Ten open-source utility skills that give AI coding agents real-world tools

The problem

The problem: agents without utility tools

An AI coding agent without utility skills is a developer who can write code but cannot:

• Create a diagram without leaving the chat
• Audit a website's SEO without paying for Semrush or Ahrefs
• Sign a PDF without opening Preview or Adobe
• Transcribe a video without a separate transcription service
• Schedule a meeting without switching to Google Calendar
• Access a remote machine without a separate SSH client
• Build a presentation without PowerPoint or Keynote
• Clean up automation without manually deleting workflows
• Reference a design system without opening Figma

These are not model capability gaps. They are tool gaps. The agent can describe how to create a Figma diagram, but it cannot create one without a tool that talks to the Figma API. Skills provide those tools.

The skills

The ten skills

Design and visualization (3)

figma-mcp-free — Create Figma diagrams on the free Starter plan. Architecture diagrams, microservice maps, user flows, UI mockups — all through the Figma Plugin API without a paid Figma plan. The skill handles the Plugin API authentication, file creation, and node placement. Use when an agent needs to produce a visual artifact in Figma without leaving the chat.

figma-diagram-visualization — A scientific algorithm for building diagrams, dashboards, and visualizations in Figma. The skill encodes a repeatable process: identify the diagram type, determine the node set, compute the layout, place the nodes, route the edges. It is not a prompt — it is a procedure the agent follows step by step. Composes with figma-mcp-free for the actual Figma API calls.

pantheon-style — Design system tokens reference. WCAG accessibility tables, anti-patterns for motion and spacing, token naming conventions, and component guidelines. This is a reference skill — the agent reads it to answer questions about the design system, not to execute code. It replaces the "open Figma and look it up" workflow with an in-chat reference.

SEO and web (1)

free-seo-stack — Free alternatives to Semrush, Ahrefs, and Ubersuggest. Lighthouse for performance and accessibility, CrUX for real-user data, Google Search Console for indexing and rank tracking, Bing Webmaster Tools and IndexNow for fast indexing, backlink analysis through public APIs, rank audit, and security checks. The skill orchestrates these free tools into a single workflow that replaces a paid SEO tool subscription.

The key insight: most SEO data is available for free through public APIs. The paid tools aggregate and present it, but the underlying data — Lighthouse scores, CrUX metrics, GSC coverage, backlink indexes — is free. The skill teaches the agent to call these APIs directly and assemble the results into an audit report.

Document production (2)

pdf-signing — Add a stamp, seal, or signature to a PDF. Secure local signing assets, visual placement verification. The skill handles the PDF manipulation locally — no document leaves the machine. Use when an agent needs to sign or stamp a contract, certificate, or official document.

talk-deck-builder — Build a PDF talk deck for presentations. Brand-style templates, consistent typography, and automatic layout. The skill generates a PDF from a Markdown outline — the agent writes the content, the skill handles the design. No PowerPoint or Keynote needed.

Media and communication (2)

youtube-transcribe — Transcribe YouTube videos with AI. The pipeline: yt-dlp downloads subtitles or audio, faster-whisper large-v3 transcribes locally, and Groq API provides a fast fallback when local transcription is too slow. The skill handles the full pipeline — download, transcribe, and return the text. Use when an agent needs to extract the content of a video for analysis, summarization, or repurposing.

zoom-host — Create Zoom meetings on a shared Pro host. The meeting gets a public join_url that anyone can use, and the facilitator fetches the start_url via a Credential Gateway — the host credential never leaves the gateway. Use when an agent needs to schedule a meeting without a human opening Zoom.

Infrastructure (2)

remote-mac-access — Connect to a remote Mac through a reverse SSH tunnel via a VPN relay. Works behind NAT, firewall, or TSPU. The skill sets up the tunnel, handles the authentication, and provides the SSH command. Use when an agent needs to access a team member's Mac for diagnostics, installation, or configuration — and the Mac is behind a firewall.

n8n-temp-cleanup — Discover and delete temporary or one-shot n8n workflows. n8n workflows accumulate — every one-shot automation, every test workflow, every "let me try this" leaves a workflow behind. The skill discovers all workflows, identifies the temp/one-shot ones, and deletes them. Use when an agent needs to clean up an n8n instance that has grown cluttered with temporary workflows.

