Understanding Prompts: How to Talk to AI

Most professionals waste their first month with AI tools because they treat them like search engines. You type a question, get a mediocre answer, and conclude "AI isn't ready yet."

The problem isn't the AI. It's how you're talking to it.

Prompt engineering is the skill of structuring your requests so AI systems produce exactly what you need. Master this, and you turn ChatGPT or Claude from a novelty into a tool that drafts client memos, analyzes contracts, and builds financial models in minutes.

This guide shows you how to construct prompts that work. No theory. Just the specific techniques managing partners and operations directors use daily.

Every effective prompt has two components: the system message and the user message. Think of the system message as the job description. The user message is the specific task.

System Messages: Setting the Rules

The system message defines who the AI is, what it knows, and how it should respond. This is where you set expertise level, tone, and output format.

Basic system message:

You are a senior associate at a mid-sized law firm specializing in commercial contracts.

Better system message:

You are a senior associate at a mid-sized law firm with 8 years of experience in commercial contracts, particularly SaaS agreements and vendor contracts. You draft in plain English, flag ambiguous terms, and always identify missing standard protections (limitation of liability, indemnification, termination rights). Your output is structured with headers and uses numbered lists for action items.

The second version produces dramatically better results because it specifies expertise depth, writing style, what to watch for, and output format.

System message template for professional services:

You are a [ROLE] at a [FIRM TYPE] with [X] years of experience in [SPECIALTY]. You [KEY BEHAVIORS]. Your output is [FORMAT REQUIREMENTS]. You never [CONSTRAINTS].

User Messages: The Specific Request

The user message contains your actual task. Vague requests get vague answers. Specific requests with constraints get usable output.

Vague user message:

Review this contract.

Specific user message:

Review this vendor services agreement. Identify: (1) any terms that deviate from our standard MSA template, (2) missing indemnification language, (3) ambiguous payment terms, and (4) any auto-renewal clauses. Provide findings in a numbered list with page references.

The specific version tells the AI exactly what to look for and how to format the response. You get a usable deliverable, not a generic summary.

Start simple. Add complexity only when you need it.

Level 1: Single-Turn Prompts

One system message, one user message, one response. Use this for straightforward tasks.

Example: Client email draft

System message:

You are a client services manager at an accounting firm. You write clear, professional emails that acknowledge client concerns and propose specific next steps. You never make promises about timelines without checking with the team first.

User message:

Draft an email to Sarah Chen at Apex Manufacturing. She's concerned about delays in her Q4 financial statements. We're waiting on two missing bank statements from her team. Acknowledge the delay, explain what we need, and propose a revised delivery date of March 15.

This produces a complete, send-ready email in 10 seconds.

Level 2: Multi-Step Prompts

Break complex tasks into numbered steps. The AI executes each step in sequence.

Example: Contract analysis workflow

System message:

You are a contracts analyst at a consulting firm. You review vendor agreements for risk and compliance issues.

User message:

Analyze this IT services agreement using this process:

1. Extract key terms: contract value, term length, payment schedule, termination rights
2. Identify risk factors: unlimited liability, broad indemnification, IP ownership issues
3. Compare payment terms to our standard net-30 policy
4. List any missing standard protections from our vendor contract checklist
5. Provide a go/no-go recommendation with specific reasoning

Format each step as a separate section with a header.

The numbered steps force structured output. You get a complete analysis, not a wall of text.

Level 3: Iterative Prompts with Examples

For specialized tasks, show the AI exactly what good output looks like. Provide one or two examples, then ask it to match that format.

Example: Financial commentary generation

System message:

You are a senior accountant who writes executive summaries for monthly financial reports. Your summaries are concise (200-250 words), focus on variances over 10%, and always include specific dollar amounts.

User message:

Here's an example of the commentary style I need:

"Revenue for March reached $847K, up 12% from February's $756K. The increase was driven primarily by the Acme Corp contract ($65K) and higher-than-expected consulting hours from existing clients ($26K). However, gross margin declined from 42% to 38% due to increased subcontractor costs on the Acme project. Operating expenses held steady at $312K. Net income was $98K, down from February's $115K despite the revenue growth. Key concern: subcontractor cost overruns are eroding profitability on fixed-fee projects."

Now write commentary for April using this data:
- Revenue: $823K (down from $847K)
- Gross margin: 41% (up from 38%)
- Operating expenses: $318K (up from $312K)
- Net income: $119K (up from $98K)
- Key driver: Acme project completed, reducing subcontractor costs

The example trains the AI on your exact style, tone, and level of detail. The output matches your firm's standards without extensive editing.

Pattern 1: The Checklist Enforcer

Use this when you need the AI to verify completeness against a standard.

System: You are a compliance reviewer for client onboarding.

