original length on average, dropping details like where to find the data and what the finished file should look like. GPT-5’s work got worse. In the paper’s words, “the models struggled to figure out context.”¹

Workers who use chatbots feel that gap, even without a name for it. Among U.S. workers who have used AI chatbots for their jobs, Pew Research found that 40% call them extremely or very helpful for doing things more quickly. Only 29% say the same about improving the quality of their work.² Speed is easy to get. Quality takes context.

How do I get better results from ChatGPT?

Give ChatGPT a brief instead of a one-line prompt. Say who the work is for, what you’re working from, what a good result looks like and what it must not change. Microsoft, Google and OpenAI each publish a formula for this, and all three come down to the same handful of parts you’d cover with a new employee.

Microsoft’s Copilot prompt guide lists goal, context, source and expectations, and warns that “you’ll often need to include more than a goal to get the results you want.” Google’s Gemini guide uses persona, task, context and format. OpenAI’s ChatGPT prompting page uses goal, context, output and boundaries. Different labels, same brief.

Here is what that looks like on a job a 50-person company does every week, answering a quote request.

Before: “Write a reply to this quote request.”

After: “I run sales at a 50-person commercial HVAC contractor. A property manager we’ve done two jobs for wants a quote to replace two rooftop units; her email is below. Draft my reply. Thank her, then ask for the three things we need before we can quote: the unit model numbers, how we get onto the roof, and her preferred install window. Keep it under 150 words and friendly. Don’t mention prices and don’t promise a date.”

The first version gets a polite, generic email that you rewrite. The second gets something close to sendable, because the model now knows who’s writing, who’s reading, what the email is for and where the lines are.

My request for this post was one sentence. The brief the writer worked from ran to about 400 words: who the reader is, the argument, the examples it should use, and what the post may not say. That expansion is most of the work, for a person or a model.

Not every question needs a brief. OpenAI’s own page says “a short prompt is often enough,” and it is, when nothing about the answer depends on your company. When I was asked which job I’d leave in plain ChatGPT, my answer was one that doesn’t need company context or tools, like a daily news digest.

How do I stop re-explaining my job to ChatGPT every time?

Put the context that doesn’t change somewhere the AI reads it every time. Custom instructions carry who you are and how you like work done. A project, or notebook, holds one recurring job: its instructions, its files and its past chats. Every major tool has both now, under different names.

Tool	Standing instructions	One recurring job, with its files
ChatGPT	Custom instructions (Settings > Personalization) and memory	Projects, which “keep related chats, files, and instructions together”
Microsoft Copilot Chat	Custom instructions (Personalization) and Copilot Memory	Copilot Notebooks: chats, files, pages and meeting notes in one place, with their own instructions
Gemini	Skills, “reusable custom instructions,” which are replacing Gems (personal accounts first, work accounts later)	Gemini Notebook, the new name for NotebookLM
Claude	Instructions for Claude, and memory	Projects, with project knowledge and project instructions

Plans and settings differ by tool and by workspace.³ The habit is the same everywhere: write the context down once and let the tool reuse it.

Take the supplier emails a purchasing lead writes every week.

Before: every chat starts from zero. “Write an email to our fastener supplier. They moved our delivery to the 28th and we need it by the 21st.”

After: a “Supplier emails” project with standing instructions. “I handle purchasing for a 50-person distributor. Our suppliers are long relationships: be firm on delivery dates and warm on everything else. Never accept a price change in writing; say I’ll review it. Sign as me.” The project also holds the supplier list and each supplier’s terms. Now the request is one line, “The fastener supplier moved us to the 28th, we need the 21st. Draft the reply,” and the reply already sounds like the company.

The agent that drafted this post works the same way. Before every run it reads a brief on who we write for, a guide to how I write and a page of my standing rules. When a correction should hold for every post, it becomes a dated line on that page, so I don’t have to make it twice. A project with good instructions gives a person at the keyboard the same benefit.

