How I Trick AI Into Giving Me Brutal Feedback

The 3 AI Tricks That Make Terrible Estimates Impossible

Estimations are hard. In the last estimation post you learned how to use four Power Questions to cut through the noise and gather the required estimation information. You used the Three-Point estimation technique to home in on the estimation ambiguity and learned a few tricks on how to manage the uncertainty.

In the next post we’ll talk about the hidden danger of buffers and how to manage stakeholder expectations.

But first, there's something we have to talk about: AI.

In today's post, you'll learn how to use Large Language Models (LLM) to further improve the accuracy of your estimations. I'll cover three tips, including one that 95% of technical people don't know about. Let's get started.

Prompting for Feedback

Instead of using LLMs to write for you, use them to critique your work. There's a lot of information out there about writing high-quality prompts. And frankly, it's still a moving target.

Luckily, you don't have to know too much here. I fired up Claude and asked it to:

I want to write a high quality prompt. Ask me questions until you are highly confident that you can craft an excellent prompt for me.

I'll spare you the full back and forth. Here are some of the instructions I shared in response to the AI's questions:

I want to use AI to critique my project estimations. For context, I'm a software development manager.

We usually plan one half ahead. My team has 7 software engineers and we typically work on one big project and 1-4 small projects per half. We're using the estimation methodology described in this post: https://posts.managementdeltas.com/p/your-estimates-are-lies-heres-how My main goal is to meet or exceed stakeholder expectations.

and:

I'll share a 3-point estimation spreadsheet and another text document listing the assumptions I made, such as dependencies on external factors. In the prompt, leave a placeholder for me to include my recent estimation mistakes and pain points. I want the AI to help me identify methodological mistakes as well as technical blind spots.

This was enough for Claude to create a reasonable starting prompt:

# Project Estimation Critique Prompt

You are an expert software engineering consultant specializing in project estimation and risk assessment. I'm a software development manager leading a team of 7 engineers, and I need you to critique my project estimation for the upcoming half (6-month period).

## Context

- My team typically works on 1 large project and 1-4 smaller projects per half
- We use three-point estimation methodology (best case, likely case, worst case) with PERT distribution
- My primary goal is to meet or exceed stakeholder expectations
- I want to identify both methodological mistakes and technical blind spots

## What I'm Providing

1. **Three-Point Estimation Spreadsheet**: Contains task breakdowns with best/likely/worst case estimates
2. **Assumptions Document**: Lists dependencies, external factors, and key assumptions I've made
3. **Recent Pain Points**: [PLACEHOLDER - Insert your recent estimation mistakes, missed deadlines, or recurring issues]

## Please Analyze and Critique

### Estimation Methodology Review

- Are my task breakdowns granular enough, or am I missing intermediate steps?
- Do my best/likely/worst case scenarios seem realistic given the task descriptions?
- Are there patterns in my estimates that suggest systematic bias (too optimistic/pessimistic)?
- Am I properly accounting for integration complexity between tasks?
- Are my confidence intervals (using PERT distribution) appropriate?

### Technical Risk Assessment

- What technical risks or dependencies am I potentially overlooking?
- Are there tasks that seem underestimated given their technical complexity?
- Do I have adequate buffer for integration testing, debugging, and rework cycles?
- Are there potential performance, scalability, or security considerations missing?
- What external API dependencies or third-party integrations could cause delays?

### Team and Resource Considerations

- Given a team of 7 engineers, are there resource bottlenecks or single points of failure?
- Are there knowledge transfer requirements or skill gaps that could impact timelines?
- Do my estimates account for code review, testing, and deployment overhead?
- Are there tasks that require specific expertise that might not be immediately available?

### Stakeholder Expectation Management

- Based on my assumptions, what are the highest-risk items I should proactively communicate?
- Which estimates have the widest confidence intervals and need careful stakeholder communication?
- Are there any "unknown unknowns" that could derail the project timeline?
- What early warning indicators should I track to detect estimation drift?

## Output Format

Please provide:

1. **Overall Assessment**: A summary of the estimation quality and key concerns
2. **High-Risk Items**: Top 3-5 tasks or assumptions that are most likely to cause delays
3. **Methodological Improvements**: Specific changes to my estimation approach
4. **Technical Red Flags**: Technical risks or complexities I may have underestimated
5. **Stakeholder Communication Plan**: What to communicate upfront vs. what to monitor
6. **Follow-up Questions**: Questions I should ask my team or stakeholders to validate assumptions

Be direct and specific in your feedback. I'd rather hear uncomfortable truths now than discover them during execution.

If you want to improve this prompt, ask another AI model to critique it. Most importantly, try it out and tweak it.

Save this prompt. If you don't already have a prompt library, then you should start one. I use Obsidian, but you can use Notion or just a dedicated folder on your machine.

Prompt for Growth

You may have noticed that placeholder I asked the AI to prepare for me. One of the best uses of AI is to use it to amplify and accelerate our personal growth.

The trick to fast learning is to use it on the job. You can use LLMs to surface your growth areas and give you advice exactly when you need it. Instead of you prompting the AI for growth, you should have the AI prompt you for growth.

The way to accomplish this is to have a single global personal growth document. Make sure to feed this document into all of your prompts. This way, the AI can take into account what you're working on right now when it gives you advice on whatever it is you're asking it.

Say you are not in the details. You delegate many of your tasks and don't have the systems to know what's actually going up. This appears in your global growth list. When you use the estimation prompt, the AI will connect the dots between this personal growth area and potential misses in how you scope various tasks without consulting with your team.

Connecting dots is at the heart of modern LLMs and their attention mechanism.

Prompt for Pain

LLMs are trained to please us. If you want to get brutally honest and highly influential feedback then you have an uphill battle to fight.

"Be brutally honest."
"Give me constructive feedback."

or my favorite:

"You are my devil's advocate."

These are all reasonably starting points. They work ... okay.

There's one trick that will help you get much higher quality feedback. Instead of fighting the uphill battle, embrace it. All you have to do is lie. Instead of telling the AI that you are the software manager, tell the AI that you are the stakeholder.

Ask the AI to critique the latest estimation plan that you, the Product Manager, got from your engineering counterpart. You can still list out all of the past mistakes, just frame them as the mistakes of this specific EM. Likewise for the growth areas.

This small trick works surprisingly well. You will leverage the AI's training to please the user to trick it into giving you more brutally honest feedback. And as a nice bonus, you will also develop a deeper appreciation for what your stakeholders are going through.

Summary

We covered three LLM prompting tricks:

1. Use the AI to write the prompt for you.

2. Embed your growth areas into all of your prompts.

3. Create a real Devil's Advocate prompt by fooling the AI that you're the Devil.

What other uncommon tricks do you use?