Agent Skills Library
Reusable skills, prompts, and automation workflows for OpenClaw, Claude Code, Codex, and other coding assistants.
- Source repo
- jdrhyne/agent-skills
- Stars
- ★ 240
- Last updated
- 1mo ago
- License
- MIT
- Primary language
- Python
- FA score
- 67/100 · Some gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · Claude Code
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- A development team using Claude Code, Codex, and OpenClaw can maintain a shared collection of operating procedures and install selected packages in each assistant.
- Not a fit if
- Teams wanting a standalone app, not a skill library
- Users of assistants other than OpenClaw, Claude Code, or Codex
- Teams unwilling to set up Jira, Salesforce, or Zendesk accounts and APIs
- Source review
- 67/100 · Some gaps
What does this agent do, and when should you use it?
Agent Skills is an MIT-licensed repository of packaged skills and instruction-only prompts for OpenClaw, Claude Code, Codex, and other LLM-based coding assistants. Its content is divided into skills/, prompts/, clawdbot/, and codex/: skills may bundle scripts, templates, assets, API integrations, or multi-step workflows, while prompts contain guidance text only. The catalog covers Jira, GA4, Google Ads, Google Search Console, Gong, Salesforce, Zendesk, document processing, research loops, planning, and multi-agent orchestration. Users copy or symlink selected directories into an assistant's skill path, after which that assistant interprets SKILL.md, runs included scripts, or connects to the relevant service; the repository is not a standalone agent service. The README reports that 79% of its skills work across OpenClaw, Claude Code, and Codex, but some entries are OpenClaw-only and integrations still require the corresponding APIs, CLIs, accounts, and permissions.
Each skill is driven by a SKILL.md file and may include scripts/, templates/, assets, or a structured workflow. Individual packages can read data through the GA4 Data API, Google Search Console, Gong, Jira, Salesforce CLI, and Zendesk; supported workflows also include drafting, creating, or updating Jira work and managing Salesforce CRM records and Zendesk tickets. planner produces structured plans for task-orchestrator, while task-orchestrator performs dependency analysis, parallel tmux/Codex execution, and heartbeat monitoring. autoresearch-loop supplies a metric-driven improvement cycle, context-recovery addresses explicit compaction or truncation, and elegant-reports produces Nordic-style PDF reports. Document-oriented resources cover conversion, OCR, extraction, redaction, watermarking, signing, and inspection, although some rely on Nutrient MCP servers or a pinned plugin. Repository utilities include scripts/update-readme.sh for generating README content from metadata and scripts/validate-skills.sh for checking SKILL.md formatting.
- A development team using Claude Code, Codex, and OpenClaw can maintain a shared collection of operating procedures and install selected packages in each assistant.
- A growth or analytics practitioner can equip an assistant to query or audit GA4, Google Ads, and Google Search Console data.
- Support and revenue-operations teams can use the Jira, Salesforce, Gong, or Zendesk packages to retrieve business data and, where documented, create or update records.
- An engineering lead breaking a project into dependent work can combine
planner,parallel-task, andtask-orchestratorfor planning and concurrent execution. - A team automating PDF workflows can evaluate the Nutrient-related skill and MCP resources for OCR, extraction, redaction, signing, or inspection.
- A skill author can add domain instructions, scripts, and templates using the documented
SKILL.mdlayout and run the repository's validation script.
How do you install or deploy this agent?
Clone the repository first:
git clone https://github.com/jdrhyne/agent-skills.git ~/agent-skillsFor Claude Code, copy an individual skill or create a symlink from the repository directory:
cp -r skills/elegant-reports ~/.claude/skills/ln -s $(pwd)/skills/planner ~/.claude/skills/For Codex CLI, copy the dedicated packages and any universal packages you need:
cp -r codex/codex ~/.codex/skills/
cp -r codex/gemini ~/.codex/skills/
cp -r codex/command-creator ~/.codex/skills/
cp -r skills/jira ~/.codex/skills/
cp -r prompts/humanizer ~/.codex/skills/Alternatively, symlink the entire repository:
ln -s $(pwd) ~/.codex/agent-skillsOpenClaw users add selected directories to the OpenClaw skill paths. Credentials, service-specific CLIs, and API setup are not documented centrally in the supplied overview, so inspect the files belonging to each chosen integration.
