Skip to main content
Tushar Pathak
Let's connect

PortfolioCase study

CB-01Built · not launched

Cubicle

Your first team fits in a cubicle.

One idea.Four visible teammates.Four artifacts.

A solo founder types an idea and watches four AI teammates debate it in the open — then gets the plan.

  • 4

    AI teammates

    Prototype

    PM · researcher · designer · developer

  • 4

    fixed artifacts

    Prototype

    PRD · scan · landing copy · build plan

  • 10 days

    buildathon

    Structural

    a team of six, Sept 7 → 16, 2026

Evidence keyMeasuredStructuralPrototype

Cubicle's real interface after a run: four teammate desks — PM, Researcher, Designer, Developer — each marked Done, above the PRD they wrote.

The problem

Solo builders have no team — so ideas die before the first artifact.

AI tools answer in one generic voice and hide their reasoning. Nobody shows the “why”.

What a solo founder does today
  1. An idea
  2. A generic AI prompt
  3. A generic PRD
  4. Low trust
  5. Nothing shippednever sent to anyone

“He has pasted the idea into ChatGPT three times and got three slightly different, equally generic PRDs that he never sent to anyone.”

Aarav, the target persona — Discovery PRD

The product

Type an idea. Watch the team argue. Get four artifacts.

The debate is the product: every proposal and objection is on screen before a deliverable is written.

Cubicle's home screen: “Your first team fits in a cubicle.” above a “What are you building?” box, with four desks — PM, Researcher, Designer, Developer — each marked Ready.
Idle: waiting for an idea
One run, as designed
  1. Idea
  2. Orchestratorpicks who speaks
  3. Team debatePM · research · design · dev
  4. Stop rule
  5. 4 artifacts
Cubicle after a run: all four desks marked Done above a one-page PRD with Problem, Who it is for, Proposed solution, v1 scope, One success metric and The open question we argued about.
Done: the real UI replaying the repo’s test fixture — not a live run

What every run prints

  • PRD

    by the PM

    • Problem
    • Who it is for
    • Proposed solution
    • v1 scope: in / out
    • One success metric
    • The open question we argued about

  • Competitor scan

    by the Researcher

    • Three competitors
    • What this means for positioning
    • Confidence note

  • Landing copy

    by the Designer

    • Headline
    • Subheadline
    • Three benefits
    • Call to action
    • Two objections, answered

  • Build plan

    by the Developer

    • Smallest v1 slice
    • Suggested stack
    • Five steps with rough time
    • What we cut and why
    • Riskiest assumption to test first

Trust system

Trust first. Ownership second. Autonomy last.

  1. firstTrust
  2. secondOwnership
  3. lastAutonomy
Could have
Maximum agent autonomy
Chose
Visible debate first
Because
Founders need to see the reasoning before they hand over ownership.
How one run moves through Cubicle
  1. Ideatyped by the founder
  2. Orchestratorpicks the next speaker
  3. 4 role agentspropose · question · object
  4. Stop rules6 messages · 45 s
  5. 4 artifactswritten in parallel
  6. Databaseevery message saved
  7. Browsera live mirror

Key learnings

Three lessons from building Cubicle.

  • Visible reasoning creates trust

    The product’s bet: a founder trusts a plan they watched being argued. Untested with users.

  • Bounded debate beats endless chat

    Stop rules and four fixed artifacts keep a run’s cost and time bounded.

  • Read the diff, not only green checks

    The two most important defects were found by reading the reasoning, not by any test.

Evidence

Built and tested offline — not launched.

The QA report’s verdict: “CONDITIONALLY READY — STEPS REQUIRED”.

  • 326

    automated tests passing

    Measuredas of 12 Sep 2026

    3 skipped; typecheck, lint and audit clean in CI

  • 29 / 97

    test cases passed

    Measuredas of 12 Sep 2026

    17 blocked · 48 planned

  • 0

    test cases failed

    Measuredas of 12 Sep 2026

Not measured yet

  • A real four-agent run on a live model
  • Users or usage
  • Live latency (50–75 s is an estimate)
  • Real cost (≈ $0.04 a run is an estimate)

PrototypeNot launched

Source material

Everything here is backed by real artifacts.

  • PRD
  • Research
  • Architecture
  • Design
  • Code
  • Evaluation
  • Build ledger

Evidence behind the Cubicle case study

Private working documents are listed with what they support; only public sources link out.

  1. Discovery PRD

    PRD8 Sep 2026

    Supports: The problem, the Aarav persona, the gap insight and the trust-first sequencing bet

  2. Research notes

    Research

    Supports: 46 secondary sources, tagged High/Medium/Low

  3. Technical plan

    Architecture9 Sep 2026

    Supports: The streaming route, debate protocol, stop rules and the database as the source of truth

  4. Decisions log

    Design

    Supports: Decisions S1–S6

  5. package.json + gateway

    Code

    Supports: Gemini Flash (agents) and Flash-Lite (orchestrator); the unverified-search fallback

  6. Artifact headings

    Code

    Supports: Each artifact’s owner and its fixed headings

  7. QA report

    Evaluation12 Sep 2026

    Supports: 326 tests; 97 TC rows (29 pass, 17 blocked, 48 planned, 0 fail); “CONDITIONALLY READY”

  8. HANDOFF

    Build ledger

    Supports: Offline build finished; the first real run never happened

  9. Lessons learnt

    Build ledger12 Sep 2026

    Supports: L8: read the diff

  10. Buildathon brief

    Research

    Supports: A team of six, ten days

Links and next steps

Team buildSep 2026Built · not launched

Next

Nuptis