Lab

Experiments, taken to a working MVP.

I stay sharp by building. Each experiment started as an idea I wanted to test, and I took it to something that runs. Each story has the stack, a simple architecture diagram, what was hard, what it costs, and what is not proven yet.

Pulse

Invite-only alpha

Pulse

A proven organizational diagnostic, locked in a DOS program for two decades, rebuilt as a 40-minute voice interview and scored two ways.

Read the story184 commits, 142 merged pull requests.

Esen

Live, early access

Esen

A private coach for leaders on remote calls. It listens from your Mac, never joins the meeting, and leaves a short note.

Read the story101 commits, 45 merged pull requests.

In production

Client work that ran for real, shown with permission and under fictional names.

NovaBiome

Ran in production

NovaBiome

A weekly marketing dashboard for a DTC supplement brand, shown under a fictional name. Every source in one view, with a weekly draft a person edits.

Read the story300 commits, 15 merged pull requests.Open the demo

How I build

Some good feedback from Y Combinator's Paxel.

Paxel reads a builder's Claude Code sessions and reports how they work. Here is what it said about mine.

On Pulse

Ships production AI features fast.

On a client marketing platform

Steers Claude through messy architecture, rejects shortcuts, and turns product ambiguity into operator-ready systems.

On a client training platform

Steers Claude with sharp product judgment, rejects plausible wrong output, and turns prototypes into pull requests and deployments.

On the underwriting engine

Operating at a strong senior-builder level.

Read the Pulse report

Since the reportsThe reports also named things to tighten, and I acted on all of them. Tests run in CI on every pull request, and the newest product, Enable, has 46 test files including accessibility checks. Commit messages follow one convention almost without exception. And I remove old code as deliberately as I add new: between one line in five and one in two, up from about one in fifteen.

What every report found

  1. 01

    Start from the user

    Every build starts from what the person using it is trying to do.

  2. 02

    Steer the agent

    Claude Code in dialogue, up to three sessions at once, stopped the moment the frame is wrong.

  3. 03

    Operate what ships

    Test the product as a user, then fix what breaks until real people get through.

  4. 04

    Write it down

    Every fix and decision goes into the project notes, so the next session starts better.

  5. 05

    Say what isn't proven

    Every story ends with what I do not know yet.

Have an idea you want taken to a working MVP? Tell me about it.

cassandra@instinctai.io

Or use the form.

Instinct AI LLC, Tulum, Mexico