That qualification matters in 2026 because Skill is now an overloaded term. OpenAI describes Skills as reusable instructions and resources that can encode multi-step workflows. Anthropic's Agent Skills can also combine instructions, scripts, tool guidance, and supporting resources for complex tasks.
In this guide, Shadow Skill means Shadow's current Action Skill or Meeting Skill behavior. It does not mean every product that uses the word Skill.
The short answer
An AI agent is a runtime that can plan and carry out multiple steps with models and tools inside configured permissions. Autonomy is a spectrum. Depending on the product and task, a person may interrupt, steer, approve, pause, or stop the work. OpenAI documents those controls for ChatGPT agent and describes sandbox and approval boundaries for Codex.
A Shadow Action Skill combines a saved prompt with the voice, screen, or selected-text inputs you enable. You trigger it from your Mac, then choose whether the result is pasted at the cursor or shown in an overlay. A Shadow Meeting Skill runs a reusable prompt over configured meeting context, on demand or automatically after meeting processing completes. Supported Meeting Skills can save Markdown or send a result to a configured webhook.
The comparison is therefore agent runtime versus scoped Shadow workflow, not agents versus a universal definition of Skills.
For a current product example at the persistent end of that spectrum, compare OpenAI Dots and Shadow by work horizon, context, permissions, and review points.
Why the old binary no longer works
Earlier product language often treated an agent as a loop and a Skill as one prompt. That can still describe a specific product design, including Shadow's current Action Skills, but it is no longer a safe category definition.
- An Agent Skill from OpenAI or Anthropic can contain multiple steps, scripts, or tool guidance.
- An agent can be tightly sandboxed and require approvals.
- A Shadow Skill can run a scoped prompt while still sending enabled context to an external AI provider.
- Apple Shortcuts, Custom GPTs, Claude Projects, and Raycast AI each mix configuration, context, tools, and human control differently.
What an AI agent does
An agent runtime can receive a goal, inspect available context, call tools, observe results, and continue across multiple steps. The product decides which tools and permissions are available. The user or administrator decides which boundaries to configure, and many systems insert approvals before consequential actions.
Current examples include:
- ChatGPT agent using a browser and connected tools while allowing the user to interrupt or steer the task.
- Claude's agent tooling exposing files, commands, and MCP-connected tools to an agent loop.
- Cursor Agent editing files and using development tools.
- Codex working through code changes inside sandbox and approval boundaries.
What a Shadow Action Skill does
A Shadow Action Skill is a scoped Mac workflow:
1. Open Quick Access or use the Skill's configured shortcut. 2. Provide the inputs the Skill is allowed to use: voice, screen, selected text, or a combination you enabled. 3. Run the saved prompt. 4. Paste the result at the cursor or show it in an overlay. 5. Review the result before using it in an important workflow.
That shape is useful when the work begins with what is already in front of you. Draft a reply from the current message and a spoken instruction. Turn selected text into a concise summary. Rewrite a paragraph without moving it into a separate chat window.
The result behavior is configured up front, but the model output still needs review. Paste at cursor requires Accessibility access. Action Skills are unavailable while Listening is active for a meeting.
What a Shadow Meeting Skill does
A Shadow Meeting Skill is a reusable prompt for meeting context. It can create notes, action items, a follow-up draft, or another structured result.
You can run a Meeting Skill from a processed meeting. If a supported Skill has automation enabled, it runs after meeting processing completes, not when the call starts.
The surrounding meeting workflow has separate stages:
1. Autopilot can start Listening when Shadow detects a supported meeting, or you can start Listening manually. 2. Core audio capture, transcription, diarization, and vault storage stay local. 3. When Smart Screenshots are enabled and a capture target remains visible, Shadow saves images when meeting content changes meaningfully. Screenshots pause when the selected target is hidden. 4. After processing, a Meeting Skill can run with its configured context. 5. Supported results can be saved as Markdown or sent to a configured webhook. 6. A person reviews generated claims, decisions, owners, and dates before sharing them.
Meeting detection, screenshot capture, AI processing, and result delivery are distinct controls. Treating them as one automatic pipeline hides the privacy and review decisions that matter.
Agent versus Shadow Skill
| Decision | AI agent runtime | Shadow Action or Meeting Skill |
|---|---|---|
| Work shape | Can plan and execute multiple steps with tools | Runs a configured Shadow prompt workflow |
| Start | User goal, scheduled run, event, or another configured trigger | Action Skill shortcut or Meeting Skill control/automation |
| Context | May gather more context through allowed tools | Uses the Shadow inputs and meeting context you enable |
| Control | Can include steering, approvals, sandboxing, pause, and stop controls | User chooses inputs and result behavior, then reviews the output |
| Result | Depends on the agent's allowed tools and task | Action result goes to cursor or overlay; supported Meeting results can go to Markdown or webhook |
| Best fit | Longer work that benefits from planning, tool use, and iteration | Repeated Mac or meeting work with a known prompt and review point |
Neither column is automatically safer or more capable. The useful questions are: What context can leave the Mac? Which tools can act? Where is approval required? What happens when the model is wrong?
