Beginner's Guide to Modifying AI Agent Proposals

Introduction: Understanding AI Agents

Think of an AI agent like a very capable assistant who needs clear directions. They can do complex work, but they need you to guide them toward the specific outcome you want. This guide will teach you how to communicate effectively with AI agents to get the right result the first time.

Part 1: The Fundamentals

What is a Proposal Modification?

When an AI agent creates a proposal (their suggested action or response), you review it and provide feedback. A modification is your instruction to change what they've proposed to better match what you need.

The Golden Rule

Be as clear with the AI as you would be with a new employee on their first day. Don't assume the AI knows your preferences, your company's specific procedures, or what you're thinking.

🔑 Critical System Concepts

The "Generate New Proposal" Trigger

Sometimes the AI completes a task but doesn't automatically give you a next step. When you don't see a "Proceed" button:

Use: "Generate a new proposal to [specific action]"
This tells the AI to create the next action item in your workflow

The "Next Step" Field

This field determines where items route in your system:

Always use exact step names from your workflow
Wrong step = wrong department = delays
Example: "Set next step to 'Management Review'" (not just "review"

The Three C's of Effective Modification

1. CLEAR

Say exactly what you want
Avoid vague words like "better," "fix this," or "wrong"
Instead of: "This doesn't look right"
Try: "Change the amount to $500"

2. COMPLETE

Include all necessary information in one message
Don't make the AI guess missing details
Instead of: "Use the other account"
Try: "Use the maintenance account ending in 4521"

3. CONTEXTUAL

Explain WHY when it matters
Give background for complex changes
Instead of: "Change this"
Try: "Change to Question For Manager because Homeowner is requesting specific information"

Part 2: The Modification Framework

Step 1: Review What the AI Proposed

Before modifying, understand what the AI has suggested:

Read the entire proposal
Identify what's correct (don't re-explain this)
Note what needs to change

Step 2: Structure Your Modification

Use this simple formula:

[WHAT to change] + [HOW to change it] + [OUTCOME you expect]

Examples:

"The invoice amount should be $1,200 not $1,500 so it matches the purchase order"
"The response message needs to be shorter, maximum 2 sentences so the customer gets a quick answer"
"The next step should be 'AP Review' not 'Review' so it routes to the correct department"

Step 3: Choose Your Instruction Style

Direct Instructions (Most Effective)

Start with an action verb:

"Change the date to..."
"Add a note about..."
"Remove the section on..."
"Update the status to..."

Conditional Instructions

When there are different scenarios:

"If the amount is over $1000, then..."
"When the customer is new, include..."
"Unless already approved, send to..."

Replacement Instructions

When you want to swap content entirely:

"Instead of [current text], use [new text]"
"Replace the entire message with..."
"Delete everything after... and add..."

Part 3: Common Modification Scenarios

Scenario 1: Correcting Information

❌ Poor modification: "Wrong info" ✅ Effective modification: "The address should be 123 Main St, not 123 Main Ave"

Scenario 2: Adjusting Message Tone

❌ Poor modification: "Too formal" ✅ Effective modification: "Make the response more friendly and conversational, like talking to a neighbor"

Scenario 3: Changing Process Flow

❌ Poor modification: "Send somewhere else" ✅ Effective modification: "Route this to Manager instead of Customer Support"

Scenario 4: Adding Missing Elements

❌ Poor modification: "Incomplete" ✅ Effective modification: "Add the invoice number and due date to the response"

Scenario 5: Simplifying Complex Responses

❌ Poor modification: "Too long" ✅ Effective modification: "Summarize in 3 bullet points covering: the issue, the solution, and the timeline"

Scenario 6: When the Agent Stops Without Next Steps

❌ Poor modification: "Continue" ✅ Effective modification: "Generate a new proposal for the next action"

Example: If the agent gives you information but no "Proceed" button appears:

Say: "Create a new proposal to send this response to the customer"
Not: "Now what?" or "Continue processing"

Scenario 7: Ensuring Correct Workflow Routing

❌ Poor modification: "Fix the workflow" ✅ Effective modification: "Change the next step to [specific step name] to route to the correct team"

Example: For action items that need specific routing:

Say: "Update next step to 'Manager Review' before proceeding"
Not: "Send to the right place"

Why This Matters: The "next step" determines how items move through your system. Wrong step = wrong department = delays.

Part 4: Advanced Techniques

Technique 1: The "Instead" Method

When the AI is on the wrong track entirely:

"Instead of creating a work order, this should be a billing inquiry"
"Actually, this is a compliment not a complaint, so thank them instead"

Technique 2: The Priority Method

When multiple things need attention:

"First, correct the account number to 12345"
"Then, add a note about the payment terms"
"Finally, route to supervisor for approval"

Technique 3: The Example Method

When you want a specific format:

"Format the response like this: 'Dear [Name], [Brief acknowledgment]. [Action taken]. [Timeline].'"

Technique 4: The Boundary Method

When you need limits:

"Maximum 50 words"
"Only include charges from this month"
"Limit to business hours responses only"

Technique 5: The Continuation Method

[NEW] When the AI stops without giving you next steps:

"Generate a new proposal to [specific action]"
"Create the next action item for [specific task]"
"Proceed with creating a response to send"

Key Insight: If you don't see a "Proceed" or "Continue" button, explicitly tell the AI to generate a new proposal for the next action.

