Skip to content

Skills & Plugins

Claude Code supports skills (reusable workflow templates) and plugins (installable packages of skills). Here's what we use in this project.

Plugin: Superpowers

obra/superpowers — an agentic skills framework that adds a 7-phase development workflow to Claude Code. In the Anthropic marketplace since Jan 2026. Installed as a Claude Code plugin.

The skills auto-invoke when relevant — you don't need to call them manually.

Skill When it fires What it does
Brainstorming Before any creative work Explores intent, requirements, and design before writing code
Writing Plans Multi-step tasks Write a plan, get approval, then execute with checkpoints
Executing Plans After plan approval Follows the plan step-by-step with review gates
Test-Driven Development Before implementation Write tests first, then code to pass them
Systematic Debugging Any bug or test failure 4-phase root cause analysis: trace, hypothesize, verify, fix
Dispatching Parallel Agents 2+ independent tasks Spawns multiple Claude Code agents working simultaneously
Subagent-Driven Development Implementation plans with independent tasks Coordinates agents within a single session
Verification Before Completion Before claiming "done" Requires running commands and confirming output before success claims
Requesting Code Review After completing a step Reviews implementation against plan, reports issues by severity
Receiving Code Review When given feedback Prevents blind agreement — requires verifying feedback is technically correct
Using Git Worktrees Feature work needing isolation Creates isolated git worktrees for parallel development
Finishing a Development Branch All tests pass, ready to integrate Presents merge/PR/keep/discard options, cleans up worktree
Writing Skills Creating or editing skills Guides skill authoring and verification

Related repos:

Custom Skills (included in this repo)

This repo ships project-level skills in .claude/skills/. Anyone cloning the repo gets them automatically — Claude Code detects them on session start.

.claude/skills/
  anti-slop-guide/
    SKILL.md          # auto-detected by Claude Code

Skills can also live at the user level (~/.claude/skills/) for use across all projects.

Anti-Slop Guide

Location: .claude/skills/anti-slop-guide/SKILL.md (in this repo) Invoked: /anti-slop-guide or automatically when writing prose

Detects and removes AI writing patterns from prose. We ran it on all 32 MkDocs pages after initial generation — the difference was large. Pages went from sounding like ChatGPT marketing copy to reading like research notes.

See the full reference.

Sources: stop-slop by Hardik Pandya, Wikipedia: Signs of AI Writing, and several community guides. MIT licensed.

Ralph Wiggum

A technique for maintaining persistent background loops in Claude Code sessions. Available commands: /ralph-wiggum:help, /ralph-wiggum:ralph-loop, /ralph-wiggum:cancel-ralph.

MCP Servers

Model Context Protocol (MCP) servers extend Claude Code with new tools. You configure them per-project or per-user. Some relevant ones for research work:

Server What it provides
Google Docs MCP Live read/write access to Google Docs (could replace sync_google_docs.py)
Browser MCP Fetch web content, papers, arxiv results directly
GitHub MCP Direct PR/issue management without gh CLI
drawio-mcp Create and edit diagrams from Claude Code

See the Anthropic MCP docs for setup.

Skills Worth Building

These don't exist yet but would fit the Sutro research workflow:

Research Sprint — enforce the loop from prompting strategies: literature search, gap diagnosis, ranked experiments, one-at-a-time execution, failure analysis.

Energy Audit — run ARD analysis on an experiment script, compare against baseline, identify top memory access bottlenecks.

Google Docs Sync — wrap sync_google_docs.py: pull latest, diff, update cross-references, rebuild site.

More Skills & Resources