For an emerging event-started pattern, MCP Events and permissioned trigger contracts explains how a connected signal can begin bounded AI work without becoming a new authority grant.
For the operating-system boundary, the Mac AI permission map separates task-specific access, workflow persistence, product action scope, and exceptional Full Disk Access.
Privacy and external processing
Shadow's core meeting capture and transcription stay local. Audio, transcript processing, diarization, and finalized meeting files remain in the local vault by default.
Running Action Skills, Meeting Skills, or Ask can send enabled context to an external AI provider. Depending on the feature and settings, that context can include transcripts, screenshots, notes, selected text, prompts, Skill options, profile or calendar context, and generated results. A webhook sends the result to the external destination you configured.
For sensitive work, review each Skill's inputs, disable automation when a standing external-processing rule is inappropriate, and test with non-sensitive content first. The current Privacy and Data guide explains how to minimize external processing.
Three practical choices
Choose an agent when the work needs a tool loop
Coding is the clearest example. An agent can edit a file, run a test, inspect the failure, and revise the change. Research modes in ChatGPT and Claude can conduct multiple searches, analyze sources, and synthesize a cited report. The iteration is the value.
Use an agent when the goal is bounded, the permissions are appropriate, and the task benefits from multiple tool calls. Keep approval gates around consequential external actions.
Choose a Shadow Action Skill for in-context Mac work
Use an Action Skill when the job begins with the current screen, selected text, or a spoken instruction and should return to a known review surface. Drafting a reply, rewriting selected text, or cleaning up dictated text fits this lane.
Choose a Shadow Meeting Skill for repeatable post-meeting output
Use a Meeting Skill when you want the same reviewed structure after a meeting, such as decisions, action items, or a follow-up draft. Automation can remove the extra click after processing, but it also creates a standing rule for external AI processing. Enable it only when that rule fits the meetings in scope.
How to choose for a specific task
Ask five questions:
1. Does the work need multiple tool calls or iterative planning? If yes, an agent runtime may fit. 2. Is the input already on the current Mac screen, in selected text, or in a short voice instruction? An Action Skill may be the smaller tool. 3. Is the input a completed meeting with a repeatable output format? A Meeting Skill may fit. 4. What leaves the Mac? Check every enabled input, external AI request, and webhook destination. 5. Where does a human review or approve the result? If that answer is unclear, do not automate the consequential step.
The narrowest workflow that reliably completes the job is usually easier to review, measure, and trust.
FAQ
Is an AI Skill always smaller than an AI agent?
No. OpenAI and Anthropic use Agent Skills for reusable instruction and resource packages that may include multi-step workflows, scripts, and tool use. Shadow uses Skill as a product term for Action and Meeting workflows. Always check the product-specific behavior.
Is a Shadow Action Skill the same as a Custom GPT or Claude Project?
No. A Shadow Action Skill is a Mac-triggered product workflow with selected context and configured result handling. Custom GPTs can include knowledge, capabilities, apps, and actions. Claude Projects can include instructions, knowledge, files, memory, and agentic work. They are broader configurable AI surfaces.
Does Shadow have a general-purpose agent mode?
Shadow's current public documentation describes Action Skills and Meeting Skills, not a general-purpose autonomous agent mode. That is a statement about the current product, not a roadmap promise.
Where does Apple Intelligence fit?
Apple Intelligence no longer fits cleanly into a Skill-versus-agent binary. Apple's announced Siri AI includes screen context, personal-context search, web access, and systemwide app actions. Shortcuts also supports configurable model-powered workflows. Apple's Siri AI announcement and Shortcuts guide are the current first-party references.
Where does Raycast AI fit?
Raycast Quick AI supports typed prompts and Screen Awareness for the focused window, selected text, a selected area, or the full screen. Raycast AI Chat adds persistent history, attachments, and tools. Those surfaces do not reduce to one side of the binary either.
What does Shadow cost?
Shadow's free tier includes unlimited recording and transcription, audio recording, and Smart Screenshots. Plus adds the unlimited AI features listed on the current pricing page and costs $8 per month billed annually or $12 month-to-month.
Bottom line
"Agent" and "Skill" are labels, not sufficient specifications.
An agent runtime can carry multi-step work through models and tools inside configured permissions. A Shadow Action or Meeting Skill runs a scoped product workflow with user-selected context, configured result handling, and a clear review point.
Choose by execution shape, permissions, data flow, and approval boundaries. If a one-run Mac or post-meeting workflow does the job, use the smaller surface. If the work genuinely needs planning, tools, and iteration, use an agent with explicit limits.
Shadow is an AI interface for Mac that sees, hears, and runs. Its current public product model centers on Action Skills and Meeting Skills, with core meeting capture and transcription local by default and external AI processing controlled by the features and inputs you enable.
---
This article was written by Chad Oh, Shadow's AI writer. While we strive for accuracy, AI-generated content may contain errors. If you spot something off, let us know.