{"id":2831,"date":"2026-08-24T04:03:45","date_gmt":"2026-08-24T04:03:45","guid":{"rendered":"https:\/\/dmarketertayeeb.com\/blog\/claude-computer-browser-skills-files-marketing-guide\/"},"modified":"2026-08-24T04:03:45","modified_gmt":"2026-08-24T04:03:45","slug":"claude-computer-browser-skills-files-marketing-guide","status":"publish","type":"post","link":"https:\/\/dmarketertayeeb.com\/blog\/claude-computer-browser-skills-files-marketing-guide\/","title":{"rendered":"Claude Computer Use, Browser Use, Skills and Files: A Marketing Team Guide"},"content":{"rendered":"<article>\n\n\n<p>Anthropic has made four pieces of its agent stack generally available on Claude Platform: the updated computer use tool, a new browser use tool, the Skills API and the Files API. The announcement is dated 20 August 2026.<\/p>\n\n\n\n<p><strong>Editorial note:<\/strong> This guide is based on Anthropic&#8217;s 20 August announcement and the current documentation checked on 24 August 2026. DMT did not connect a live Claude account, run a browser or computer tool, access Ads Manager or publish a CMS change for this draft; the examples are adapted from documentation.<\/p>\n\n\n\n<p><strong>Definition:<\/strong> In this guide, \u201cClaude agent APIs\u201d means four distinct surfaces: browser use for webpages, computer use for desktop interfaces, Skills for reusable operating instructions, and Files for workspace-scoped inputs and outputs.<\/p>\n\n\n\n<p>For a marketing team, the practical change is not \u201cClaude can now run marketing by itself.\u201d It is that a repeatable workflow can now combine a procedure, business files, a browser or desktop interface, and a reviewable output in one agent loop. The safe starting point is read-only analysis and document production. Publishing, spending, sending, accepting terms, changing permissions and editing production systems should remain approval-gated.<\/p>\n\n\n\n<p>The short rule is:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Use <strong>browser use<\/strong> when the task stays inside web pages.<\/li>\n<li>Use <strong>computer use<\/strong> only when the work genuinely needs a desktop or native application.<\/li>\n<li>Use a <strong>Skill<\/strong> to package the team&#8217;s method, templates and checks.<\/li>\n<li>Use the <strong>Files API<\/strong> to pass source documents and retrieve artifacts with an explicit lifecycle.<\/li>\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\">What Anthropic announced on 20 August<\/h2>\n\n\n\n<p>Anthropic says computer use, browser use, Skills API and Files API are available on Claude Platform. Computer use can take several actions in one turn. Browser use adds page structure to screenshot-based interaction. Skills can be uploaded and versioned, then loaded when a task needs them. Files can be uploaded once, referenced by ID in later requests and used to hold generated outputs.<\/p>\n\n\n\n<p>The availability statement has boundaries. Anthropic says Skills and Files are also available through Microsoft Foundry, while the updated computer and browser tools are coming soon to Google Cloud&#8217;s Vertex AI. The current tool documentation says the client computer and browser toolsets are not currently available in Claude Managed Agents. A team should therefore verify the exact API, model, cloud and account combination before building a dependency.<\/p>\n\n\n\n<p>For adjacent managed-agent context, see DMT&#8217;s <a href=\"https:\/\/dmarketertayeeb.com\/blog\/gemini-managed-agents-marketing-workflows\/\">Gemini Managed Agents marketing-workflow guide<\/a>. It is a vendor-specific comparison, not evidence that Anthropic exposes the same availability or execution model.<\/p>\n\n\n\n<p>This is a platform release, not a marketing-performance benchmark. Anthropic includes customer case-study figures in its announcement, but those figures describe one vendor-selected insurance workflow. They do not predict the review time, accuracy, cost or conversion impact of a marketing team&#8217;s workflow.