Potential deliverables
- Review of its intended purpose and instructions.
- Realistic test scenarios.
- Prioritized findings.
- Practical improvement recommendations.
AI Evaluation • Prompt Engineering • Workflow Improvement
Your AI sounds confident. Is it getting things right?
A curious mind. A constructive skeptic. A practical approach.
01 · Prompt
“Qualify this lead and decide the next step.”
02 · AI answer
“That sounds like a great fit! Let’s book a call.”
? fit never confirmed03 · Check again
- ✓ Asks about team size
- ✓ Asks current tools
- ? Handles vague replies
The problem
A convincing answer isn’t always a useful one.
An AI workflow can sound polished while missing instructions, making unsupported claims, asking the wrong questions, or producing inconsistent results. Finding those gaps is the first step to narrowing them.
-
Confident, unsupported answers.
The tone is certain, but the claim has nothing behind it.
-
Inconsistent responses.
The same question, asked twice, gets two different answers.
-
Prompts that miss the real task.
The instructions describe one job; the people using it need another.
-
No clear plan for difficult cases.
When a question is ambiguous or out of scope, nobody decided what should happen.
Services
Three ways to work together.
Each starts with a conversation about your workflow. These are proposed offerings; scope is agreed with you.
Prompt Design & Improvement
For teams whose AI outputs are inconsistent or not meeting their needs.
Improve a WorkflowPotential deliverables
- Clearer instructions and reusable prompt templates.
- Examples and response guidelines.
- Rules for asking clarifying questions.
- Guidance on when to escalate to a person.
- Before-and-after testing.
Ongoing AI Quality Checks
For teams that want to keep evaluating a workflow as it changes.
Discuss Ongoing SupportPotential deliverables
- Recurring test runs.
- Review of new failure cases.
- Prompt revisions.
- A concise findings and recommendations report.
What this work is not
It is not cybersecurity penetration testing, and it is not compliance certification. It is also not a guarantee of accuracy: testing helps find problems and guide improvements, but no review can remove every error.
Approach
How I work
Writing a clever prompt is a small part of it. The larger part is testing how a workflow behaves and checking whether the changes helped.
-
01
Understand
Define the task, the users, and what a good result should look like.
-
02
Challenge
Test realistic questions, ambiguous instructions, and difficult cases.
-
03
Improve
Refine prompts, examples, response rules, and handoffs.
-
04
Check Again
Compare results against agreed criteria and document remaining limitations.
Illustrative example—not a client case study
What an AI review can uncover
Scenario: a lead-qualification assistant recommends a sales call before confirming whether the prospect is a good fit. The dialogue below is invented for illustration.
Prospect: We’re a small team and we’re looking at tools for scheduling.
Assistant: That sounds like a great fit! Let’s get you booked in for a call with our sales team.
The issue: missing qualification questions. The call is recommended before anyone knows whether it’s a fit.
Prospect: We’re a small team and we’re looking at tools for scheduling.
Assistant: Thanks. Before I suggest a next step, a couple of quick questions: roughly how many people would use it, and what are you using today?
The proposed change: clearer qualification rules and instructions to ask for missing information.
What to test: whether the assistant follows those rules across varied prospect conversations, including vague, incomplete and off-topic ones.
About Natalie
Curious enough to investigate. Practical enough to improve it.
I’m interested in the gap between what AI appears to understand and what it actually delivers.
My approach brings together careful questioning, creative experimentation, and clear communication. I look for overlooked assumptions—and turn what I find into practical improvements.
FAQ
Questions worth asking.
What is AI evaluation?
It means testing an AI’s outputs against defined expectations, then reviewing where the workflow succeeds and where it falls short.
How is this different from writing prompts?
Writing a prompt is one step. I start by understanding the task, then test how the workflow behaves, refine the instructions, and check the results against what we agreed a good outcome looks like.
Do I need an existing AI workflow?
Not necessarily. A first conversation can clarify whether the project is about improving a workflow you already have or defining an initial prompt-based prototype. I don’t offer software development or integrations.
Can you guarantee accurate answers?
No. Testing can identify problems and guide improvements, but it cannot eliminate every error. I’ll document remaining limitations so you know what to watch.
What about sensitive information?
Please don’t submit confidential or personal data through the contact form. Data-handling requirements should be agreed before any project materials are shared.
Contact
Find out what your AI is missing.
Tell me what you’re building, what isn’t working, and what better would look like.