How do you use this agent?
Enable only the directories relevant to the intended task. The documented Codex configuration is:
[skills]
paths = [
"~/.codex/skills/codex",
"~/.codex/skills/gemini",
"~/.codex/skills/jira"
]The documented OpenClaw path configuration is:
skills:
paths:
- /path/to/agent-skills/skills/elegant-reports
- /path/to/agent-skills/prompts/frontend-design
- /path/to/agent-skills/clawdbot/todo-trackerOnce configured, ask the host assistant for a task matching the installed package, such as reading GA4 reporting data, searching Jira work, or producing a structured project plan. Initial invocation details and credential requirements belong to each package's SKILL.md; the repository overview does not define one universal run command.
What are this agent's strengths and limitations?
- It separates universal skills, instruction-only prompts, OpenClaw-specific packages, and Codex-specific packages into clear directories.
- The catalog contains concrete workflows spanning analytics, ticketing, CRM, PDF processing, and multi-agent orchestration rather than only generic prompt templates.
- A skill can ship its instructions together with scripts, templates, and assets, and the repository includes metadata-driven README generation and format validation scripts.
- The README states that every skill is scanned by AgentVerus for prompt injection, data exfiltration, and hidden threats, with results published from the latest scan.
- This is a library rather than a standalone application; adopters must select packages, wire them into an existing assistant, and maintain that installation.
- Portability is uneven: the README reports 79% cross-platform compatibility and separately identifies seven OpenClaw-only skills.
- Integrations with Jira, GA4, Salesforce, Zendesk, Gong, Google Ads, and Nutrient introduce service-specific accounts, networking, permissions, APIs, or CLI dependencies.
- The overview does not provide consistent credential names, runtime versions, first-call examples, or failure behavior for every integration, requiring package-by-package review.
task-orchestratorexplicitly relies on tmux/Codex execution, so that workflow should not be treated as portable to every supported assistant without caveats.
How does this agent compare with similar options?
Compared with the author's lobster-workflows repository, Agent Skills focuses on instructions and packages interpreted by an assistant. lobster-workflows is presented as the option for deterministic, repeatable typed pipelines with approval gates and state tracking. This repository is the better fit when users want assistant-directed skills and prompts; the companion project is positioned for workflows that require tighter execution control and repeatability.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Agent Skills Library This agent | 67 · Some gaps | Agent plugin / skillFree + model costs | ★ 240 | 1mo ago | Python | Codex · Claude Code |
| Agency Agents Chinese Expert Library | 28 · Major gaps | CLIFree + model costs | ★ 22k | 3d ago | Shell | Codex · Claude Code |
| Code Abyss: Persona and Engineering Skills for Coding Agents | 66 · Some gaps | CLI | ★ 244 | 3mo ago | JavaScript | Codex · Claude Code |
| WUPHF — A Collaborative Office for AI Employees with a Shared Brain | 66 · Some gaps | CLIFree + model costs | ★ 1.5k | 11d ago | Go | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
The release workflow uses read-only repository permissions, exact-commit binding, disabled persisted checkout credentials, pinned tools, a preview stage, and SHA-bound human publication gates. It creates the temporary token configuration with mode 0600 and removes it unconditionally. TODO tests also cover deletion confirmation, locking, atomic writes, backups, and an opt-in heartbeat. Deductions apply because these strong controls are evidenced mainly for the ten-package release process and one TODO utility, not for every API, scraper, updater, or CRM-writing skill in the collection. Repository-level documentation gives only limited detail about per-skill data destinations, retention, and external effects. Author, repository, paths, and commit provenance are explicit, but the publisher is unverified and the referenced AgentVerus scan results are not included as independent evidence.