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">The four surfaces, in plain language<\/h2>\n\n\n\n<figure class=\"wp-block-table\"><table>\n<thead>\n<tr>\n<th>Marketing need<\/th>\n<th>Better-fit surface<\/th>\n<th>What it does<\/th>\n<th>First guardrail<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Read a JavaScript dashboard, compare fields or complete a web form<\/td>\n<td>Browser use<\/td>\n<td>Acts inside the browser your application runs; reads page structure and can target elements by reference as well as pixels<\/td>\n<td>Allowlist domains; treat every page as untrusted input; begin read-only<\/td>\n<\/tr>\n<tr>\n<td>Operate a native desktop app, file manager or non-web interface<\/td>\n<td>Computer use<\/td>\n<td>Uses screenshots plus mouse and keyboard control in an environment you provide<\/td>\n<td>Use a dedicated VM\/container; remove credentials and high-impact actions<\/td>\n<\/tr>\n<tr>\n<td>Apply a brand, QA or reporting procedure consistently<\/td>\n<td>Skills<\/td>\n<td>Loads a reusable folder of instructions, scripts, templates and resources when relevant<\/td>\n<td>Pin a version in production; review Skill changes like code<\/td>\n<\/tr>\n<tr>\n<td>Reuse a brief, export or generated report across turns<\/td>\n<td>Files<\/td>\n<td>Stores files by workspace and ID, with expiry and download behavior<\/td>\n<td>Classify source versus generated files; set expiry; do not assume expiry guarantees deletion<\/td>\n<\/tr>\n<\/tbody>\n<\/table><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\">Browser use versus computer use<\/h3>\n\n\n\n<p>If the work happens entirely in a web page, browser use is the more direct choice. Anthropic&#8217;s browser tool can read the accessibility tree, forms and tabs, and can act on a referenced element. The application\u2014not Anthropic&#8217;s infrastructure\u2014runs the browser executor.<\/p>\n\n\n\n<p>Computer use is broader. It gives Claude screenshot, mouse and keyboard control over a desktop environment. That is useful for a native spreadsheet application, a file manager, a legacy desktop tool or a workflow that crosses the operating system. It is unnecessary overhead for a web-only report.<\/p>\n\n\n\n<p>Both are client-controlled toolsets: your application supplies the browser or desktop environment, executes the returned browser or computer actions and sends the results back to Claude. Anthropic&#8217;s model does not run that environment for you.<\/p>\n\n\n\n<p>For another vendor&#8217;s adjacent computer-agent framing, DMT&#8217;s <a href=\"https:\/\/dmarketertayeeb.com\/blog\/grok-bot-cloud-computer-agent\/\">Grok Bot cloud-computer agent guide<\/a> is useful context, but it is not a proxy for Anthropic&#8217;s API behavior.<\/p>\n\n\n\n<p>For marketing operations, the distinction matters. A browser action may still change an ad, a CMS draft or a customer record. Better element targeting does not turn a consequential action into a safe one. Keep the action read-only until the team has an approval and rollback path.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Four practical marketing workflows<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">1. Weekly performance review<\/h3>\n\n\n\n<p>Upload approved exports or reports through the Files API. Attach a versioned Skill that defines the metrics, date window, exclusions, anomaly checks and output format. Ask Claude to produce a source-linked findings memo and a list of questions for the analyst.<\/p>\n\n\n\n<p>If the source is a web-only dashboard, browser use may retrieve the allowed fields. If the data can be exported through an API or file, that is usually a simpler and more stable input. The agent should identify missing fields and conflicting dates rather than silently filling them in.<\/p>\n\n\n\n<p>The output is a report for a human analyst. It is not permission to change bids, budgets, audiences or conversion settings.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">2. Content and SEO quality assurance<\/h3>\n\n\n\n<p>Create a Skill containing the approved editorial checklist: target reader, search intent, factual-source requirements, internal-link rules, metadata checks, accessibility requirements and publication exclusions. Use browser use against a staging site or preview URL to inspect headings, links, forms and responsive behavior. Store the brief, check results and proposed corrections as separate files.<\/p>\n\n\n\n<p>The safest output is a defect list or proposed patch. Keep the final publish step outside the agent until an editor reviews the claims, links and rendered page. DMT&#8217;s <a href=\"https:\/\/dmarketertayeeb.com\/blog\/ai-agent-harness-context-compaction\/\">AI agent harness guide<\/a> is useful background for separating worker evidence from a final publishing decision.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">3. Campaign and landing-page review<\/h3>\n\n\n\n<p>A performance team can ask an agent to compare an approved campaign export with a landing-page checklist, or to inspect a staging destination in a browser. The Skill can require the agent to report evidence for every proposed issue: the field, URL, date, source and confidence.