How to record a skill, and which recorder to use
You can now show an AI a task instead of describing it. Claude's Record a skill works in Cowork in Claude for Mac, OpenAI's Record & Replay in the ChatGPT desktop app and Codex, and Microsoft's open-source Skill Recorder builds skills for Microsoft's agents. Ditto AI records a task on your Mac and turns it into an agent with a checked tool for each step, installed into Claude Code or Kiro. Pick the one that lives where your agent runs.
The short answer
Recording a skill means doing a task once while a tool watches, then letting a model write the procedure an agent can follow next time. Four tools do this today, and the biggest difference between them is where the result runs.
| Recorder | Where it runs | What it makes | Where the result runs |
|---|---|---|---|
| Claude's Record a skill | Cowork in Claude for Mac, on Pro, Max and Team plans (Anthropic) | A proposed skill you save in Customize > Skills | Claude |
| OpenAI's Record & Replay | The ChatGPT desktop app (Work) or Codex, on macOS, with Computer Use enabled (OpenAI) | A skill: when to use it, its inputs, steps and how to verify | ChatGPT or Codex |
| Microsoft's Skill Recorder | A desktop app, open source under MIT (GitHub) | A SKILL.md procedure or a scheduled automation, built with the GitHub Copilot CLI | Microsoft Scout, Copilot Cowork or Copilot Studio |
| Ditto AI | A Mac app, macOS 15 or later on Apple silicon | An agent: steps with evidence, inputs, a checked tool per step and a permission policy | Ditto, Claude Code or Kiro CLI |
How recording a skill works, whichever tool you use
- Pick a task you repeat, with the same shape every time and details that change. Two or three apps is the sweet spot.
- Close what shouldn't be seen. Anthropic's advice holds for every recorder: "Don't type passwords or secrets, or display sensitive information or private conversations while recording" (Anthropic).
- Do the task once, at a normal pace, and talk through it. Saying which values change next time helps the model tell an input from a constant.
- Read what it wrote before you use it. Every recorder here shows you the result first; this is where a wrong step gets fixed cheaply.
- Run it on a new input and compare the result with what you'd have done by hand.
Claude's Record a skill
"Recording a skill is available on Pro, Max, and Team plans, in Cowork in Claude for Mac. It isn't available in chat, on Windows, or on Free and Enterprise plans." You send "a video of your screen, clicks, typing, and voice", for "about 10 minutes" at most, and Claude proposes a skill, or changes to a skill you already have (Anthropic). "The video and audio from your recording aren't retained"; screenshots stay in the Cowork task. The best fit if your work already happens in Cowork. Compared with Ditto AI.
OpenAI's Record & Replay
"Record & Replay lets you demonstrate a workflow on your Mac and turn it into a reusable skill." You start it from Plugins, then Record a skill, in the ChatGPT desktop app or Codex. Replay works through "Computer Use, browser actions, and installed plugins" (OpenAI). The best fit if you work in ChatGPT or Codex. Compared with Ditto AI.
Microsoft's Skill Recorder
An open-source desktop app that "records your on-screen work session and uses the GitHub Copilot CLI to reconstruct it as an intent + ordered steps, then builds a reusable Skill or Automation for Microsoft Scout, Microsoft Copilot Cowork, or Copilot Studio". Its README says both outputs prefer the agent's native tools over replaying clicks (GitHub). The best fit for teams on Microsoft's agents, or anyone who wants to read the recorder's source.
Ditto AI
Ditto records a task across your Mac apps and writes down the steps, each citing the frames and clicks it came from. Then it does the part the others leave to the agent: it finds a real integration for every step (an MCP server from the official registry or npm, or a headless browser for a public page), checks each one with a single read-only call, and sets permissions from a fixed table, where anything unclassifiable counts as a write. The agent runs in Ditto or installs into Claude Code or Kiro with its servers and allowlist. It doesn't export to Codex yet.
Which one to use
- You work in Cowork: Claude's Record a skill.
- You work in ChatGPT or Codex: Record & Replay.
- Your agents are Microsoft's: Skill Recorder.
- You work in Claude Code or Kiro, or want each step on a checked tool with a fixed policy: Ditto AI.
Honest limits, for all of them
- A recording is one example. Branches you didn't show, like an invoice in another currency, aren't in the skill until you add them.
- Recordings capture your screen. Whatever was visible may be sent to the model that reads it, so record with care.
- Check the first runs. Keep writes as drafts, or run on test data, until the skill has done the task right a few times.
To try Ditto's version, install Ditto AI, or read how it works.