Claude's "Record a Skill" and ChatGPT Codex's "Record & Replay": The Future of Business Automation

AI Learns Work by Watching Your Screen

As of July 22, 2026, Anthropic's "Record a skill" added to Claude Cowork is a reality. According to Anthropic's official post, users can record their screen while working and explain what they are doing; Claude then converts that content into a reusable skill. This is available via the "+" menu in the Claude desktop app for Pro, Max, and Team plans.

First, Anthropic specifies "recording the screen and talking while working." While clicks and inputs are visible, it's unclear if it saves raw event data like a keylogger or analyzes video and audio. Describing it as "letting Claude analyze clicks and typing" is generally accurate for users but shouldn't be stated as a technical certainty.

Second, "no additional charge" means it's included in existing subscriptions. While Claude Cowork is part of Pro, Max, and Team plans, Anthropic notes that Cowork consumes usage limits faster than regular chat. Thus, it is not "completely free" or "unlimited."

Third, the competition with OpenAI is not just a future prediction. On June 18, 2026, OpenAI announced "Record & Replay" as a feature for Codex on macOS with a nearly identical concept. Users demonstrate a workflow once, and it converts into a reusable skill. OpenAI's public announcement preceded Anthropic's by about a month.

In short: Anthropic added "Record a skill" to Claude Cowork, while OpenAI's ChatGPT Work/Codex already has "Record & Replay." The competition has shifted to "showing the AI how to work instead of writing prompts."

1. This is Not a "Macro," but a System for Learning the Meaning of Work

At first glance, Record a skill looks like old-fashioned macro recorders or RPA. Traditional macros simply replay actions like "click here" or "wait 3 seconds," which break if the UI changes.

In contrast, Claude and Codex skills are "AI capability packages" that include:

• When to use the skill
• Required inputs
• Step-by-step procedures and decision logic
• Prohibitions and verification methods
• Necessary scripts, templates, and tools

OpenAI explains that Codex-generated skills include usage context, input requirements, and result verification. For re-execution, you only need to provide the new values (files, dates, etc.). Codex then executes the workflow using Computer Use, browser operations, and plugins.

2. Differences Between Claude and Codex

While similar, there are key differences:

• Claude Cowork: Uses "Record a skill" via the "+" menu. Focuses on knowledge work like research and document creation. Available for Pro, Max, and Team plans.
• ChatGPT Work/Codex: Uses "Record & Replay" via Plugins. Focuses on general business and software development. Initially limited to macOS and specific regions (excluding EEA/UK/Swiss).

3. Why "Not Writing Prompts" Matters

Previously, automation was limited to those who could articulate workflows in text. Now, experts can externalize their "tacit knowledge"—the intuitive judgments made during a task—simply by speaking while doing. Record a skill lowers the cost of verbalizing one's work, allowing the AI to create a draft that humans then refine.

4. Areas with the Greatest Impact

Automation works best for high-frequency, stable tasks with clear completion criteria:

• Sales: From research to CRM updates and follow-up drafts.
• Accounting: Downloading reports, reconciling data, and budget variance analysis.
• Marketing: Standardizing publishing procedures (titles, thumbnails, UTM parameters).
• Support: Categorizing inquiries and drafting responses.
• HR: Cross-system onboarding tasks.
• Data Analysis: Automating the path from data retrieval to insight summary.
• Dev/QA: Reproducing bugs, UI diffing, and release note generation.

5. How to Record a Good Skill

1. One task per skill: Avoid overly broad recordings.
2. State purpose and rules first: Tell the AI the goal and prohibitions before starting.
3. Explain "Why," not just "What": Voice the reasoning behind your clicks.
4. Include exceptions: Don't hide errors; show the AI how to handle them.
5. Use realistic test data: Avoid sensitive info but keep the complexity real.
6. Edit the result: Treat the AI's generated skill as a first draft.
7. Test failures: See how the skill handles missing data or UI changes.
8. Automate last: Start with manual triggers before moving to scheduled execution.

6. Recording Template

Start your recording by stating: "The purpose of this skill is [Goal]. It should be triggered when [Condition]. Inputs are [Data]. Rules include [Logic]. Never [Prohibition]. It is complete when [Criteria]."

7. Risks in Corporate Implementation

• Confidentiality: Screen recordings might capture notifications or private data.
• Standardizing Bad Habits: Inefficient workflows can be codified as "correct."
• UI Changes: AI might fail if system layouts change significantly.
• Prompt Injection: Skills reading external data could be manipulated by malicious instructions.
• Skill Bloat: Organizations need governance to manage redundant or outdated skills.

8. Claude or Codex: Which to Use?

• Claude Cowork: Best for non-technical departments (Sales, HR, Marketing) handling documents and research across multiple tools.
• ChatGPT Work/Codex: Best for development-integrated workflows, utilizing CLI, IDE, and scripts.

9. 30-Day Implementation Roadmap

• Week 1: Identify candidate tasks based on frequency and risk.
• Week 2: Record three pilot skills (Data, Document, and Multi-app types).
• Week 3: Test, improve, and measure time savings vs. error rates.
• Week 4: Share validated skills with the team and move to limited automation.

10. The Essence of Competition

The real value for companies isn't the recording file itself, but the "organizational knowledge"—how the company makes decisions and completes work. Record a skill and Record & Replay are the gateways to turning individual expertise into verifiable, updatable organizational capabilities. The biggest beneficiaries are not just engineers, but the domain experts who know the work best.