A Beginner's Guide to Using AI Chatbots at Work
This guide is for anyone using Claude, ChatGPT, or Gemini through their chat interface. It covers two types of usage: one-off prompts for daily tasks, and reusable prompts you can install in a Project or custom GPT.
Two reminders before we start.
Confidentiality — Use your enterprise subscription. Never paste client data, identifiable amounts, or internal documents into a personal AI account. Your professional subscription is what keeps your data private and GDPR-compliant.
AI makes mistakes, and that's normal. Current models produce errors in roughly 20–25% of their responses. To put that in perspective: a human who is right 80% of the time is already performing well. AI is a remarkably fast and versatile junior assistant — but it's an assistant, not a decision-maker. Verification, judgment, and the final call remain yours. That's precisely where your value lies.
Part 1 — The Daily Prompt
Eight reflexes for a good result in 1 to 3 exchanges.
A daily prompt is a message you type (or dictate) into an AI chat for a one-off task: research, drafting a note, reviewing a document, summarizing a report.
1. Be curious — AI can teach you how to use it better
This is the most important advice. AI evolves quickly, and the best way to improve is to explore. Test things, ask about its capabilities, ask it to suggest how to frame your request.
"I need to analyze a 30-page fund manager report. What's the best way to submit it to get a useful analysis? What questions should I be asking?"
"Which AI model would be best suited for comparing term sheets — Claude, ChatGPT, Gemini? What are your strengths and limits on this type of task?"
"I don't know your capabilities well yet. What can you do that I'm probably not using?"
2. Talk to it — voice mode is your ally
You don't have to type everything. Voice modes let you dictate your prompt, provide context, and give feedback out loud. It's faster, more natural, and lets you give far more context without effort.
In practice: open voice mode, explain your situation in 30 seconds as if you were briefing an intern, then ask for the deliverable.
As of March 2026: voice mode is available on the Claude and ChatGPT mobile apps. On desktop, it's available in the ChatGPT app but not yet in Claude.
3. Say precisely what you want
Replace vague requests with a clear brief: what, for whom, what format, what length.
❌ "Summarize this annual report."
✅ "Summarize this annual report in 5 key points for the investment committee. Each point in 2 sentences max. Highlight significant changes compared to last year."
4. Specify the audience
A memo for a client and an email to a banker are very different: tone, level of detail, vocabulary. Give context.
"This note is for a 70-year-old family patriarch, experienced with markets but not with structuring jargon. Stay factual, avoid technical terms, and close with a clear recommendation."
5. One clear task per message — break complex requests into steps
Pasting a 50-page document works well — as long as you specify precisely what you're looking for. Mixing multiple tasks in one message degrades quality. The AI ends up choosing what it thinks is the priority, and that may not match yours.
A concrete example — analyzing a PDF bank statement:
Prompt 1: "Here is a bank statement in PDF. Convert it to Markdown format."
Prompt 2: "Now organize this data into a table with columns: Date | Description | Amount | Balance. Verify that opening balance + sum of transactions = closing balance. Flag any discrepancy."
Prompt 3: "Classify each transaction into these categories: Management fees | Investments | Income | Tax | Other. Show the total per category."
This step-by-step approach is more reliable than a single prompt asking for everything at once. And if you do this type of analysis regularly, you can save these 3 prompts as a reusable sequence.
6. If you don't know where to start, ask questions first
When you're not sure how to approach a topic, reverse the role — let the AI help you frame your thinking.
"I need to prepare a note on restructuring a client's real estate assets. I'm not sure where to start. Ask me the 7 most important questions to properly frame this note."
7. Iterate instead of starting over
The first result is a draft. Say precisely what's wrong.
"Good content but too technical for the client. Simplify the vocabulary, cut by 30%, and add a 'next steps' paragraph at the end."
8. Keep a critical eye
AI tends to answer with confidence even when it's not sure. Three reflexes to stay in control.
Allow it to doubt. Without this permission, models invent rather than admit a gap.
"If you're not sure about a point, say so clearly and indicate what I should verify. Don't fabricate an answer."
Ask for a confidence score — then challenge it.
"For each point in your response, give me a confidence score out of 10."
Then dig in: "You gave 6/10 on the tax treatment of capital gains. Why? What are you missing to be more confident?"
Ask for a 'things to verify' section. The AI identifies the points it's least sure about, focusing your review where it matters most.
"At the end of your response, add a 'Things to verify' section listing the claims you're least certain about and the sources I should consult."
Part 2 — The Structural Prompt
Building reusable prompts for recurring tasks.
1. Ask the AI to help you write the prompt
You don't need to master the art of prompting — explain your context and objective, and ask the AI to propose a structured prompt.
"I work in an investment team. Every month I receive fund manager reports (10–30 pages). I need to extract: performance vs benchmark, main portfolio moves, identified risks, and generate questions for the next call with the manager. Help me create a reusable prompt I can install in a Project to run this analysis on each report."
Test it on a real report, note what's missing or off, and ask for improvements. In 3–4 iterations, you'll have a rules file — the core of your structural prompt.
2. Build your rules file
The rules file is the set of permanent instructions that frame the AI's behavior at every use. Here's an example structure: CONTEXT: You assist an analyst at a multi-family office. ROLE: Buy-side analyst, rigorous, cautious, capital-preservation oriented. RULES:
Base yourself only on the document provided. Always distinguish facts from interpretations. Flag anything that seems inconsistent or unusual. Conclude with 5 questions to ask the manager. Max 500 words unless otherwise specified. SOURCE OF TRUTH: The report provided. If information isn't there, state "Not mentioned in the report." OUTPUT FORMAT: Table — Performance vs benchmark | Key moves | Points of attention | Questions for the manager
This file goes into the "Project Instructions" of a Claude Project or the instructions of a custom GPT.
3. After each conversation, enrich your rules file
This is the mechanism that makes your prompts better over time.
"We just finished this analysis. Taking into account all the feedback and corrections I gave you during this conversation, what improvements should I make to my rules file so the next analysis is better from the first draft? Propose an updated version."
The AI will identify friction points, missing instructions, and phrasing to sharpen. Update your file, and your next use will be better. Every use improves the prompt.
4. Build quality control into your prompt
"Before giving me your final result, verify that: (1) all figures cited are present in the source document, (2) allocations total 100%, (3) the stated performance is consistent with quarterly figures."
If you're not sure how to frame these checks, ask the AI:
"How would you verify the reliability of this analysis? What controls would you put in place?"
Quality control is the single biggest lever for reducing error rate.
5. Identify 3 to 5 recurring tasks and build a structural prompt for each
Don't try to automate everything at once. Start with what you do most often and where a time gain would be immediately visible: drafting emails in your style, analyzing a recurring document type, preparing a committee agenda, running a sanity check on financial data.
6. Test, learn, improve — continuously
A structural prompt is never "finished." It improves with every use if you maintain the discipline of testing and correcting.
The virtuous cycle:
- Use the prompt on a real case
- Note what works and what drifts
- Ask the AI to propose improvements to the rules file
- Update the file
- Test the new version on the next case
7. Where to install your structural prompts
| Tool | How |
|---|---|
| Claude Project | Create a Project, paste the rules file into "Project Instructions", add your reference documents |
| ChatGPT GPT | Create a custom GPT with the rules file as instructions |