dotclaude

.claude/

Goes to ~/.claude/. Applies to every Claude Code session on the machine.

File	Purpose
CLAUDE.md	Global instructions applied to every project
TROPES.md	AI writing patterns to avoid (source by ossama.is)
settings.json	Claude Code settings
statusline-command.sh	Custom status line showing model and context usage
skills/	Skills available across all projects

Skills

Each skill uses disable-model-invocation: true, so it only fires when you type the slash command or the skill name verbatim. The one exception is scholar, which also auto-invokes when no Semantic Scholar MCP server is available in the session.

Skill	Purpose
/brainstorm	Interview-driven design or planning conversation
/research	Literature + codebase + interview research, outputs RESEARCH.md
/to-tasks	Convert PLAN.md (and optional RESEARCH.md) into TASKS.md for the autonomous loop
/handover	Compact the current conversation into HANDOVER.md for a fresh session
/scholar	Semantic Scholar API wrapper (paper search, author analysis, BibTeX export)

autonomous-implementation/

A coding-agent loop that completes one task per Claude session, with a fresh context each iteration. Drop the contents into a project root.

Based on the original idea of the Ralph Wiggum loop by Geoffrey Huntley.

Contents

File	Purpose
loop.sh	The loop driver which runs N iterations of claude -p
PROMPT.md	Per-iteration instructions for the agent
CLAUDE.md	Project-specific stack, commands, and conventions (you fill this in)
PATTERNS.md	Human-curated corrections accumulated over time
TASKS.md	Task list (placeholder until /to-tasks generates it)
ACTIVITY.md	Session log appended by the agent each iteration
.claude/settings.local.json	Minimal sandbox config (writes confined to the working directory)

Workflow

PLAN.md
    │
    ▼
/to-tasks  →  TASKS.md
    │
    ▼
./loop.sh N   (each iteration: read task, execute, verify, log, commit)

1. Write a PLAN.md in the project root.
2. Run /to-tasks to convert it to TASKS.md.
3. Run ./loop.sh <iterations> from the project root.

Each iteration spawns a fresh Claude session that:

• Reads CLAUDE.md, TASKS.md, ACTIVITY.md, PATTERNS.md.
• Finds the first task in TASKS.md where **Passes:** false.
• Executes it, runs verification, appends to ACTIVITY.md, commits.
• Terminates.

The loop stops when all tasks pass or the iteration count is exhausted.

Running the loop

# from project root
./loop.sh 5     # five iterations, one task each

Requires the claude CLI, git, and jq (optional, for token usage stats). Runs with --dangerously-skip-permissions for autonomous operation.

Sandboxing

The included .claude/settings.local.json enables Claude Code's native sandbox:

{
  "sandbox": {
    "enabled": true
  }
}

This confines all bash writes to the current working directory. Reads remain system-wide; new network domains trigger first-access prompts. See Claude Code sandboxing docs for the full schema.

For stronger isolation, run the loop inside a Docker container, dev container, or cloud sandbox (E2B, Fly Machines). At minimum, run on a dedicated branch and review diffs before merging.

License

MIT