Find the
friction
How I built SparkFlow with AI
Three ways we use computing power
The agent is powerful.
It only knows what you tell it.
I talk to agents all day.
I was typing all of it.
Speak
Say the request the way you think it
Clean up
Filler words removed, sentences formatted
Copy
One click
Paste
Straight into any AI agent
Build the context infrastructure with three Cs
Context
Why it exists, and what matters most
Components
How the project breaks into parts
Criteria
Rules the agent turns into tests
The exact wording of the prompt doesn't matter. The content does.
Tell the agent why, and what matters most
Break the project into parts
Speech capture
Browser speech recognition
Text cleanup
Remove fillers, format sentences
History and topics
Save and organize thoughts
Time saved
Activity rings and goals
Hosting
A static web page, nothing to install
Usage tracking
Anonymous adoption counts
Each part is a small, focused task. The agent solves one at a time, so each step is more likely to succeed, and so is the whole project.
The bigger the project, the more this matters.
Turn criteria into tests the agent runs
Each criterion becomes a test. The tests check every part after each build, so you only look at what fails.
You won't know everything up front
I didn't know either
My first version stacked every block in one long column.
I found the criteria by trying
Each problem I noticed became a new rule.
It gets faster every time
By the second app, you start with a much better sense of what you need.
With natural language, the bottleneck is how clearly we define and check what we want.
What's your friction?
You'll start saving what you learn
Style guide
Palette, buttons, layout rules
Development workflow
Criteria before building, checks before pushing
Usage-tracking playbook
Written for an agent, reusable on the next app
The agent reads them on the next app. No re-exploring.