ROLE SIMS

FOUR JOBS · ONE ACCOUNT OF YOUR OWN JUDGEMENT · NO INSTALL

Each of these puts you in a seat for a fixed span of time with a scarce resource, a hidden ground truth, and people who push back. None of them is about the tools — a chart, a GPU queue, a feature board and a ticket queue are just surfaces. Each is built around one specific way that confident people reason badly, and each scores you against a truth you cannot see until the end. They pause at gates and hand you to an LLM chat window, where the rest of the cast lives.

01

DAY TRADER

Prop desk · three sequential sessions

Three real market sessions, rebased and anonymised, traded on one account that carries your P&L forward. Day 2 is built to punish exactly what Day 1 taught you.

THE ENEMYnoise — a run of luck looks identical to an edge
Capital$25,000
Span3 sessions
Hard stop−$1,500 and risk pulls your card
Mandatorya typed thesis on every entry

GATES → pre-open pitch · midday risk check · EOD P&L review

LAUNCH → readme playbook debrief
02

ML RESEARCHER

Pretraining team · a five-day research week

Recommend at most four of eight interventions for a 70B run. Every experiment you can afford is at least an order of magnitude below the scale you are deciding for.

THE ENEMYscale extrapolation — the best cheap result is worth nothing at size
Budget6,000 GPU-hours · 4 slots
Span5 days × 10 h
Decide for70B — where you have no data
Mandatorya predicted effect + CI per run

GATES → Monday plan · Wednesday midweek · Friday readout

LAUNCH → readme playbook debrief
03

PRODUCT MANAGER

B2B SaaS · one full quarter

Ten features, seven ways of finding out what is true, and five people who all want different things and none of whom reports to you. Unfinished work is worth exactly zero.

THE ENEMYbias — every instrument lies in a knowable direction
Capacity48 eng-weeks
Span12 weeks
Held toactivation 31.4 → 40.0
Second scoretrust — below 40 caps your grade

GATES → roadmap · midquarter · ship-or-cut · QBR

LAUNCH → readme playbook debrief
04

JUNIOR ENGINEER (PRE-AI)

Dispatch team · your first sprint

Six tickets in 412k lines you did not write. Two of them look identical on the board; one is sitting in the code and the other is only in your mentor's head.

THE ENEMYnot knowing what you don't know — and what asking costs
Focus time60 h over 10 days
Real budgetDeepa's 10 hours
Gate to shipunderstanding, not time
Mandatoryan estimate before you start

GATES → kickoff · standup · 1:1 · retro

LAUNCH → readme playbook debrief

HOW EACH ONE RUNS

  1. Open it and read the brief. The start button is at the bottom of that screen.
  2. The first gate fires immediately and asks you to post a plan. Bring it to a Claude window along with that sim's playbook — each playbook is written to work cold in a fresh session.
  3. Work the session. The clock runs, people interrupt, and nothing tells you whether you are being careful or being stubborn.
  4. At the last gate you paste a tearsheet — the review, the readout, the QBR, the retro — and get interrogated on your reasoning rather than your result.
  5. Then the debrief reveals the ground truth against what you actually did. That screen is the product; everything before it exists to earn it.

In three of the four I do not know the ground truth during the gates and will not pretend to — I reason only from the evidence you put in front of me.

Everything runs offline from file://. No build step, no server, no network. Progress persists in browser localStorage, so you can close a tab and come back to it.  ·  Component previews with synthetic data: board · plots · viz · chart