Anthropic shipped a feature yesterday that removes the hardest part of building Claude automations: you no longer need to write a single line of SKILL.md to teach Claude a workflow. Record a Skill, available now in Claude Cowork on Pro, Max, and Team plans, turns a screen recording with voice narration into a reusable, runnable skill. Show Claude once. It handles the rest.
How Record a Skill Works
The entry point is the + menu inside the Claude Desktop app. Select Record a skill, then perform whatever workflow you want to automate â while narrating your reasoning aloud. Claude captures screen activity, mouse clicks, keystrokes, and voice commentary in parallel. When you stop the recording, it synthesizes that multimodal evidence into a structured skill your library can run on demand.
The voice requirement is not a cosmetic detail. Codex Record & Replay â OpenAI’s functionally equivalent feature, which shipped on June 17, 34 days earlier â captures clicks and keystrokes but ignores voice entirely. That matters: a skill built from mechanics can replay a sequence. A skill built from explained intent can adapt when inputs change. Claude’s approach is betting on the latter.
What You Actually Get: The SKILL.md File
The output is not a black box. Every skill Claude generates is a folder containing a human-readable SKILL.md file â plain markdown with YAML frontmatter at the top. A generated file looks roughly like this:
---
name: Weekly metrics report
description: Pulls dashboard data and formats the Tuesday standup report
---
## Steps
1. Open the metrics dashboard
2. Export last 7 days as CSV
3. Format top 3 KPIs into the standup templateBecause SKILL.md is plain text, the generated output is auditable and editable. You can review it, modify steps, drop it into version control, and submit pull requests when the underlying workflow changes. Optional subdirectories â templates/, references/, scripts/ â let more complex skills carry their own supporting files. This is not a proprietary recording format: it is standard markdown that any developer can read, fork, and improve.
The Competitive Picture
OpenAI beat Anthropic to this feature by over a month, and that context is worth knowing. Codex Record & Replay is macOS-only, excludes users in the EEA, UK, and Switzerland, and generates skills from visual actions alone. Claude’s implementation runs on macOS, Windows, and Linux, has no reported geographic restrictions, and incorporates voice narration to capture workflow intent.
The substantive difference is not the 34-day gap â it is the voice layer. If your automation involves conditional logic (“I add a note here because the client name does not match the invoice format”), a click replay will silently skip that judgment. A voice-narrated skill that understood why the step exists has a chance of handling the edge case. Whether that holds up in practice is something each team will need to validate, but the design philosophy is coherent.
Team Plan: One Demo, Org-Wide Automation
On Team plans, skills can be shared across an organization. Sharing is off by default and must be enabled per skill â a reasonable default that prevents accidental broad exposure of sensitive workflows. The implication for engineering teams is significant: a senior developer can record an internal runbook once and distribute a runnable version to the entire team. This is living documentation that also executes.
SKILL.md files are packageable as ZIP archives and deployable across Cowork instances, which means they can move through standard CI/CD pipelines. A skills library managed in git, tested in staging, and released to production is a viable architecture. Anthropic’s skill authoring best practices cover the underlying format in detail.
Before You Hit Record: The Privacy Checklist
Screen recordings capture everything visible on screen â not just the workflow steps Claude needs. That means email threads, internal dashboards, credentials in browser tabs, and colleague names are all fair game. Anthropic has not yet published a dedicated retention and training-data policy for Record a Skill content, and until that policy exists, the safest posture is to treat every recording as if it could be reviewed.
- Close email, Slack, and any password manager windows before starting
- Disable desktop notifications during the recording session
- Use synthetic test data â fake names, masked account numbers
- Remove or mask API keys and tokens from any visible windows
- Review the generated SKILL.md for any sensitive identifiers Claude copied
- Do not record workflows that surface HIPAA-covered data, financial account numbers, or credentials
The convenience is real; so is the exposure. Until Anthropic clarifies what happens to recording content under its data retention and model training policies, these hygiene steps are not optional for teams with compliance obligations.
The Bottom Line
Record a Skill is a genuine workflow shift for teams already invested in Claude Cowork. It does not add new capabilities to Claude â skills have existed for months â but it removes the documentation burden that was blocking adoption. The generated SKILL.md is auditable and version-controllable, the voice narration captures intent rather than just mechanics, and the team sharing architecture is thoughtfully conservative.
The feature is available now from the + menu in Claude Desktop for Pro, Max, and Team subscribers. Try it on a low-stakes repeatable workflow first, review the output before running it anywhere critical, and address the privacy checklist before you record anything that touches real user data.