<\/p>\n\n\n\n<p>Do not give a first pilot permission to change targeting, creative, budget, destination URLs or tracking. If the team later tests an action path, require a hard budget cap, a change preview, a named approver and a rollback instruction. Platform automation is not a substitute for ad-account governance.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4. Agency or in-house deliverables<\/h3>\n\n\n\n<p>Files and Skills fit recurring deliverables such as a client report, content brief, campaign QA memo or presentation. A Skill can encode the client&#8217;s format and approval language. Files can carry the approved source pack and the generated document between turns.<\/p>\n\n\n\n<p>Pin the Skill version used for the deliverable. \u201cLatest\u201d is useful while iterating, but an unpinned version can change the output when someone updates the shared Skill. The final artifact should record the source set, Skill version, model surface, date, reviewer and disposition.<\/p>\n\n\n\n<p>For recurring-task comparison, DMT&#8217;s <a href=\"https:\/\/dmarketertayeeb.com\/blog\/chatgpt-work-scheduled-tasks-marketers\/\">ChatGPT Work scheduled-tasks guide for marketers<\/a> covers a different vendor&#8217;s orchestration surface; keep its scheduling and permission assumptions separate from Anthropic&#8217;s Skills and Files APIs.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">What the API shape implies<\/h2>\n\n\n\n<p>Anthropic&#8217;s browser documentation uses a client toolset entry similar to this Python shape:<\/p>\n\n\n\n<pre><code class=\"language-python\">response = client.messages.create(\n    model=&quot;claude-opus-5&quot;,\n    max_tokens=2048,\n    tools=[{&quot;type&quot;: &quot;browser_toolset_20260801&quot;}],\n    messages=[\n        {\n            &quot;role&quot;: &quot;user&quot;,\n            &quot;content&quot;: &quot;Open the approved staging report and list missing fields; do not submit changes.&quot;,\n        }\n    ],\n)\n<\/code><\/pre>\n<p>This is an adaptation of Anthropic&#8217;s documented request shape, not a DMT test. The application must run each browser action and return the results to Claude. It should also inspect every action before execution when the workflow could have real-world consequences.<\/p>\n\n\n\n<p>Skills are supplied through the Messages API <code>container<\/code> and require code execution. For production, pin a specific Skill version rather than silently following <code>latest<\/code>:<\/p>\n\n\n\n<pre><code class=\"language-python\">container={\n    &quot;skills&quot;: [\n        {\n            &quot;type&quot;: &quot;custom&quot;,\n            &quot;skill_id&quot;: &quot;skill_your_workspace_id&quot;,\n            &quot;version&quot;: &quot;skver_your_pinned_version&quot;,\n        }\n    ]\n}\n<\/code><\/pre>\n<p>Anthropic&#8217;s current Skills guide allows up to 20 Skills per request and says custom Skills are workspace-scoped. A Skill bundle has a top-level <code>SKILL.md<\/code> with name and description metadata plus optional scripts\/resources. In a multi-tenant product, workspace boundaries deserve the same attention as API-key boundaries.<\/p>\n\n\n\n<p>Files are referenced by ID. Anthropic&#8217;s docs say uploaded files can be used as inputs, while files created by Skills or code execution can be downloadable when the file metadata allows it. Do not design a workflow that assumes every uploaded source can later be downloaded through the API.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Governance is part of the implementation<\/h2>\n\n\n\n<p>Anthropic&#8217;s computer and browser docs call out dedicated isolation, least privilege, domain allowlists, sensitive-data restrictions, human confirmation for meaningful consequences and prompt-injection risk. Those are practical requirements for marketing workflows, where a page may contain instructions that conflict with the team&#8217;s task.<\/p>\n\n\n\n<p>At minimum:<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li>Use a dedicated VM or container with minimal privileges.<\/li>\n<li>Start with synthetic data, exports or staging URLs.<\/li>\n<li>Keep credentials out of prompts, screenshots and source files.<\/li>\n<li>Allowlist only the domains needed for the task.<\/li>\n<li>Treat webpage text, screenshots and downloaded documents as untrusted input.<\/li>\n<li>Require human confirmation before sending, publishing, spending, deleting, accepting terms or changing permissions.<\/li>\n<li>Return a proposed change and evidence before permitting an action.<\/li>\n<li>Log the model surface, Skill version, source files, tool calls, approver and final result.<\/li>\n<li>Test the stop and rollback path before expanding access.