Technique 6: The Workflow Direction Method

When routing through your system:

"Set next step to '[Exact Step Name]' for proper routing"
"Change workflow status to '[Exact Status]' before proceeding"
"Update action item type to '[Specific Type]' to trigger correct process"

Remember: The "next step" field is like a GPS destination - be exact, or you'll end up in the wrong place!

Part 5: Common Mistakes to Avoid

Mistake 1: Being Too Polite

The AI doesn't need social pleasantries in modifications.

Skip: "Could you maybe possibly change..."
Use: "Change..."

Mistake 2: Explaining What's Right

Don't waste time confirming correct parts.

Skip: "The first part is good but..."
Use: "Change the second paragraph to..."

Mistake 3: Using Pronouns Without Context

The AI might not know what "it," "that," or "this" refers to.

Skip: "Add it to that"
Use: "Add the invoice total to the summary line"

Mistake 4: Giving Conflicting Instructions

Make sure your instructions don't contradict each other.

Skip: "Make it shorter but include all details"
Use: "Summarize the key points in 3 sentences"

Mistake 5: Assuming Knowledge

Don't assume the AI knows your internal processes.

Skip: "Process this the usual way"
Use: "Process using standard approval workflow: manager then director"

Part 6: Quick Decision Trees

When the Amount/Number is Wrong:

State the correct amount clearly
Include units if relevant ($, %, days, etc.)
Mention if this affects other calculations

When the Recipient is Wrong:

Name the correct recipient explicitly
Include their role if relevant
Specify if copy/CC needed to others

When the Tone is Wrong:

Describe the desired tone (professional, friendly, firm, etc.)
Give a sample phrase if helpful
Mention the audience (customer, vendor, internal)

When Information is Missing:

List exactly what to add
Specify where to add it
Indicate the source if not obvious

When the Process is Wrong:

Name the correct process/workflow
Identify the correct first step
Note any special conditions

Part 7: Building Your Modification Skills

Practice Exercise 1: Specificity

Transform these vague modifications into clear ones:

Vague: "Fix the date"
Clear: "Change the date from May 1 to May 15"

Practice Exercise 2: Completeness

What's missing from this modification?

"Send to John"
Missing: John's last name/department, when to send, any message to include

Practice Exercise 3: Action-Oriented

Rewrite these to start with action verbs:

"The amount might need to be different"
Better: "Change the amount to..."

Self-Check Questions:

Before sending your modification, ask yourself:

Would a new employee understand exactly what to do?
Have I included all specific values, names, or numbers needed?
Is there any ambiguity in my instruction?
Have I explained the outcome I expect?

Part 8: Quick Reference Card

Power Starters

"Generate a new proposal to...”
"Change... to..."
"Replace... with..."
"Add... after..."
"Remove... from..."
"Update... to show..."
"Correct... to read..."
"Move... to..."
"Include..."
"Exclude..."
"Combine..."
"Set next step to...”

System Navigation Commands

"Generate a new proposal for [action]" - Creates next action when system stops
"Set next step to '[exact step name]'" - Routes to correct department
"Create separate action item for [issue]" - Splits multiple issues

Precision Words

Specific: "the third paragraph" not "that part"
Exact: "$542.50" not "about $500"
Named: "the maintenance department" not "them"
Timed: "by 3 PM today" not "soon"
Counted: "in 2 sentences" not "briefly"

Result Indicators

"so that..."
"in order to..."
"to ensure..."
"which will..."
"resulting in..."

Part 9: Real-World Application

Before You Modify, Consider:

Is this a one-time exception or a new rule?
- One-time: Give specific instruction for this case
- New rule: Explain the pattern for future similar cases
Is the issue with content or process?
- Content: Focus on what information to change
- Process: Focus on workflow and routing
Could this confuse the AI?
- If yes, break into smaller, simpler steps
- Number your instructions if multiple changes needed

The 30-Second Check

Before submitting your modification, you should be able to:

Point to the exact words/numbers being changed
Explain the change in one sentence
Predict the exact outcome you'll see

Part 10: Troubleshooting Guide

If the AI Keeps Getting It Wrong:

Check for ambiguity
- Are you using "it," "that," "this" without clear references?
- Could your instruction mean two different things?
Verify completeness
- Did you provide all necessary information?
- Are you assuming the AI knows something it doesn't?
Simplify the instruction
- Break complex modifications into steps
- Focus on one change at a time
Provide an example
- Show exactly what the output should look like
- Use quotation marks for exact text
Specify the format
- If structure matters, describe it explicitly
- Use templates or patterns when helpful

If the System Seems Stuck:

No "Proceed" Button?
- Tell the AI: "Generate a new proposal for [next action]"
- Be specific about what proposal type you need
- Example: "Create a new proposal to respond to the customer"
Wrong Department/Team Routing?
- Check the "next step" field in your modification
- Use exact step names from your workflow
- Example: "Change next step to 'AP Review' not 'Review'"

Success Metrics

You know you're writing good modifications when:

✅ The AI gets it right on the first try
✅ You rarely need follow-up modifications
✅ Others can understand your instructions
✅ The AI's responses consistently match your expectations

Remember:

The AI agent is a powerful tool, but it needs your clear direction. Every modification is a chance to teach the AI exactly what you need. With practice, you'll develop an intuition for writing modifications that work the first time, every time.

The goal isn't to write perfect modifications—it's to communicate clearly enough that the AI can deliver what you need.