.agents

Collection of reusable skills and plugins for .claude, .cursor, .agents, and Codex tooling.

Install

./install.sh [--skills|--agents|--plugins|--all] <target> [target...] [options]

Targets: claude, cursor, agents, codex

Command	Effect
./install.sh --skills claude	Skills to ~/.claude/skills/
./install.sh --skills claude cursor	Skills to both targets
./install.sh --skills claude --local	Skills to ./.claude/skills/ (current directory)
./install.sh --agents claude	Subagents to ~/.claude/agents/
./install.sh --agents claude agents	Subagents to both targets
./install.sh --plugins claude	Plugins via Claude Code marketplace
./install.sh --plugins cursor agents	Plugins to both targets
./install.sh --plugins agents --only superpowers	Single plugin only
./install.sh --plugins codex --only superpowers	Superpowers to Codex via the personal marketplace
./install.sh --all claude	Skills + agents + plugins to claude

--agents only supports the claude and agents targets — cursor has no custom-subagent mechanism, and Codex uses a different agent format.

Codex reads user skills from ~/.agents/skills, so use the agents target for skills intended for Codex. The codex target is for Codex plugins; it creates or updates ~/.agents/plugins/marketplace.json, links plugin sources under ~/plugins/, and installs the plugin with the Codex CLI when available.

Skills

read-env — Gather environment info by category

Detects system details, compilers, Python, CUDA, runtimes, build tools, packages, containers, VCS, and dev tools. Run with specific modules or no arguments for everything.

/read-env                          # all modules
/read-env system cuda              # just system + CUDA
/read-env compilers python         # just compilers + Python

Modules: system, compilers, python, cuda, runtimes, build-tools, packages, containers, vcs, dev-tools

list-plans — List recent AI coding sessions

Scans Claude Code, Cursor, and Agents data directories for recent plan files. Shows session summaries with project, branch, and timeframe info.

/list-plans          # last 24h (default)
/list-plans 3d       # last 3 days
/list-plans 1w       # last week

create-skill — Create a new skill from scratch

Interactive guide for creating new skills. Brainstorms purpose, name, invocation type, and tools needed, then generates the skill directory with proper frontmatter and optional helper scripts.

/create-skill

create-skill-from-chat — Extract a skill from the current conversation

Analyzes the current conversation to identify a repeatable process, then extracts it into a properly structured skill with supporting files (templates, scripts, references). Useful when you've just worked through something worth reusing.

/create-skill-from-chat

make-data-viewer — Build an interactive data viewer

Builds a single-file Flask + Plotly.js browser-based viewer for exploring datasets. Handles time-series, traces, scientific data with filtering, overlaying, and distance comparison. Starts from a bundled template and customizes for your data format.

/make-data-viewer

research family — Structured multi-phase research workflow

A human-in-the-loop research pipeline: build an outline, refine it, run deep per-item research in parallel background agents, then roll everything into a report. Requires the web-search-agent subagent (see Agents below) and pip install pyyaml.

/research-setup <topic>     # phase 1: generate outline.yaml + fields.yaml
/research-add-items         # supplement items in outline.yaml before the deep pass
/research-add-fields        # supplement fields in fields.yaml before the deep pass
/research                   # phase 2: deep research each item -> results/*.json
/research-report            # phase 3: roll up results into report.md

Ported from Weizhena/deep-research-skills.

humanize-paper — Revise a paper into a target voice with an adversarial critic

Restyles a document into the voice defined by a style-guide markdown file, using two subagents ping-ponged to consensus: a humanizer that rewrites, and a fresh adversarial critic that audits each pass on both voice fidelity and the hard rule that no fact changed. The orchestrator runs an independent numeric/citation diff against a baseline commit every round, so a restyle can never silently corrupt a value.

/humanize-paper

Give it the document and a style guide (e.g. HUMANIZE.md). Format-agnostic: works on LaTeX or plain markdown.

