Claude Record a Skill: Turn Screen Recordings Into Reusable Workflows

Direct Answer

Claude Cowork's Record a Skill feature lets you demonstrate a task instead of writing its operating procedure from scratch. Start a recording in the Claude desktop app, perform the workflow, explain what you are doing and why, then let Claude turn the demonstration into a reusable Skill.

The important correction is that Claude is not permanently retraining its base model on you. It is creating an inspectable package of instructions, examples, scripts, and resources that Claude can load on future runs. That can feel like teaching a colleague, but the durable asset is the Skill, not a private set of newly trained model weights.

Video credit: Miles Deutscher, "This NEW Claude Feature Is The Best Thing I've Used In 3 Years". Follow @milesdeutscher on X. Miles demonstrates client onboarding and YouTube outlier research, then shares further sales, finance, content, ecommerce, and operations ideas. His experience is a creator test; this article checks the product framing against Claude's official launch post, Skills guidance, Cowork safety documentation, and privacy information.

Source Note

Claude announced Record a Skill on 21 July 2026: record a task, narrate it, and Claude turns it into a Skill that can run again. The launch post lists Pro, Max, and Team plans and places the feature under the plus menu in Cowork inside the desktop app.

Anthropic had not published a dedicated Record a Skill help page when this article was checked on 23 July. The mechanics in this guide therefore come from the official launch post and Miles' live demonstration. The operating boundaries come from Anthropic's current documentation for Skills, Cowork computer use, and consumer data controls. Feature availability, limits, and policy can change during rollout.

Link Map

Claim Check: What Does "Claude Learns by Watching" Mean?

What Record a Skill Actually Creates

A Skill is closer to a living operating procedure than a rigid macro. Anthropic describes Skills as folders containing instructions, scripts, and resources. Claude first sees the Skill's name and description, then loads the full instructions and supporting files only when the task matches.

Record a Skill changes how the first draft is authored. Instead of manually writing every step, you provide a multimodal example:

• The screen shows the sequence: which app, folder, page, field, or document you use.
• Your narration supplies judgment: why you choose one source, format, threshold, or exception over another.
• The generated Skill captures the procedure: triggers, ordered steps, outputs, and any reusable rules Claude inferred.
• Tools make the procedure executable: connectors, local files, Claude in Chrome, computer use, scripts, or MCP servers provide actual access.
• Tests decide whether it is trustworthy: a fresh input should produce the right artifact without silently skipping or inventing steps.

How to Record a Useful Claude Skill

1. Pick one bounded task. Choose a repeated job with a clear start and finish, such as turning approved meeting notes into a client folder and draft kickoff document.
2. Prepare clean sample data. Replace real names, account numbers, confidential terms, and personal information with realistic placeholders.
3. Open Cowork in the latest Claude desktop app. Use the plus menu and select Record a skill. If the option is missing, update and restart the app; rollout may still be phased.
4. State the goal before clicking. Explain what triggers the workflow, what inputs are required, and what a successful result looks like.
5. Narrate rules, not just movements. Say why a value is selected, which details vary each run, what must never be changed, and when Claude should stop and ask.
6. Finish at a safe boundary. Draft the email, report, or update, but do not demonstrate an irreversible send, payment, deletion, or publication unless the approval step itself is the lesson.
7. Save and inspect the Skill. Check the generated trigger, steps, assumptions, tool access, examples, and output format. Remove client-specific details that Claude mistook for universal rules.
8. Run a fresh test. Use a second synthetic case with different names, dates, missing fields, and one deliberate exception.
9. Back up the accepted version. Download it to a versioned skills folder or repository with an owner, date, and change log.

A Narration Template That Produces Better Skills

Record a Skill demonstration brief

Goal:
- This workflow turns [input] into [verified output].

Trigger:
- Run it when [specific condition is true].
- Do not run it when [exclusion or conflicting state].

Inputs:
- Required: [files, fields, source systems].
- Optional: [context that may be absent].
- Source of truth: [system or document].

As I demonstrate:
- This click is only navigation, not a business rule.
- This value changes every run: [variable].
- This format never changes: [invariant].
- I choose this option because [decision rule].
- If [exception], stop and ask [owner] instead of guessing.