Should I ask ChatGPT to ask me questions first?

Yes, have ChatGPT ask you questions first on anything longer than a quick reply. End the brief with “Before you start, ask me the questions you need answered, one at a time.” The model is good at spotting what your brief left out, and a minute of answers up front beats a confident first draft built on guesses.

OpenAI writes it into its own example. The sample project instruction on its Projects help page is “Act like my marketing mentor. Be concise. Use bullet points. Ask clarifying questions.”

Try it on a document nobody wants to write from scratch.

Before: “Write an onboarding checklist for a new dispatcher.”

After: “I manage dispatch at a 60-truck delivery company. Write the first-week checklist for a new dispatcher. Before you write anything, ask me up to six questions about how we work, one at a time. Then draft the checklist by day.”

The questions are the useful part. Which system do we dispatch in? Who signs off on overtime? What does a new dispatcher handle alone in week one? Your answers turn a generic checklist into yours, and they’re the same answers a new person would have needed from you anyway.

I hold my agents to the same rule from the other side. The one that drafts this blog may not guess a fact about my company. When it has no source, it leaves a question for me in the draft, and those questions come to me with the draft for review.

How do I check AI’s work for numbers and facts?

Check every number, name, date and quote against its source before it leaves your desk. Make the AI show its work: ask it to say where each figure came from and to flag anything it couldn’t verify. Accuracy is the top problem small businesses report with AI in the Federal Reserve’s survey, and it’s yours to catch.

In the Federal Reserve’s Small Business Credit Survey, 46% of small employer firms using AI named accuracy as a challenge, more than any other.⁴ OpenAI’s own help page lists “fabricated quotes, studies, citations or references to non-existent sources” among the mistakes to watch for, and says: “Use ChatGPT as a first draft, not a final source.”

Being good with the tool doesn’t fix this. In a field experiment with 758 BCG consultants, those using GPT-4 finished 12.2% more tasks and worked 25.1% faster. On one task that sat outside what the model could do well, the AI users were 19 percentage points less likely to get the right answer. The group that also got a prompt-engineering overview did worse on that task than the group with AI alone.⁵ Knowing how to prompt didn’t protect anyone from a wrong answer.

The same habit works on the report every company runs.

Before: “Summarize this month’s P&L for the owners.”

After: “Here are the October and September P&L exports. Write a five-line summary for the two owners. Use only numbers from these files, and after each number name the line it came from. Flag every line that moved more than 10% and anything that doesn’t reconcile. Don’t guess at causes the data doesn’t show.”

Then do your part. Pick the three numbers the owners will act on and find each one in the file yourself.

On this blog, a second agent checks every fact against its source before a draft reaches me, and any number it can’t trace comes out. My invoicing agent drafted invoices for three months before it sent one on its own. Check the work closely at first, then less, the way you would with a person. Trust comes from checking.

How do I use AI safely at work?

To use AI safely at work, use a business account and never paste in anything your company’s AI policy rules out. If there’s no policy yet, start with customer personal data, health and financial records, passwords and anything under an NDA. Personal ChatGPT, Gemini and Claude accounts can use what you type to improve their models, depending on settings; business accounts don’t by default.

Here is what each vendor says about its business tier:

• ChatGPT Business and Enterprise: “By default, we do not use your business data for training our models” (OpenAI). On personal Free, Plus and Pro accounts, OpenAI “may use your content to train our models” unless you turn off “Improve the model for everyone” (OpenAI).
• Microsoft Copilot Chat with a work account: “the prompts, responses, and data accessed through Microsoft Graph aren’t used to train foundation models” (Microsoft).
• Gemini in Google Workspace: “Your content is not human reviewed or otherwise used for Generative AI model training outside your domain without permission” (Google). For personal Gemini, Google’s advice is blunter: “Please don’t enter confidential information that you wouldn’t want a reviewer to see” (Google).
• Claude for Work: “By default, we will not use your inputs or outputs from our commercial products (e.g. Claude for Work, Anthropic API, Claude Gov, etc.) to train our models” (Anthropic). On Free, Pro and Max accounts, Anthropic trains on chats when the user’s model-training setting is on (Anthropic).