User: Review this client intake form against our standard checklist:
- Legal business name and DBA
- Federal EIN
- Primary contact with title and email
- Billing address
- Engagement letter signed and dated
- Conflicts check completed
- W-9 on file

List any missing items. If complete, respond with "Intake complete - ready for setup."

Pattern 2: The Format Converter

Use this to transform data from one format to another.

System: You are a data analyst who converts unstructured information into structured formats.

User: Convert these meeting notes into a project task list with columns: Task, Owner, Due Date, Status.

[PASTE MEETING NOTES]

Use "Not assigned" if no owner is mentioned. Use "TBD" if no due date is mentioned. Set all statuses to "Not started."

Pattern 3: The Quality Checker

Use this to review your own work before sending it to clients.

System: You are a senior editor reviewing client deliverables for quality issues.

User: Review this client memo for:
- Spelling and grammar errors
- Inconsistent terminology
- Vague recommendations (flag anything that says "consider" or "may want to")
- Missing specifics (dates, amounts, names)
- Passive voice

List issues found with the sentence or paragraph where each appears.

Pattern 4: The Template Filler

Use this to populate standard documents with client-specific information.

System: You are a legal assistant who prepares engagement letters.

User: Fill in this engagement letter template with information from the client intake form below.

Template:
[PASTE TEMPLATE]

Client information:
[PASTE CLIENT DATA]

Replace all [BRACKETED FIELDS] with the appropriate information. If any required information is missing from the client data, list those fields at the end.

Your first prompt rarely produces perfect output. Here's how to improve it systematically.

Step 1: Identify the specific problem

Don't just say "this isn't right." Pinpoint exactly what's wrong:

• Too long/short?
• Wrong tone?
• Missing specific information?
• Wrong format?

Step 2: Add one constraint at a time

If the output is too long:

[Original prompt] Keep the response under 200 words.

If the tone is too casual:

[Original prompt] Use formal business language appropriate for a client-facing document.

If it's missing specifics:

[Original prompt] Include specific dollar amounts and percentages for all financial figures.

Step 3: Show, don't tell

If adding constraints doesn't work, provide an example of exactly what you want. The AI learns faster from examples than from descriptions.

Mistake 1: Asking the AI to "be creative"

AI doesn't do creative well. It does pattern-matching well. Give it a pattern to match.

Bad: "Write a creative proposal for this client." Good: "Write a proposal following our standard structure: problem statement, proposed solution, timeline, pricing, next steps. Use the Acme Corp proposal as a reference for tone and detail level."

Mistake 2: Combining multiple unrelated tasks

One prompt, one task. If you need three things done, use three prompts.

Bad: "Review this contract, draft a response email, and update the project tracker." Good: Three separate prompts, each with a clear single objective.

Mistake 3: Assuming the AI knows your context

The AI doesn't know your firm's policies, your client's history, or your internal terminology. You must provide that context explicitly.

Bad: "Draft the standard NDA." Good: "Draft a mutual NDA using our standard template (attached). This is for a potential vendor relationship with a software company. Term should be 2 years. Exclude the non-solicitation clause we normally include for competitors."

Mistake 4: Accepting the first output

The first response is a draft. Always iterate at least once. Add constraints, fix formatting, adjust tone. The second version is usually 3x better than the first.

Start with these five prompts. Modify them for your specific needs.

1. Email response generator

System: You are a client services professional at [YOUR FIRM TYPE].

User: Draft a response to this client email: [PASTE EMAIL]. Acknowledge their concern about [SPECIFIC ISSUE], explain that [YOUR EXPLANATION], and propose [YOUR SOLUTION]. Keep it under 150 words. Professional but warm tone.

2. Meeting notes summarizer

System: You are an executive assistant who creates structured meeting summaries.

User: Summarize these meeting notes into three sections: Decisions Made, Action Items (with owners), and Open Questions. Use bullet points. [PASTE NOTES]

3. Document reviewer

System: You are a quality control reviewer.

User: Review this [DOCUMENT TYPE] for: (1) spelling/grammar errors, (2) inconsistent terminology, (3) missing information based on our standard template. List findings with specific locations. [PASTE DOCUMENT]

4. Data formatter

System: You are a data analyst who creates clean spreadsheet-ready formats.

User: Convert this information into a table with columns: [LIST COLUMNS]. [PASTE UNSTRUCTURED DATA]

5. First draft generator

System: You are a [YOUR ROLE] who drafts [DOCUMENT TYPE] for [AUDIENCE].

User: Create a first draft of a [DOCUMENT TYPE] that [SPECIFIC PURPOSE]. Include sections for [LIST SECTIONS]. Keep each section to 2-3 paragraphs. [PROVIDE ANY RELEVANT DATA OR CONTEXT]

That's prompt engineering. Not magic. Just structured communication that turns AI from a toy into a tool that saves you 5-10 hours per week.

The following prompts are production-ready for the most common professional services tasks. Copy, modify for your context, and add to your internal prompt library.

