Cheatsheet¶

The one-page take-home. Print it, pin it.

Getting in and out¶

Command	Does
`claude`	Start a session in this directory
`claude --continue`	Resume the most recent session here
`claude --resume`	Pick an older session to resume
`claude -p "..."`	One-shot, non-interactive (scriptable)
`Ctrl`+`C`	Interrupt the agent mid-flight
`/help`	Everything else

The dials¶

Action	How
Plan mode (read-only, propose first)	`Shift`+`Tab` to cycle modes
Auto-accept edits	`Shift`+`Tab` again
Compact the conversation	`/compact [what to focus on]`
Think harder on the next step	say "think hard" / "ultrathink"
Spawn parallel subagents	just ask: "use N subagents to…"

The seven patterns, one line each¶

Where it runs — tmux on the machine with the data; detach and let it work.
Modes — plan when wrong is expensive; auto-accept when a loop catches errors.
Subagents — fan out reading/searching; keep conclusions, not file dumps.
Context — finite. Put durable state in NOTES.md/CLAUDE.md, not the chat.
Skills/MCP — corrections you've made twice become a CLAUDE.md line or a skill.
Feedback loops — goal + measure + iterate. Build the test harness first.
Artifacts — results to artifacts/ with metrics.json + REPORT.md; make the agent look at its own plots.

Prompts that work¶

# Goal + loop (the big one)
Implement X. Verify against [known limit / synthetic truth / published
figure]; iterate until [tolerance]. Show me the comparison plot.

# Context hygiene
Write NOTES.md capturing what we decided and why, before we continue.

# Honest delegation
Don't write code yet — plan the approach and list what could go wrong.

# Visual feedback
Look at artifacts/fig3.png. Anything pathological?

# Skeptical cross-check
Spawn a subagent whose only job is to refute this result.

A minimal CLAUDE.md for a physics project¶

- Python via uv (`uv run ...`). Tests with `uv run pytest` — run before
  declaring success. Never modify tests/ to make them pass.
- Natural units (G=c=1); document any unit conversion at the boundary.
- Every physics function: docstring with equation + paper reference.
- Results to artifacts/<task>/ with metrics.json and REPORT.md.
- Plots saved to file, never plt.show(). Look at plots after making them.

Pricing intuition (from the intro)¶

Input (prefill) tokens: cheap, cacheable. Long stable context costs little if it doesn't change between turns.
Output + reasoning tokens: the real cost. Spend them on decisions, not boilerplate.
A focused session beats a sprawling one on both cost and quality — compaction and subagents exist to keep sessions focused.