explainx / blog
Anthropic launched "Record a Skill" in Claude Cowork on July 21, 2026 — record a workflow like filing expenses, and Claude turns it into a reusable, rerunnable skill. Here's how it works, early use cases, and the privacy tradeoffs.
May 26, 2026
Comprehensive guide to Anthropic's Knowledge Work Plugins for Claude Cowork and Claude Code. Learn how to install, customize, and build plugins for sales, engineering, product, marketing, and more.
May 12, 2026
The /goal command lets you set a completion condition and Claude keeps working across turns until it's met. Works in interactive, -p, and Remote Control. Tracks elapsed time, turns, and tokens. Integrated into Codex and Claude Code Mobile.
Jul 22, 2026
The largest copyright recovery in history is final: $3,000 per pirated book, split between authors and publishers, plus a legal fee cut from $187.5M to $101M. Training on books stays fair use — buying or pirating them doesn't.
Claude Cowork just removed the last step that made "vibe automation" annoying: writing the prompt. On Tuesday, July 21, 2026, Anthropic shipped "Record a Skill" inside Claude Cowork — a feature that lets Pro, Max, and Team users screen-record themselves doing a task once, then has Claude analyze that recording and turn it into a reusable, rerunnable skill. No prompt engineering, no manually describing the steps — you just do the thing, and Claude figures out how to automate it.
This is the latest expansion of Cowork's shared workspace model — a Claude-controlled environment that already edits files, organizes folders, and interacts with apps, running on Anthropic's own servers rather than the user's local machine. Here's how the feature actually works, what people are automating with it in the first day, and the privacy questions that come with recording your screen for an AI to study.
| Question | Answer |
|---|---|
| What launched? | "Record a Skill" in Claude Cowork |
| When? | Tuesday, July 21, 2026 |
| Who has access? | Pro, Max, Team users — phased rollout |
| How does it work? | Screen-record a workflow → Claude analyzes it → generates a rerunnable skill |
| Prompt required? | No — the recording itself is the specification |
| Suggests connectors? | Yes, when one exists for a step in the recording |
| Works with? | Browser and computer use |
| Early use cases? | Trading alerts, LinkedIn tasks, email replies, expense filing |
| Open concern? | Privacy — what happens to recordings, and whether they train future models |
The mechanic is straightforward on the surface: open Claude Cowork's shared workspace, hit record, and perform a workflow the way you normally would — clicking through menus, filling in fields, navigating between apps. Anthropic's own example is filing an expense report, a task most people repeat identically every month or quarter but never bother scripting because writing that automation from scratch isn't worth the time for a once-a-month chore.
Once the recording is captured, Claude analyzes it — the sequence of clicks, navigation, and actions — and converts that into a skill definition it can execute again later. Anthropic engineer Lydia Hallie described the practical value directly: "I've been using it for workflows I always meant to automate but never did, or that had no connectors. Just record the workflow, and Claude turns it into a skill you can rerun. It also suggests connectors if they exist! Works with browser & computer use."
That connector-suggestion behavior is a meaningful design choice. Rather than always automating the raw pixel-level UI interaction — which is brittle if a website's layout changes — Claude checks whether a proper API-based connector already exists for a step in your recorded workflow and prefers that when available, falling back to browser or computer-use automation only when no structured integration exists.
One of the more interesting early reactions came from researcher Tobin South, who tested how forgiving the feature is about messy, imprecise recordings: "pro tip: you'd be amazed how dumb you can be while navigating your computer and still have Claude create a coherent skill." That's a meaningful usability claim — it suggests Claude is extracting the intent of the workflow from the recording rather than replaying a literal, brittle sequence of exact clicks and coordinates, which would break the moment a button moved a few pixels.
South also proposed a cost-optimization pattern worth noting for teams planning to use this at scale: "use a big model to generate the skill and a cheap model to execute it (like mini distillation)." That mirrors a pattern already common in agent harness design — a capable model does the expensive, one-time work of understanding and encoding a workflow, while a cheaper, faster model handles repeated execution once the skill is well-defined. It's the same tiered-model economics explainx.ai has covered in the context of agent skill security and Microsoft's SkillOpt — generation and execution don't need the same model tier.