Why utility skills

Why utility skills matter

Engineering skills — the ones in github.com/g-shevchenko/agentic-engineering-skills — teach the agent how to work. Utility skills teach the agent what to work with. Both are necessary. An agent that knows how to run a proof loop but cannot create a Figma diagram is a developer who can write code but cannot communicate the architecture. An agent that can create diagrams but cannot run a proof loop is a designer who cannot ship.

The composition is intentional. The github.com/g-shevchenko/agentic-engineering-skills repository provides the engineering patterns. The github.com/g-shevchenko/utility-skills repository provides the utility tools. The github.com/g-shevchenko/agentic-quality-skills repository provides the quality gates. The github.com/g-shevchenko/mcp-token-savers repository provides the MCP servers. Together, they form a complete agent stack — engineering, utility, quality, and infrastructure.

Key pattern

The free-seo-stack pattern

The most economically significant skill in this collection is free-seo-stack. SEO tools — Semrush, Ahrefs, Ubersuggest — cost $100-300/month per seat. For a team of 10, that is $1,000-3,000/month. The underlying data is mostly free:

• Lighthouse — Google's open-source tool for performance, accessibility, and SEO auditing. Free, runs locally or via API.
• CrUX — Chrome User Experience Report. Real-user performance data from millions of Chrome users. Free via API.
• Google Search Console — Indexing status, search queries, click-through rates, and coverage reports. Free.
• Bing Webmaster Tools — Indexing, backlinks, and search analytics. Free.
• IndexNow — Instant indexing for Bing and Yandex. Free.
• Backlink data — Available through public APIs (Google, Bing, and third-party indexes).

The paid tools aggregate this data and present it in a unified dashboard. The skill teaches the agent to call these APIs directly and assemble the results into a report. The agent becomes the dashboard.

Key pattern

The youtube-transcribe pipeline

The youtube-transcribe skill is a three-stage pipeline with a fallback:

1. yt-dlp — Download subtitles if available. If the video has human-written captions, the pipeline stops here. This is the fastest path.
2. faster-whisper large-v3 — If no subtitles, download the audio and transcribe locally. This is the high-quality path — large-v3 is the most accurate Whisper model.
3. Groq API — If local transcription is too slow (long video, slow machine), fall back to Groq's hosted Whisper. Groq's hardware is optimized for inference — transcription that takes 10 minutes locally takes seconds on Groq.

The fallback is the key design pattern. The skill does not fail when local transcription is slow — it falls back to a faster path. And the Groq API key is optional — the skill works without it, just slower.

Key pattern

The remote-mac-access tunnel

The remote-mac-access skill solves a problem that sounds simple but is not: connecting to a Mac that is behind a firewall. The Mac is behind NAT. The Mac is behind a firewall. The Mac is behind TSPU (Russia's DPI system, which blocks many VPN protocols). The solution is a reverse SSH tunnel through a VPN relay:

1. The Mac establishes an SSH connection to a public relay server.
2. The relay server opens a port that forwards to the Mac's SSH port.
3. The agent connects to the relay server's port, which tunnels to the Mac.

This works because the Mac initiates the connection — it does not need a public IP or open port. The relay server can be any VPS with a public IP. The VPN handles the network obfuscation — the relay connection looks like normal HTTPS traffic to a firewall.

The skill handles the tunnel setup, authentication, and SSH command generation. The agent just says "connect to the Mac" and gets an SSH session.

Key pattern

The zoom-host credential pattern

The zoom-host skill solves a credential isolation problem. Zoom meetings need a start_url — a URL that lets the host start the meeting. The start_url contains the host's OAuth token. If the agent has the start_url, the agent has the host's Zoom access.

The skill solves this with a Credential Gateway pattern:

1. The agent calls the skill to create a meeting.
2. The skill calls the Zoom API with the host's credential — but the credential is stored in a gateway, not in the agent's context.
3. The skill returns only the join_url (public, anyone can use it) to the agent.
4. The facilitator — a human or a separate process — fetches the start_url from the gateway when it's time to start the meeting.

The agent never sees the host credential. The agent never sees the start_url. The agent can create meetings but cannot start them. This is capability separation — the agent has the "create meeting" capability but not the "start meeting" capability.

