Claude’s Record a Skill: Train Agents by Demonstration

Claude’s Record a Skill, shipped inside Claude Cowork on July 21, 2026, replaces prompt writing with demonstration: you record your screen while doing a task, talk through your reasoning as you go, and Claude converts the walkthrough into a reusable skill it can run again on demand. No instruction files, no prompt engineering — the demo is the spec.

The stakes are bigger than one feature. OpenAI shipped the same idea — Record & Replay in the Codex desktop app — five weeks earlier, on June 18. Two frontier labs converging on demonstration-based skill capture within a single product cycle is a signal about where office automation is heading: away from typed instructions, toward watching how your team actually works.

This guide covers what launched and for whom, how the recording flow works, a structured head-to-head against Codex Record & Replay that no outlet has published, and the lens the general coverage missed entirely: which CRM and back-office tasks demo well, which need judgment, and how to govern recordings that inevitably point a camera at business data.

Anthropic announced Record a Skill on July 21, 2026 via its official @claudeai account, and coverage from The Decoder, CyberSecurity News, and Search Engine Journal corroborated the details the same week. The feature lives in one place: the “+” menu of the Claude Desktop app, inside a Cowork session. It is not in the web app and not on mobile.

Access is gated to paid tiers. Every paid individual and team plan gets it; the Free plan is excluded entirely. That makes the practical entry price $20 per month — the same floor OpenAI set for Codex Record & Replay with its Plus tier.

The workflow Anthropic recommends, as summarized by Coursiv’s launch coverage, runs in eight steps: update and open the Claude Desktop app, start a Cowork task, open the “+” menu, choose Record a skill, perform the workflow yourself, narrate your choices, rules, and exceptions aloud as you work, let Claude convert the demonstration into a Skill, and then review and test that Skill on a fresh example before relying on it.

The narration is the part teams underestimate. The recorder captures what you click and type, but the spoken track is where the rules live — “we always check the branch field before assigning,” “if the amount is over the threshold, stop and escalate.” A silent demo produces a skill that mimics motions; a narrated demo produces one that carries reasoning.

Under the hood, the output is a Claude Skill — a pre-existing product concept Anthropic describes in its Skills tutorial as a reusable package of task-specific instructions Claude loads automatically when the context matches. Recording doesn’t change what a Skill is; it changes how one gets made. Before July 21 there were already four creation paths — recording is the fifth, and the first that requires no writing at all.

Coursiv’s comparison of Claude’s customization primitives is a useful mental model here: Instructions set broad rules, Project knowledge holds per-project context, Memory carries facts across conversations — and Skills are loadable, task-specific procedures. Record a Skill is the newest way to create that last category, not a replacement for the other three.

OpenAI got there first. Codex Record & Replay shipped on June 18, 2026 in the Codex desktop app (v26.616) — 33 days, just under five weeks, before Anthropic’s July 21 launch. The Decoder put it plainly: “OpenAI launched a comparable Record and Replay feature in June, making demonstration-based skill capture an industry standard within a single product cycle.”

Existing coverage namechecks the parallel in a sentence and moves on. The structured comparison matters more than the namecheck, because the two implementations differ on exactly the dimensions an operations team has to care about: platform, regions, session limits, and — critically — what each vendor has actually disclosed about how the recording is represented.

Two asymmetries stand out. First, reach: Claude’s version ships on Windows and macOS with no disclosed regional exclusions, while Codex is macOS-only and dark across the EEA, UK, and Switzerland — for a European or mixed-fleet operations team, Claude is currently the only one of the two you can actually roll out. Second, transparency runs the other way: OpenAI has documented how its recorder represents a demo — semantic intent as JSON, not pixel coordinates — while Anthropic has disclosed nothing about its internal representation. Whether Claude’s recorded skills survive UI changes the way Codex claims its replays do is, for now, an open question you can only answer by testing.

