ai-safety-plugins

Claude Code plugins for AI safety research: experiment design, academic writing, code review, workflow management, and visualization.

Quick Start

# Add the marketplace
/plugin marketplace add yulonglin/ai-safety-plugins

# Install core (recommended for everyone)
/plugin install core@ai-safety-plugins

# Install domain-specific plugins
/plugin install research@ai-safety-plugins
/plugin install code@ai-safety-plugins
/plugin install writing@ai-safety-plugins
/plugin install workflow@ai-safety-plugins
/plugin install viz@ai-safety-plugins

Plugins

core (recommended: always-on)

Foundational agents and skills that other plugins depend on.

Agents:

Agent	Purpose	External Deps
efficient-explorer	Context-efficient codebase exploration	None
context-summariser	Conversation compression	None
claude	Judgment-heavy delegation via Claude Code CLI	None
codex	Implementation delegation via Codex CLI	Codex CLI + OpenAI key (optional)
gemini-cli	Large context delegation via Gemini CLI	Gemini CLI + Google key (optional)

Skills:

Skill	Purpose
/docs-search	Fast grep-based docs search (fd + rg)
/fast-cli	Modern CLI tool mappings (eza, fd, rg, bat, etc.)
/spec-interview	Interview-based spec development
/task-management	Timestamped task tracking

Hooks (safety guards):

Hook	Event	Purpose
check_destructive_commands	PreToolUse:Bash	Blocks sudo rm, xargs kill, etc.
check_secrets	PreToolUse:Bash	Blocks committing API keys/tokens
check_read_size	PreToolUse:Read	Warns on reading large files without offset/limit
task_force_background	PreToolUse:Task	Forces subagent calls to background
truncate_output	PostToolUse:Bash	Truncates verbose output
pre_plan_create	PreToolUse:Write	Enforces per-project plans
pre_task_create	PreToolUse:TaskCreate	Enforces per-project tasks

research

AI safety research workflows.

Agents: experiment-designer, research-engineer, data-analyst, literature-scout, research-advisor, research-skeptic

Skills: /spec-interview-research, /experiment-setup, /run-experiment, /api-experiments, /read-paper, /reproducibility-report, /generate-research-spec, /mats-slurm, /audit-docs, /new-experiment, /reflect

writing

Academic writing and presentations.

Agents: paper-writer, application-writer, humanizer, writing-clarity, writing-facts, writing-narrative, writing-redteam

Skills: /review-draft, /review-paper, /research-presentation, /slidev, /fix-slide, /clear-writing, /humanize-draft, /strategic-communication

code

Development workflow, code review, and delegation.

Agents: code-reviewer, codex-reviewer, plan-critic, debugger, performance-optimizer, tooling-engineer

Skills: /bulk-edit, /fix-merge-conflict, /deslop

Depends on: core (codex and gemini-cli agents)

workflow

Agent orchestration and conversation management.

Skills: /agent-teams, /custom-compact, /externalise-handover, /custom-insights

Hooks (convenience, opt-out via env var):

Hook	Event	Purpose	Disable
auto_background	PreToolUse:Bash	Auto-backgrounds long commands	CLAUDE_AUTOBACKGROUND=0
check_pipe_buffering	PreToolUse:Bash	Warns about piping anti-patterns	Warn only
auto_log	Pre/PostToolUse:Bash	Audit trail of commands	Async, non-blocking

Depends on: core (context-summariser, gemini-cli agents)

viz

TikZ diagrams and Anthropic-style visualization.

Skills: /tikz-diagrams

Requires: LaTeX (pdflatex/xelatex)

Dependency Matrix

Dependency	core	research	writing	code	workflow	viz
Claude Code v2.1+	REQ	REQ	REQ	REQ	REQ	REQ
Codex CLI	opt	—	—	REQ	—	—
Gemini CLI	opt	—	—	opt	REQ*	—
fd + rg	opt	—	—	—	—	—
bun/bunx	—	—	opt	—	—	—
LaTeX	—	—	opt	—	—	REQ
Modern CLI tools**	opt	—	—	—	—	—
OpenAI API key	opt	opt	—	opt	—	—
Google API key	opt	—	—	opt	opt	—
Python 3.9+	—	REQ	—	—	REQ*	—

* For /custom-insights skill only. ** eza, bat, dust, duf, fzf, zoxide, delta, jq.

Full Setup

# macOS
brew install codex gemini-cli fd ripgrep
brew install --cask mactex  # Only for viz / presentations

# Auth
codex auth           # OpenAI key (for Codex delegation)
gh auth login        # GitHub token (optional)

Optional Reference Docs

Some agents reference documentation from ~/.claude/docs/ for enhanced functionality. These docs are optional — agents work without them but provide richer output when they're available. If you use the dotfiles repo, these are deployed automatically. Otherwise, bundled copies are included in each plugin's agents/references/ directory.

Enable Always-On

Add to your Claude Code settings (~/.claude/settings.json):

{
  "enabledPlugins": {
    "core@ai-safety-plugins": true
  }
}

License

MIT