Key pattern

The n8n-temp-cleanup lifecycle

n8n workflows accumulate. Every one-shot automation, every test workflow, every "let me try this" leaves a workflow behind. Over months, an n8n instance grows from 20 workflows to 200, most of which are dead. The n8n-temp-cleanup skill solves this:

1. Discover — List all workflows in the n8n instance.
2. Identify — Flag workflows that are temp or one-shot. The skill uses heuristics: workflows with "test" or "temp" in the name, workflows that have not been executed in 30 days, workflows with no trigger node (manual-only).
3. Delete — Remove the flagged workflows. The skill asks for confirmation before deleting — this is a destructive operation.

This is a lifecycle skill — it runs periodically to keep the n8n instance clean. It composes with the MCP stack token economy measurement framework to ensure that the remaining workflows are valid.

Install

Installation and verification

Install

git clone https://github.com/g-shevchenko/utility-skills.git
cd utility-skills
bash scripts/install.sh

Then in your agent chat:

use utility skills stack

The installer copies skills to the target directory (default: $HOME/.codex/skills), optionally writes a managed block to AGENTS.md, refuses unsafe targets, and creates a backup before replacing an existing skill.

Verify before install

Clone and inspect first:

git clone https://github.com/g-shevchenko/utility-skills.git
cd utility-skills
bash scripts/doctor.sh
bash scripts/audit-public-surface.sh

• doctor.sh — verifies all skills have required files and frontmatter
• audit-public-surface.sh — scans for secrets, private paths, and placeholder markers

See github.com/g-shevchenko/utility-skills/blob/main/VERIFY_BEFORE_INSTALL.md and github.com/g-shevchenko/utility-skills/blob/main/SECURITY.md for details.

Compatibility

Dependencies

External dependencies

Some skills reference external services or tools. All endpoints are configurable — no hardcoded internal URLs.

Boundaries

Why public

These skills were extracted from production operating patterns. The internal versions had dependencies on private infrastructure — internal API endpoints, private credentials, hardcoded URLs. The public versions have none of that. Every endpoint is configurable. Every credential is optional. Every dependency is a public tool or a free API.

The publication boundary is strict: the skills teach the agent how to use a tool, not which tool to use. The internal version says "use the HWAI Figma account." The public version says "use your Figma account." The internal version says "connect to the HWAI Mac Mini." The public version says "connect to your Mac." The pattern is the same. The credential is different.

This is the same boundary used in the github.com/g-shevchenko/agentic-engineering-skills and github.com/g-shevchenko/mcp-token-savers releases: publish the pattern, keep the infrastructure private.

AEO findings

AEO findings

AEO Magic collection run 259 was executed on 2026-09-09 with 85 queries (35 utility-skills + 50 from previous articles) against ChatGPT UI, collected from Dubai, UAE. All 85 queries were measured and answered, with zero failures.

ChatGPT produced substantive answers about Figma diagrams, SEO tools, PDF signing, YouTube transcription, Zoom hosting, remote Mac access, n8n cleanup, and design system references — but did not mention gregshevchenko.com, did not cite any gregshevchenko.com page, and did not mention utility-skills by name. Zero brand mentions, zero citations, zero top domains returned.

This is consistent with the article-1 and article-2 findings: the project is new and not yet established as a citable source. ChatGPT recognizes the tools and workflows described in the skills — Figma Plugin API, Lighthouse, CrUX, yt-dlp, faster-whisper, Zoom API, reverse SSH tunnels, n8n workflow management — but does not attribute these workflows to any specific open-source package. The article itself serves as the primary source for the project.

References

References

• github.com/g-shevchenko/utility-skills
• github.com/g-shevchenko/agentic-engineering-skills
• github.com/g-shevchenko/agentic-quality-skills
• github.com/g-shevchenko/mcp-token-savers
• Agentic Engineering Skills — open-source patterns for AI agents
• MCP Token Router — measurement-driven compressor routing
• Part 1 — How I cut my Claude Code token usage by 75.5% with 17 local MCPs
• Part 2 — Receipts on real production
• Part 3 — When MCPs save tokens (N=100)
• Part 4 — Measuring a dead-code detector honestly