Output:
- Create [artifact] in [location].
- Include [required sections or fields].
- Cite or link every source used.

Approvals:
- Claude may prepare [drafts or previews].
- A human must approve before [send, publish, pay, delete, deploy].

Verification:
- The run is complete only when [observable success condition].
- Report missing inputs, failed steps, and assumptions separately.

After Claude creates the Skill, run this audit prompt before trusting it:

Review the Skill you just created as an operations engineer.

1. Separate variables from fixed rules.
2. List every assumption inferred from this one demonstration.
3. Identify hidden dependencies, credentials, and tool permissions.
4. Add missing-input handling and explicit stop conditions.
5. Put sends, publishing, payments, deletions, contract changes,
   and financial status updates behind human approval.
6. Define the output contract and evidence of completion.
7. Create three acceptance tests:
   - normal case
   - missing-data case
   - conflicting-data case
8. Show me the revised Skill before saving it.

Workflow One: Make Client Onboarding Safe to Replay

Miles' first demonstration starts with meeting notes, creates a client folder and terms document, drafts an email, updates a tracker, watches for payment confirmation, and stores future meeting notes. That is a strong candidate because it repeats and leaves visible artifacts. It also crosses several trust boundaries.

The best version produces a review pack: proposed folder path, draft document, draft email, tracker diff, evidence used, missing information, and actions waiting for approval. That saves the work while keeping accountability visible.

Workflow Two: Outlier Research Without Copying Creators

Miles' second Skill studies YouTube channels, identifies videos that outperform a channel's baseline, gathers transcripts, explains why the packaging worked, and proposes adjacent ideas. His key editorial point is good: a small channel jumping from normal traffic to an exceptional result can be a more useful signal than a large channel's ordinary distribution.

A responsible research Skill should turn that observation into a transparent process:

1. Collect the source URL, channel, publish date, views at observation time, and the channel's recent baseline.
2. Calculate an outlier ratio instead of comparing raw view counts across differently sized channels.
3. Separate topic timing, title promise, thumbnail pattern, format, audience, and distribution hypotheses.
4. Use transcripts for research and attribution, not for reproducing another creator's script or distinctive expression.
5. Generate original angles for the user's expertise, audience, evidence, and semantic territory.
6. Return source links, assumptions, and confidence rather than presenting a causal guess as fact.
7. Keep final topic selection, claims, and creative direction with a human editor.

Platform access matters too. Use official exports, APIs, or permitted tools where possible, respect rate limits and terms, and do not let a recorded browser habit become an uncontrolled scraping routine.

Teaching Judgment, Not Just Copywriting Movements

Miles suggests recording himself writing posts each morning while explaining his choices. That is where demonstration can be more valuable than a style prompt. The clicks are unimportant; the useful information is the decision process:

• What makes this audience care now?
• Which claim is evidence-backed, and which is opinion?
• Why does the hook promise a specific payoff?
• Which phrases sound generic or unlike the creator?
• How is the draft made distinct from the research sources?
• What would make the creator reject the post?

Capture accepted examples and rejected examples with reasons. A pile of polished posts teaches surface style; paired examples teach boundaries. Then measure the Skill on edit distance, factual corrections, approval rate, and whether the draft sounds original rather than merely fluent.

The Four-Layer Automation Design

This separation prevents a common mistake: putting credentials, client facts, and procedural rules into one giant Skill. Context changes by project. Tools change by environment. The procedure should remain portable, while approvals remain explicit.

Access, Permissions, and Privacy

Claude's launch post lists Record a Skill for Pro, Max, and Team users in the desktop app. Treat that as availability for creating the Skill, not a promise that every observed action can be replayed through raw screen control. Anthropic's separate Cowork computer-use guide currently describes computer use as a Pro and Max research preview, says Claude prefers connectors first, and warns that complex screen workflows can need retries.

The same guide explains that screen interaction has no sandbox between Claude and approved applications. Claude can see information visible on screen and asks for app permission, but safeguards are not perfect. Browsing content can also contain prompt injection that tries to redirect the agent.