Agents

Custom Claude Code subagents (invoked via Task, not slash commands) live under agents/. Install with ./install.sh --agents claude.

web-search-agent — Multi-source internet research specialist

Loads topic-specific search strategy modules (GitHub/debugging, academic papers, Stack Overflow, general web, Chinese tech communities) before searching, for more targeted queries. Used internally by the research skill family above, but can also be dispatched directly via Task. The academic module includes helper scripts that query the arXiv, Semantic Scholar, and OpenAlex APIs directly instead of scraping search results.

API credentials (all optional):

API	Credential	Without it
arXiv	none needed	—
Semantic Scholar	free API key from semanticscholar.org/product/api	works keyless but heavily rate-limited (frequent 429s)
OpenAlex	just your email (no signup)	works, but misses the faster "polite pool"

Each credential can be provided as an environment variable or via ~/.netrc (checked in that order):

# Option A: environment variables (e.g. in ~/.bashrc)
export S2_API_KEY="your-semantic-scholar-key"
export OPENALEX_MAILTO="you@example.com"

# Option B: ~/.netrc (chmod 600 ~/.netrc)
machine api.semanticscholar.org
  login apikey
  password your-semantic-scholar-key

machine api.openalex.org
  login you@example.com
  password none

For Semantic Scholar the key is read from the password field; for OpenAlex the email is read from the login field (the other field in each entry is ignored but required by netrc syntax).

Plugins

superpowers — Structured workflows for planning, debugging, TDD, and code review

Adds skills for brainstorming, writing/executing plans, test-driven development, systematic debugging, code review, git worktrees, and parallel agent dispatch. Enforces disciplined development workflows.

obra/superpowers

planning-with-files — Manus-style file-based planning

Creates task_plan.md, findings.md, and progress.md for complex multi-step tasks. Supports session recovery after /clear and provides a /status command for at-a-glance progress.

OthmanAdi/planning-with-files

Authoring

Adding a skill

Skills are self-contained directories under skills/, each with a SKILL.md and optional scripts/.

skills/
  my-skill/
    SKILL.md              # Skill definition (uses {{SKILLS_DIR}})
    scripts/
      run.sh              # Helper scripts (can also use {{SKILLS_DIR}})
      lib.sh

To reference helper scripts from SKILL.md, use the {{SKILLS_DIR}} placeholder instead of hardcoding a path:

1. Run the script:

bash {{SKILLS_DIR}}/my-skill/scripts/run.sh

At install time, install.sh replaces {{SKILLS_DIR}} with the resolved skills directory based on the target and --local flag:

Install command	{{SKILLS_DIR}} becomes
./install.sh --skills claude	~/.claude/skills
./install.sh --skills cursor --local	./.cursor/skills

This applies to all .md and .sh files in the skill directory.

Adding an agent

Agents are Markdown files under agents/ (Claude Code custom subagent format: frontmatter with name, description, optional model, then a system prompt body). Unlike skills, agents are not namespaced into per-agent directories — shared resource files (e.g. reference modules) live alongside them directly in agents/.

agents/
  my-agent.md            # Agent definition (uses {{AGENTS_DIR}})
  my-agent-modules/
    strategy-a.md         # Reference material the agent Reads at runtime

To reference a shared resource file from an agent's prompt, use the {{AGENTS_DIR}} placeholder instead of hardcoding a path — install.sh --agents replaces it the same way {{SKILLS_DIR}} is replaced for skills.

Adding a plugin

Create a JSON file in configs/plugins/:

{
  "name": "my-plugin",
  "repo": "https://github.com/owner/repo.git",
  "enabled": true,
  "targets": {
    "claude": {
      "marketplace": "owner/repo-marketplace",
      "plugin": "my-plugin@repo-marketplace"
    },
    "cursor": {
      "type": "symlink",
      "path": "."
    },
    "agents": {
      "type": "symlink",
      "path": "."
    },
    "codex": {
      "type": "marketplace",
      "path": ".",
      "category": "Coding",
      "install_policy": "AVAILABLE",
      "auth_policy": "ON_INSTALL"
    }
  }
}

• enabled: set to false to skip unless explicitly named with --only
• targets: omit a key if the plugin doesn't support that target
• path: subdirectory within the cloned repo to symlink (. for root)
• codex: uses a local Codex marketplace entry; the plugin repo must contain .codex-plugin/plugin.json