Claude Cowork Record a Skill — Launched Without Retention Policy

Anthropic shipped “Record a Skill” in Claude Cowork on July 21, 2026. You screen-record yourself doing a task, narrate what you are doing and why, and Claude converts your demonstration into a reusable SKILL.md workflow it can execute again. The feature captures screen activity, mouse clicks, keystrokes, and voice. It runs on Anthropic’s servers, not your local machine. And as of today, Anthropic has published zero documentation on what happens to those recordings after the skill is generated.

TL;DR

• What: Claude Cowork now learns workflows by watching your screen recording and listening to your narration
• Problem: Recordings capture everything visible — credentials, PII, internal URLs — and Anthropic has published no retention or deletion policy for them
• Comparison: OpenAI’s Codex Record & Replay shipped June 18 with explicit regional blocks (EEA, UK, Switzerland); Anthropic’s version has none
• Action: Do not use this feature on screens containing sensitive data until Anthropic publishes a recording-specific privacy policy

Record a Skill — What Happened

The UX is genuinely clever. Instead of writing instructions or building automation configs, you show Claude what you want done. Hit record, perform your workflow, talk through your reasoning as you go. Claude watches the screen capture, parses your narration for intent, and generates a SKILL.md file — a structured, inspectable, version-controllable document that describes the workflow in a format Claude can replay.

The generated SKILL.md is the good part of this story. It is plain text. You can read it, edit it, commit it to git, deploy it through CI/CD. On Team plans, skills can be shared across your organization, though sharing is off by default and must be explicitly enabled per skill. The output artifact is transparent and auditable.

The input artifact — the recording itself — is not.

When you record a skill, Cowork captures everything on your screen. Not just the application you are demonstrating. Everything visible. Your browser tabs. Your Slack sidebar. The dashboard behind the window you are clicking through. The terminal output scrolling in the background. Your password manager if it happens to be unlocked. Your client’s PII if a CRM is open on another monitor.

All of this travels to Anthropic’s servers. Cowork runs server-side, not locally. That is not speculation — it is confirmed architecture.

Cowork runs on Anthropic’s servers. Screen recordings are transmitted and processed remotely. If your screen shows credentials, client data, or internal tool URLs during recording, that data leaves your machine.

Why This Matters

I have been waiting for someone to ship this UX. Demonstration-to-automation is the right interface for non-technical skill capture. Writing prompt instructions is lossy — you forget steps, misjudge what is obvious versus what needs explicit specification, and end up debugging the instruction rather than the workflow. Showing what you do and narrating why is how humans actually teach each other. The SKILL.md output format is inspectable and composable. The narration captures intent, not just mechanics. This is good product design.

But Anthropic launched it with a gaping policy hole.

Multiple independent analyses published between July 21 and 24 confirm the same finding: Anthropic’s launch announcement contains no recording-duration limit, no file-size cap, no storage or retention policy, and no documentation on what administrative controls apply to recorded content. There is no dedicated technical page explaining how recordings are stored, how long they persist, or whether they are deleted after SKILL.md generation completes.

This is not a minor documentation oversight. This is a data governance vacuum around one of the most sensitive input types an AI product can accept. Screen recordings are not structured data you can sanitize. They are raw pixel streams containing whatever happens to be on screen. The surface area for unintentional data exposure is enormous — and unlike a text prompt, you cannot review a recording frame-by-frame before submitting it to check what you accidentally included.

The comparison to OpenAI’s equivalent feature makes the gap starker. OpenAI shipped Codex Record & Replay on June 18, 2026, as part of Codex app version 26.616. It is Mac-only. It is explicitly blocked in the EEA, the United Kingdom, and Switzerland — the three jurisdictions with the most stringent data protection enforcement. Whatever you think of that restriction, it signals that OpenAI’s legal team evaluated the regulatory exposure of screen-capture-to-automation and decided certain markets required explicit compliance work before launch.

Anthropic’s version has no stated regional restrictions. Record a Skill is available to Pro, Max, and Team users globally. Either Anthropic’s legal team reached a different conclusion about GDPR and UK GDPR exposure, or they did not do the analysis. Neither possibility is reassuring.

OpenAI’s Codex Record & Replay: Mac-only, blocked in EEA/UK/Switzerland. Anthropic’s Record a Skill: all platforms, no regional restrictions, no published retention policy. The gap is not subtle.

The team-sharing dimension adds another layer. On Team plans, a skill generated from a recording of a sensitive workflow can be shared organization-wide with a single toggle. The SKILL.md that gets shared is auditable — you can read it and verify what it does. But the recording that produced it is not versioned, not reviewable after the fact, and has no documented lifecycle. If the recording persists on Anthropic’s servers after skill generation, every team member who triggers a shared skill is implicitly trusting that the original recording did not capture something it should not have.

The training-data question is the sharpest edge. Anthropic’s general privacy documentation covers model-improvement consent for Claude consumer products. Whether recorded demonstrations fall within that scope is unanswered. If recordings feed into training pipelines — even as visual-understanding training data rather than conversational data — the exposure is categorically different from text-based interactions. A text prompt might accidentally contain an API key. A screen recording contains the entire visual context of a user’s working environment.

The Take

This feature is off-limits for anyone operating under SOC 2, HIPAA, or basic enterprise data hygiene requirements. Not because the feature is broken — the SKILL.md output format is genuinely well-designed, and demonstration-based skill capture is a better interface than instruction writing. It is off-limits because Anthropic shipped a screen-capture ingestion pipeline to their own servers without publishing what happens to the captured data.

The fix is straightforward: publish a recording-specific retention policy. State how long recordings persist. State whether they feed into training. State what happens when a user deletes a skill — does the source recording get deleted too? Provide admin controls for Team plans that let org admins disable recording-based skill creation or enforce review workflows before skills are shared. None of this is technically difficult. It is a documentation and policy deliverable.

Until that documentation exists in writing, treat Record a Skill as a personal productivity toy for non-sensitive workflows only. Do not record anything containing credentials, client data, internal infrastructure URLs, or regulated information. If you are on a Team plan, do not share skills generated from recordings until you can verify the recording lifecycle.

The UX is right. The policy is missing. Ship the policy, Anthropic.

If you want the demonstration-to-automation pattern without the data-governance risk: record your screen locally, write the SKILL.md manually from your own recording, and load it into Claude Skills without using the Record a Skill feature. You lose the automation but keep control of the recording.