• Use a clean desktop, test workspace, sample clients, and disposable documents for the recording.
• Disable notifications and close password managers, chat apps, finance tools, health records, legal files, and unrelated client windows.
• Do not type passwords, API keys, recovery codes, card numbers, or private tokens while recording.
• Prefer a connector or API with scoped permissions over screen control when one is available.
• Deny sensitive apps and approve only the minimum applications required for the task.
• Review the generated Skill for copied personal data, literal filenames, recipients, account IDs, and hidden credentials.

Anthropic had not documented a retention policy specific to Record a Skill at the time of this check. Do not assume a recording is local-only. For consumer Free, Pro, and Max accounts, Anthropic says chats and coding sessions may be used to improve models when the user's model-improvement setting is on; with it off, its Claude Code data guide lists a 30-day retention period. Commercial Team and Enterprise data is not used for generative-model training by default unless the organization opts in. Check the current privacy controls and your organization's agreement before recording real business material.

The Skill Acceptance Checklist

• Trigger: the description says exactly when to use and not use the Skill.
• Input contract: required fields, optional fields, accepted formats, and source of truth are named.
• Output contract: file type, structure, destination, naming, and completion evidence are explicit.
• Variables: names, dates, prices, recipients, folder paths, and time windows are not hard-coded from the demonstration.
• Gotchas: known failure modes and app quirks are documented. Anthropic calls these among the highest-signal parts of a Skill.
• Permissions: allowed tools and denied apps are clear; credentials are external to the Skill.
• Approvals: sends, payments, publishing, deletion, deployment, legal changes, and financial updates require a human.
• Stop conditions: missing, stale, ambiguous, or conflicting data causes a question, not a guess.
• Verification: Claude checks the resulting state, not only whether it clicked through the steps.
• Tests: normal, missing-data, conflicting-data, and changed-interface cases are recorded.
• Ownership: one person owns review, versioning, access, and retirement of the Skill.
• Metric: measure accepted outputs, human correction time, failures, and cost per successful run.

What to Record First

A Seven-Day Rollout

1. Day 1: inventory repetition. List tasks repeated at least twice a month. Score each by time, stability, sensitivity, and ease of verification.
2. Day 2: choose one low-risk task. Define its trigger, inputs, output, owner, and success metric.
3. Day 3: prepare the demonstration. Create sample data, close sensitive apps, and write the narration brief.
4. Day 4: record and audit. Generate the Skill, remove accidental specifics, add gotchas, and define approvals.
5. Day 5: test variation. Run normal, missing-data, and conflicting-data cases. Record every correction.
6. Day 6: connect the safest tool. Prefer a scoped connector or API; use browser or computer control only when the precise tool does not exist.
7. Day 7: pilot with review. Run five real cases with a human checking every output. Promote only the steps that repeatedly pass.

The first goal is not "automate my whole business." It is one repeated workflow, five accepted runs, one measurable reduction in handling time, and zero unauthorized actions.

Video Chapters

Bottom Line

Miles Deutscher is right about the important part: Record a Skill makes AI workflow capture much easier for people who can demonstrate a job more clearly than they can write an automation specification. It can turn tacit habits into a reusable business asset while the work is already being done.

The durable advantage is not the recording button alone. It is the operating discipline around the generated Skill: clean inputs, narrated judgment, explicit boundaries, scoped tools, human approvals, fresh-case tests, failure logs, and versioned improvements.

Show Claude the work once. Then inspect what it learned, test what it can repeat, and keep control of what it is allowed to finish.

Sources

• Creator walkthrough: Miles Deutscher, "This NEW Claude Feature Is The Best Thing I've Used In 3 Years", plus the attached transcript. Follow @milesdeutscher on X.
• Official launch: Claude, "New in Claude Cowork: teach Claude a skill", 21 July 2026.
• Skills overview: Claude Skills and Skills explained.
• Skills practice: Anthropic, Lessons from building Claude Code: How we use Skills, 3 June 2026.
• Skills help: Use Skills in Claude.
• Cowork safety: Let Claude use your computer in Cowork.
• Consumer privacy: Is my data used for model training?.
• Related workflow: Codex Record & Replay Turns Screen Workflows Into Reusable Skills.