Every outlet covered this as a general productivity story. The more useful frame for operations teams is narrower: which CRM and back-office tasks actually demo well? Early coverage converges on a pattern — strong fits are repeatable, rule-based, and verifiable (weekly reporting, spreadsheet normalization, support-triage classification, publishing checklists); weak fits are one-off tasks, workflows driven by un-narrated intuition, unverifiable outputs, and anything touching money transfers, account recovery, contracts, passwords, or customer PII.

Mapped onto the workflows we build in CRM automation engagements, that pattern turns into a decision table. The dividing line isn’t task complexity — it’s whether the rules can be spoken aloud completely during a single demonstration.

The narration test is the practical heuristic: if a competent new hire could do the task correctly after watching you do it once while you talked through the rules, it’s a candidate. If they’d need three months of context to know when the rules bend, it isn’t — and no recording feature from any vendor changes that.

Here’s the honest gap in this launch: as of the announcement window, Anthropic had not published a dedicated technical or help page for Record a Skill — nothing covering recording retention, editing controls, or full rollout detail. Coursiv flagged the absence explicitly, and it means every governance decision right now has to lean on Anthropic’s pre-existing Cowork guidance rather than feature-specific policy.

That pre-existing guidance is worth reading before your first recording. Anthropic’s Use Claude Cowork safely page notes that Cowork can read files, browse the web, run code, and use connected apps, and that remote-session work is processed on Anthropic’s servers. Its computer-use guidance separately warns that information visible on screen during agentic use can include personal or confidential data. A feature whose entire input is your screen and your voice inherits both warnings at full strength.

AI Weekly’s launch analysis lists the questions the reporting doesn’t answer: what happens when the underlying app’s UI changes and a replay breaks, how the recording handles credentials or sensitive data scrolling past on screen, and whether skills can be shared or edited across a Team plan’s users. Until Anthropic documents those answers, the conservative reading is the right one — the same posture we argued for around enterprise agent-deployment governance: capability ships first, controls catch up, and the gap is where incidents happen.

Why did both labs ship the same feature within a single product cycle? Because demonstration solves the cold-start problem that has throttled office-work automation since the first prompt was typed: most operational knowledge is tacit. The people who run your back office can’t write a complete spec of what they do — but they can show it, and they narrate the exceptions naturally while doing it. Recording converts tacit knowledge into an executable artifact at the cost of a single walkthrough, which is the same show-don’t-tell logic behind Claude’s shift toward shareable, show-me interfaces elsewhere in the product line.

Our read: the skill library is the real product. Once a team has recorded twenty workflows into one vendor’s desktop app — on the same Cowork surface where Anthropic’s flagship Fable 5 model already runs — switching costs stop being about model quality and start being about re-demonstrating institutional knowledge. Expect both vendors to push team-level sharing, versioning, and admin controls next, because that’s what converts a personal convenience into an organizational asset the org can’t easily leave.

Temper the projection with the reliability record, though. As we noted in our Codex coverage, the 2026 AI Index put average AI-agent success on the OSWorld computer-use benchmark at 66% — demonstration narrows the gap between what you meant and what the agent does, but it doesn’t make replays reliable on its own. And the anxiety is already public: Search Engine Journal’s coverage quoted one user reaction — “This has to be the easiest way to be replaced and lose your job man.” — which is sentiment, not data, but it’s the sentiment ops leaders will be managing in every rollout conversation. The teams that win this transition will frame recorded skills as removing the work people complain about, and keep the judgment calls visibly human.

If you run a CRM-centric operation, this feature is worth a pilot this quarter — scoped tightly. Sort your candidate workflows into four lanes and move one lane at a time.

Two disciplines make the pilot real. First, review every generated skill like code: read it, test it on a fresh example, and check it for copied identifiers before it enters a shared library. Second, assign an owner — recorded skills are process documentation that executes, and unowned executable documentation rots into risk. If you want help scoping which workflows clear the bar and building the governance around them, that’s the assessment we run at the start of AI transformation engagements.