Two Tools That Are Helping Me Build Meeako AI
In the last post I wrote about Meeako AI slipping into training data instead of its own library. Getting responses grounded and accurate is my highest priority right now. Before I experiment with anything new, I want to lock this down.
Two areas to work on:
give Meeako an accurate snapshot of what's in its library for the question being asked, and
experiment with how to control what the model responds with.
Decision Table
I have used Decision Tables before in my product work, mainly in root cause analysis scenarios. For example, when a key metric drops (say "activation rate of app B has gone down by 5%") and your goal is to systematically figure out why. It maps every possible condition that could explain the drop and what to investigate for each one.
The method has a defined sequence. Identify conditions — the inputs affecting a decision. Define the possible states for each condition. Define the actions — the outcomes. Map every combination. Fill the table. Then simplify by removing redundancies.
For Meeako, the challenge is that a user flow or feature can be referenced across multiple dimensions:
Flow name: "Success metrics for the Submit Prompt flow"
Flow name + app: "What are some success metrics for the Submit Prompt flow of ChatGPT?"
App alone: "How does Etsy measure its performance?"
App category: "Success metrics for Social Media apps"
Flow category: "Metrics for Content Creation flows"
Each dimension can lead to a different result — no flow found, more than one found, the flow exists in a different app than the one asked about. And each result needs a different action from Meeako — ask a clarifying question, inform the user something doesn't exist, or answer the question directly.
The Decision Table makes all of that visible before I write a single prompt. It also includes a sample user question and expected result for each combination, which means it doubles as a test plan for verifying Meeako's behavior.
Decision Table for Flow & App Dimensions of Meeako AI
Claude's Prompt Playground
I am embarrassed I didn't know this existed until recently. But it has become one of my favorite ways to spend time on Meeako. When I'm not actively building, I find myself in the Playground tweaking prompts to see what happens.
It lets me:
decouple prompt engineering from Meeako's codebase. Fewer moving parts, fewer things to go wrong, and one less thing to think about.
iterate and refine prompts faster for different use cases such as routing, tool usage and grounding instructions without the overhead of deploying changes.
The demo shows the same question tested twice in the Playground:
"Can you predict what feature they'll ship next?"
Before refining the prompt, Meeako returns a long, generic response asking multiple questions at once. After adding a simple instruction to ask only one clarifying question, it responds with a single focused line.
Before and after demo of Meeako AI asking a clarifying question
Why these two tools matter together
Being able to reliably communicate with models in a way they understand and control their output is how I meet the Trust principle I wrote about in post, Meeako Needs Guardrails.
The Decision Table tells me what I need the model to do. The Playground is where I figure out how to make it actually do that.