Your staff are probably using personal accounts already. Only 25% of U.S. employees say their organization has communicated a clear plan for AI, according to Gallup.⁶ In BCG’s 2025 survey, 54% of employees said they would use AI tools even if the company hadn’t authorized them.⁷ Microsoft found in 2024 that 78% of AI users bring their own tools to work, and 80% at small and medium-sized companies.⁸

You may already pay for the safer option. Microsoft says Copilot Chat “is available at no extra cost for eligible users who sign in with a Microsoft Entra work or school account,” including Microsoft 365 Business Basic, Standard and Premium. Google’s Workspace plans include Gemini from Business Starter up. Before anyone buys anything, find out what’s already in the plans you have, turn it on, and tell people which account to use.

A one-page starter covers most of it: which tools and accounts are approved, what never gets pasted in, that a person checks anything before it goes to a customer, and who to ask. Have your lawyer read it before you adopt it.

What comes after using AI well?

After you use AI well, the next step is giving it one recurring job that starts without you, the way an email landing or the first of the month starts work for a person. A chat still needs someone at the keyboard to open it, write the brief and read the result.

The four things don’t change. It needs to know how the job is done here, reach the files and systems the job uses, know what it may do on its own, and have its work checked until it has earned trust. Write those down for one recurring job and you have the job description for an AI agent. Helping companies hire that first agent is the work behind the AI automation practice at Fifth Nine (fifthnine.com).

Start with the brief. Pick the task you do most often this week, write down what a new hire would need to know to do it, and paste that in before you type the request.

---

Sources

How this was written: drafted with AI from my notes and Fifth Nine’s operating records, then edited and approved by me.

Footnotes

1. OpenAI, GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks, October 2025, appendix A.2.7. The shorter prompts “omitted additional context such as where to locate specific data within reference files, how to approach the problem, or detailed formatting expectations for the final deliverable.” The paper notes this experiment ran on an earlier version of its task set. ↩

2. Pew Research Center, survey of 5,395 U.S. workers, October 7 to 13, 2024, published February 25, 2025. Among regular chatbot users the figures were 54% for speed and 41% for quality. ↩

3. Examples: ChatGPT projects hold 5 files on Free and up to 40 on Business and Enterprise, per OpenAI’s Projects page. Copilot Notebooks need a SharePoint or OneDrive license. Google is moving Gems to skills for personal accounts in November 2026 and for Workspace business accounts by March 2027. Claude limits Free users to five projects. ↩

4. The 2025 Small Business Credit Survey was fielded September 3 to November 14, 2025 and published in 2026 as the 2026 Report on Employer Firms. The challenge, worded “Accuracy (e.g., bias or misinformation),” was named by 46% of firms using AI (2,203 responses), ahead of finding or adapting tools to meet business needs (43%) and data security and privacy (28%). Report page 28. ↩

5. Dell’Acqua and coauthors, “Navigating the Jagged Technological Frontier,” Organization Science, 2026. Inside the frontier, across 18 realistic business tasks, AI also raised performance by more than 30%. The outside-the-frontier comparison: the control group was correct about 84.5% of the time, the two AI groups 60% and 70.6%. ↩

6. Gallup, AI in the workplace indicator, data as of May 2026. ↩

7. BCG, AI at Work 2025: Momentum Builds, but Gaps Remain, June 2025, survey of 10,635 employees in 11 countries and regions. The same survey found only about a third of employees say they’ve been properly trained. ↩

8. Microsoft and LinkedIn, 2024 Work Trend Index, May 2024, survey of 31,000 knowledge workers in 31 markets. ↩