Appendix

Cheatsheet · Claude Code

Structural reference, not a tutorial — dense, lookup-first. Commands and flags drift with releases; treat this as a starting point and confirm against…

Structural reference, not a tutorial — dense, lookup-first. Commands and flags drift with releases; treat this as a starting point and confirm against the live docs when something here feels stale. This is the mechanical-layer half of the course; the lasting layer (the six parts, the four layers) doesn't change even when these commands do.

Install & entry points

Command	Does
`claude`	interactive session in the current directory
`claude -p "<prompt>"`	headless, one-shot — the shape every schedule/cron wrapper uses
`claude --resume`	reattach to the most recent session
`claude --model <name>`	pin a model for the session
`claude mcp add <name> ...`	register an MCP server

Config file shapes

File	Role
`CLAUDE.md`	the rules file — context layer; read every session, mandatory in this repo
`.claude/settings.json`	project harness config — permissions, hooks; committed, shared
`.claude/settings.local.json`	personal overrides — gitignored
`.claude/commands/*.md`	custom slash commands
`.claude/skills/<name>/SKILL.md`	a skill — see Step 9
`.claude/agents/*.md`	subagent definitions (tools, model, prompt)
`.mcp.json`	project-scoped MCP server declarations

The loop primitives

Mechanism	Shape
`/loop <interval> <prompt>`	in-session heartbeat; interval syntax like `5m`, `1h`
`/loop <prompt>` (no interval)	self-paced, run-until-done — the conditional shape
`CronCreate` / `CronList` / `CronDelete`	scheduled-task tools behind Routines; 50-task cap, 3-day expiry, jitter applied, no catch-up on a missed fire
`CLAUDE_CODE_DISABLE_CRON`	env var to hard-disable scheduling — the kill switch's harness layer
`--worktree` / `isolation: worktree`	run a task in an isolated `git worktree`

Permissions & harness guarantees

Directive	Effect
`/permissions`	interactive permission editor for the session
`permissions.allow` / `permissions.deny` in `settings.json`	the enforced list — this is where "never touch `main`" or "never read `.env*`" actually lives, not in `CLAUDE.md` prose
Hooks (`PreToolUse`, `PostToolUse`, etc.)	shell commands the harness runs around tool calls — the lowest-level guarantee short of denying the tool outright

Sub-agents & MCP

Concept	Shape
Agent tool	spawns a sub-agent with its own tool subset, model, and context — used for research/parallel work, never as a call between two independently-scheduled loops
MCP servers	connectors for external systems; declared in `.mcp.json` or via `claude mcp add`; keep the tool count small and each write idempotent (Step 10)

Gotchas

CLAUDE.md is a request, not a guarantee — anything safety-critical belongs in settings.json permissions or a hook, per safety.md.
The 50-task cron cap and 3-day expiry are silent unless you check them — a Routine that "stopped working" is often just expired, not broken.
--resume reattaches to session memory, not your spine — the spine is still the file you wrote, not the transcript.

Sources: Claude Code mechanics — /loop, Cron* tools, permissions, hooks, skills, subagents, MCP — from Panaversity's Agentic Coding Crash Course (S2) and Scheduled Tasks: The Loop Skill & Cron Tools (S4); live reference: docs.claude.com/en/docs/claude-code. Full attribution: resources/sources.md.