In the first hours after launch, reported early use cases clustered around exactly the kind of repetitive, individually-low-value tasks that were never worth manually scripting: trading alerts, LinkedIn tasks, and email replies were specifically mentioned by users testing the feature. Those are a useful signal for where Record a Skill fits — not complex, high-stakes automation, but the long tail of small, repeated browser workflows that accumulate real time cost across weeks and months without ever individually justifying a custom script.
Not everyone was uniformly enthusiastic. Researcher Kiana Ehsani raised a broader framing critique that applies to Record a Skill as much as to prompt-based interaction generally: "I'm a big believer that current form of asking users to write down prompts is not the way human-AI interaction should look like... People should not have to write prompts! People should not have to think about what they can / can not delegate." Read charitably, Record a Skill is a partial answer to exactly that critique — it removes the prompt-writing step entirely — though it still requires the user to correctly judge which tasks are worth recording and delegating in the first place, a decision Ehsani's argument suggests shouldn't fall on the user at all.
Others noted this isn't an entirely novel category. Physician and researcher Derya Unutmaz pointed out he'd seen a similar capability elsewhere, linking to ChatGPT's own "record and replay" documentation — a reminder that Anthropic and OpenAI are converging on similar "watch me do it once" automation patterns roughly in parallel, rather than either side having a clean first-mover claim on the concept.
Availability is described as a phased rollout rather than a simultaneous flip for all eligible users — several people reported not yet having access to the feature shortly after the Tuesday launch, consistent with how Anthropic has staged other Cowork feature releases.
The more substantive open question is privacy. Screen recordings, by nature, capture whatever is visible during the workflow — which for tasks like expense filing or email replies could include account numbers, internal tool interfaces, colleague names, or other sensitive information incidental to the recorded task, not just the click sequence Claude needs to learn the automation. Some users flagged this directly, noting that "skills may help improve models" — raising the same question explainx.ai covered in the Claude Cowork safety guide: what data-retention and training-use policies apply to content captured inside a Cowork session, and whether recording a workflow involving regulated data (financial records, health information, credentials) carries different obligations than a normal chat conversation.
Until Anthropic publishes more specific retention and training-use terms for recorded skills, the practical guidance mirrors general screen-share hygiene: don't record workflows that surface credentials, unmasked financial account numbers, or regulated personal data on screen during the capture, even if the task itself is otherwise safe to automate. For workflows that do involve sensitive data, consider whether a connector-based integration — which Record a Skill will suggest when available — avoids exposing the raw UI (and whatever's on screen alongside it) altogether.
Record a Skill sits in a different category from most workflow-automation tools that came before it. Traditional robotic process automation (RPA) platforms and browser macro recorders have offered literal click-and-coordinate recording for years — but those recordings are famously brittle, breaking the moment a button moves, a page reflows, or a dialog appears in a different order. Record a Skill's pitch, backed by Tobin South's "you'd be amazed how dumb you can be" observation, is that Claude is extracting the semantic intent of the workflow rather than replaying a literal input trace, which should make it more resilient to minor UI drift than classic RPA macros — though that resilience claim hasn't yet been stress-tested publicly against real-world interface changes over weeks or months.
It's also worth separating this from general agent skills as explainx.ai has covered them elsewhere — reusable capability packages a developer writes explicitly, with defined inputs, outputs, and logic. Record a Skill inverts that authoring model: instead of a developer or power user writing a skill definition by hand, the skill is inferred entirely from an example recording. That's a lower floor for who can create automation (no scripting knowledge required) but a less transparent one — a hand-written skill's logic is auditable by reading it; a recording-derived skill's actual generalization behavior is something Claude inferred, and isn't necessarily obvious to the user who created it just by watching the source recording once.
Record a Skill extends a pattern Claude Cowork has been building toward since its earlier releases: a shared, server-side workspace where Claude has persistent access to files, folders, and app interactions rather than being invoked fresh for each isolated request. explainx.ai's Claude Cowork safety guide and is Claude Cowork safe coverage both flagged the tradeoff this architecture implies: more persistent, capable automation in exchange for a larger trust surface, since Claude's access spans an entire workspace rather than a single scoped task. Record a Skill is a natural extension of that model — it turns one-off automation into reusable, saved capability, which raises the stakes on getting the underlying access and retention model right, since a recorded skill isn't a single request anymore, it's a standing automation that runs again and again.
Feature details reflect Anthropic's July 21, 2026 launch and early user reports through July 22, 2026. Rollout is phased — availability, connector suggestions, and data-handling terms may change as the feature reaches general availability.
