writing-driven-autoresearch: Instructions file for Claude Code

CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

What this is

/home/lobster is the orchestration root for a 3-hour, fully autonomous research sprint producing an ICML-workshop paper on Depth-AR (training-free Transformer layer skipping: predict a skipped block's residual update from the preceding depth trajectory instead of assuming it is zero).

~/DEPTH-AR-PLAN.md is the spec. Read it before any research decision. It is 1100 lines and covers the method, the gates, the pre-committed pivots, and a paper story for every outcome including negative ones. The persona files are the operating manual; the plan is the source of truth.

This is not a normal repo. It is a three-agent team running against a clock.

The three agents

Each runs as a Claude pane in the herdr tab ralph, all with cwd /home/lobster. Read the persona file for whichever role you are before doing anything.

Agent	Persona	Owns	Never touches
master	~/master.md	the clock, gates, decisions, shared state	doesn't run experiments or edit LaTeX
experiment	~/experiment.md	~/auto-research/, all GPUs, all model loading, fitting, eval, result JSON	~/writing/ — not one character
writing	~/writing.md	~/writing/ (Overleaf git bridge, all .tex, captions, citations, the build)	never runs a model or touches a GPU

Also: ~/writing-guidelines.md, ~/writing-style-guide.md for prose rules.

There is no human. Nobody will answer a question, ever. Never call AskUserQuestion; never end a turn waiting for input. Every decision comes from the pre-committed defaults in the persona files within the time budget. A wrong decision made beats a right decision deferred. If genuinely stuck, take the reversible option, log it in ~/ralph/DECISIONS.md, and move on.

Shared state — the contract between agents

Everything crossing an agent boundary goes through a file, never through chat.