Lead Qualification Scoring

System: You are a sales qualification specialist. Evaluate inbound leads against the following criteria: [YOUR CRITERIA]. Return a score 0-100, a tier (Qualified/Review/Disqualified), and a one-paragraph summary of your reasoning. Output as JSON.

User: Evaluate this inquiry: [PASTE INQUIRY]

Contract Clause Extraction

System: You are a contract analyst. Extract the following fields from the contract provided: party names, effective date, payment terms, liability cap, termination rights, auto-renewal clause, governing law. Return as a structured table.

User: [PASTE CONTRACT SECTION]

Client Status Report Draft

System: You are a client relationship manager. Draft a concise project status update (under 300 words) from the project data provided. Include: current status, progress since last update, any risks or blockers, and next milestones. Professional tone.

User: Project: [NAME]. Data: [PASTE NOTES OR <GlossaryTerm term="CRM">CRM</GlossaryTerm> DATA]

Meeting Brief Preparation

System: You are a research analyst preparing executive briefings. Summarize the most important information about the following company and contact for a 30-minute business development meeting. Include: company overview, recent news, known pain points for this industry, and 3 suggested discussion angles.

User: Company: [NAME]. Contact: [TITLE]. Industry: [SECTOR]. CRM history: [PASTE NOTES]

Invoice Follow-Up

System: You are a billing coordinator. Draft a professional overdue invoice follow-up email. Warm but direct tone. Do not use the word "delinquent." Include the invoice number, amount, due date, and a clear call to action with your payment portal link.

User: Client name: [NAME]. Invoice #: [NUMBER]. Amount: $[X]. Due: [DATE]. Days overdue: [N]. Payment link: [URL].

For practitioners new to AI tools, the four definitions that underpin everything in this resource site:

Large Language Model (LLM LLMLarge Language Model. The engine behind AI writing and reasoning tools. Examples: GPT, Claude, Gemini.): A statistical model trained on large text corpora that predicts the most likely next token (word or word fragment) given the context. When you send a prompt, the model generates the response based on patterns from its training data. It does not search the internet (unless tool access is explicitly provided), does not have real-time data, and does not know anything about your firm unless you include that context in the prompt.

Prompt: The input you provide to the LLM. A prompt contains at minimum your instruction (user message). More effective prompts include a system message defining the AI's role, context about the specific task, and constraints on the output format. The quality of the prompt is the primary determinant of output quality.

Temperature: A parameter controlling the randomness of the model's outputs. Temperature 0 produces the most deterministic, consistent output - use this for extraction and classification tasks. Temperature 0.7–1.0 produces more varied, creative output - use this for draft generation and synthesis tasks. Most platforms let you adjust this in settings.

Context Window: The amount of text the model can process in a single interaction, measured in tokens (roughly 0.75 words per token). GPT-4o has a 128,000-token context window. Claude Sonnet has 200,000 tokens. Longer documents require either fitting within the context window or chunking + retrieval via a RAG RAGRetrieval-Augmented Generation. An AI pattern where the model looks up your documents before answering, instead of relying on training data alone. pipeline. See What is a RAG Pipeline.

What is prompt engineering and why does it matter? Prompt engineering is structuring requests to AI systems so they produce the exact output you need. The same model produces dramatically different results from a vague prompt versus a well-structured one with a defined role, specific task instructions, output format requirements, and constraints.

What is the difference between a system prompt and a user prompt? The system prompt defines the AI's role, expertise, output format, and constraints - it's the standing job description. The user prompt is the specific task you are asking the AI to perform. Think of the system prompt as permanent context, and the user prompt as the specific request that changes each time.

How do I write a better AI prompt? Four elements: (1) Define a specific role in the system message. (2) Give explicit format instructions - list, table, paragraph, or JSON. (3) Add constraints - word limits, what to exclude, required elements. (4) Provide an example of the output you want when words alone aren't sufficient.

What temperature setting should I use for business AI tasks? Temperature 0 for extraction, classification, structured output, and compliance-sensitive analysis. Temperature 0.7-0.8 for draft emails, proposal sections, and creative reframes. Temperature above 0.9 for brainstorming only.

How do I fix bad AI output? The systematic approach: (1) Identify the specific problem. (2) Add one constraint at a time to the prompt and observe the change. (3) Provide an example of exactly what you want when constraints don't fix it. (4) For consistently poor output on a specific task, restructure as a multi-step prompt with numbered steps.

Related Resources

Learning Center

AI for Non-Technical Leaders (Video Course / Guide)

Multi-part explainer: what AI actually is, how LLMs work (conceptually), what agents do, why this is different from chatbots.

Learning Center

Confidence Thresholds Explained

What AI confidence scores mean, how to set thresholds, how to calibrate over time.

Learning Center

CRM Data Cleanup with AI (Before You Build Anything)

How to use AI to classify, deduplicate, and standardize CRM data. Austin PE firm approach.