The documented layout, platform positioning, and release batch are broadly consistent with the workflow. Release tests cover manifest drift, duplicate keys, paths, versions, gates, and secret printing, while TODO and Munger contract tests exercise substantial error and boundary behavior. The release chain pins Python, Node, PyYAML, and ClawHub versions. Deductions apply because no execution results are supplied, the elegant-reports workflow shows npm ci/npm test without the lockfile or test contents, and dependency availability, retry behavior, and user-facing failure messages are not demonstrated individually for most skills.
The source clearly separates universal skills, instruction-only prompts, OpenClaw-specific material, and Codex-specific material, with installation guidance for OpenClaw, Claude Code, Codex, and manual use. The Munger contract demonstrates precise activation, bounded history access, and scheduling boundaries, while the release workflow is restricted to a fixed ten-package set. Deductions apply because similarly precise boundaries are not evidenced for most skills, the underlying matrix for the 79% portability claim is absent, and repository-level documentation does not enumerate each skill's API, browser, CLI, credential, or operating-environment requirements.
The repository has clear skills/prompts/clawdbot/codex organization, structural examples, generated catalogs, and substantial multi-platform installation notes. The full MIT text agrees with the license metadata and merits full license credit. Versioned release commands and changelog summaries give ten packages stable identities, and the Clawdbot-to-OpenClaw rename is explained. Deductions apply for the lack of a complete FAQ, centralized known-limitations section, and repository-wide changelog. Examples are largely installation-oriented, and maintenance responsibility is implied by the author, contribution route, and release owner without a support policy, response commitment, deprecation cadence, or long-term maintenance statement.
The catalog addresses concrete workflows across analytics, documents, planning, CRM, advertising, reporting, and orchestration. Its classifications, short descriptions, and copyable installation commands make the collection practically approachable, while distinguishing structured skills from pure prompts aids selection. Deductions apply because the evidence contains few representative outputs, end-to-end usage cases, or comparative quality results for most skills. It also does not quantify time, token, API-quota, cloud-service, or maintenance costs; only isolated release notes mention quota or credit confirmation, so cost-benefit support remains thin.
Release paths, versions, owner, commits, and commands are traceable to the workflow. Separate test sources corroborate the release safety contract, TODO state behavior, and Munger evidence boundaries, including explicit separation of observation, inference, evidence, counterevidence, and confidence. Deductions apply because the supplied material omits the principal scripts, most SKILL.md files, the compatibility matrix, dependency lockfiles, and scanner reports, preventing repository-wide claim-by-claim verification. The 79% portability figure and claim that every skill is scanned therefore remain primarily project assertions.
- Do not treat the README's AgentVerus badge or claim that every skill is scanned as a verified security result in this assessment; the scan reports were not supplied.
- Before installation, inspect each selected SKILL.md, supporting script, dependency lockfile, credential scope, and network data destination, especially for auto-update, gallery scraping, CRM, advertising, and document-upload skills.
- The release controls cover a fixed ten-package ClawHub batch and should not be generalized to every other skill's confirmation, pinning, secret handling, or rollback behavior.
- Validate external writes and paid operations in read-only, preview, or sandbox modes first, and establish resource-specific confirmation and recovery plans for systems such as Jira, Salesforce, Zendesk, Google Ads, and Nudocs.
FAQ
Is the skill library itself paid?
Will every package run unchanged in Codex and Claude Code?
COMPATIBILITY.md and the selected package before adoption.Is an MCP server mandatory?
What sensitive systems can these packages access?
SKILL.md, scripts, and external calls before enabling it.Is there one shared recovery mechanism for failed runs?
task-orchestrator mentions self-healing heartbeat monitoring and context-recovery handles missing conversational context, but other integrations require individual inspection for retry and error behavior.