Agent-Native Hackathon · Builder.io

Drag two notes.
The agent sorts the rest.

A feedback board where the interface and the agent share the same actions, the same data and the same screen state.

stigi.suitz.xyzBuilt with Agent-Native and Claude Code
The problem

Hundreds of customer comments. Hours of sticky notes.

  • Product teams sit on raw feedback: tickets, NPS verbatims, app-store reviews.
  • Turning it into themes is manual affinity sorting, so it happens rarely or never.
  • A one-shot AI summary can't be corrected. Fix one grouping by hand and the rest of the pile doesn't follow.
Make the correction
the instruction.
Stigi02 / 09
The idea

Teach it by dragging. No prompt.

Two notes dragged into a new cluster while the agent is naming it
Real app, sample data: two seed notes in a new cluster, the agent "naming this…"

  • Drag two notes into a new cluster.

  • The agent names the theme and pulls every matching note in from across the board.

  • Each note it moves carries a one-line reason. You stay in control.

Stigi03 / 09
The payoff

One gesture. Fourteen notes move.

The Plan and billing confusion cluster filled by the agent, each note with an Agent reason chip
From the recorded run on the live app (sample data): "Plan & billing confusion" named by the agent, notes carry "Agent:" reasons.
  • The agent named the cluster Plan & billing confusion and moved 14 notes in one turn.
  • Notes glide across the board as the board refetches.
  • Sort with agent then files the unsorted pile in batches of up to 25.
Stigi04 / 09
How it uses Agent-Native

One action layer. Two ways to call it.

You

Drag a note in the UI

→

defineAction()

move-notes
create-cluster
rename-cluster
save-report
view-screen

Agent

Calls the same action as a tool

→

Shared Postgres (Neon)

Every note, cluster and report. Plus app-state: your selection, view-screen, navigate.

  • Same validation, same permissions, same code for the UI and the agent.
  • The action refuses an agent move without a reason.
  • The agent reads your selection through view-screen and can open the report with navigate.
  • The board refreshes on every agent tool call, with fast polling so it works on serverless Vercel.
Stigi05 / 09
The whole flow

From a CSV to an insight report.

  • Import a CSV, or load the labeled sample set of 200 comments.

  • Sort with agent in small batches, each move with a reason.

  • Teach it by dragging when it gets a theme wrong.

  • Write the report: themes, quotes, next steps. Edit it and export Markdown.

The Stigi board with theme clusters and an unsorted tray
Real app, sample data, pre-sorted starting layout with 50 notes left unsorted.
Stigi06 / 09
Built to hold up in public

A live app you can sign up to.

  • Sign-in required. Each account only sees its own board.
  • 40 agent turns per account per day, with a friendly message when they run out.
  • The model key stays server-side only.
  • Every agent turn is one small batch that fits the hosted time limit, checked against a 35-second budget.
  • 123 unit and component tests, plus Playwright end-to-end tests run against the deployed app.
  • Deployed on Vercel + Neon at stigi.suitz.xyz.
  • Sample data is synthetic and labeled everywhere it appears.
Stigi07 / 09
Against the judging criteria

Where Stigi earns its points.

CriterionWeightWhat Stigi shows
Usefulness and Execution40%A complete flow on a deployed app: import, sort, teach, report, export. Signed-in, capped, tested.
Creativity and Design20%The drag-to-teach gesture, a paper-board look, and notes that glide to their new cluster.
Innovation20%The human teaches by doing and the agent generalizes, instead of re-prompting from scratch.
Cleverness and Use of Agent-Native20%One defineAction() layer shared by the UI and the agent, shared app-state, live sync.
Stigi08 / 09
Try it
stigi.suitz.xyz

Sign up with any email and password, load the sample data, then drag two notes together.

Live appPublic demo on ClipsSample data is synthetic
Built with Agent-Native and Claude Code09 / 09