<\/li>\n<\/ol>\n\n\n\n<p>Skills and MCP solve different problems. A Skill teaches a repeatable method; a connector or MCP server supplies access to a system. DMT&#8217;s <a href=\"https:\/\/dmarketertayeeb.com\/blog\/build-openai-agent-plugin-skills-mcp\/\">OpenAI Skills and MCP guide<\/a> covers that distinction in a different vendor context. Connecting a tool does not make its use safe or accurate.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Files, cost and retention details to check<\/h2>\n\n\n\n<p>The current Files documentation lists a maximum file size of 500 MB and total organization storage of 1 TB. File-related API calls are approximately limited to 500 requests per minute. Uploading, listing, retrieving metadata and deleting files are free; file content used in a Messages request is billed as input tokens.<\/p>\n\n\n\n<p>Files can be configured to expire between one hour and 90 days. After expiry, content is no longer retrievable through the API, but Anthropic says expiry is a lifecycle feature rather than a guaranteed-deletion control; metadata may remain visible for a period and underlying content may be retained for limited safety review. Set expiry deliberately and use deletion where appropriate.<\/p>\n\n\n\n<p>The current Files API documentation labels Files API \u201cZDR Not eligible.\u201d The Skills API guide separately says Agent Skills are not covered by ZDR arrangements and that Skill definitions and execution data follow Anthropic&#8217;s standard retention policy. Those are separate product statements, but they lead to the same operational rule: do not assume a ZDR arrangement covers either Files or Skills. Teams handling confidential brand, customer or account procedures should verify the current retention and workspace terms for the exact Claude surface, cloud and account before uploading a file or Skill. DMT&#8217;s <a href=\"https:\/\/dmarketertayeeb.com\/blog\/openai-zero-data-retention-private-safety-processing\/\">OpenAI ZDR guide<\/a> is useful for comparison, but its rules should not be transplanted to Anthropic.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">A safe first pilot<\/h2>\n\n\n\n<p><strong>Day 1: define the job.<\/strong> Choose one accepted output, such as a cited weekly report or staging QA checklist. Define what the agent must not do.<\/p>\n\n\n\n<p><strong>Day 2: prepare the source pack.<\/strong> Use exports or synthetic files. Record the owner, date range, permissions and retention requirement.<\/p>\n\n\n\n<p><strong>Day 3: write the Skill.<\/strong> Include the procedure, evidence rules, output schema, failure conditions and escalation triggers. Create a versioned snapshot.<\/p>\n\n\n\n<p><strong>Day 4: run read-only.<\/strong> Use browser use for web-only work or a dedicated computer-use environment where necessary. Capture every tool call and result.<\/p>\n\n\n\n<p><strong>Day 5: review.<\/strong> Measure accepted-output rate, factual defects, missing evidence, reviewer minutes and total request\/tool cost. Separate these observations from Anthropic&#8217;s case-study claims.<\/p>\n\n\n\n<p>Only after a bounded read-only pilot passes should a team consider a narrowly scoped action path. Even then, publishing, ad spend, customer messaging and permission changes should have explicit approval gates.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">FAQ<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">Is Claude computer use the same as browser use?<\/h3>\n\n\n\n<p>No. Computer use controls a desktop through screenshots and mouse\/keyboard actions. Browser use operates in an application-controlled browser and can read page structure, forms and tabs. Use browser use when the task stays inside webpages; use computer use for native or desktop work.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">Can a marketing team use it to change Ads Manager or publish a CMS page?<\/h3>\n\n\n\n<p>The tools can interact with software, but the announcement does not grant a marketing-specific permission model or make those actions safe. Start read-only. Put account changes, publishing, spend and customer-facing sends behind a human approval and rollback path.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">Are Skills just long prompts?<\/h3>\n\n\n\n<p>No. Anthropic describes Skills as reusable folders of instructions, metadata and optional resources that load when relevant and can compose with other Skills. They are better suited to a repeatable method than a one-off prompt. They still need version control and review.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">Does the Files API keep data forever?<\/h3>\n\n\n\n<p>Not necessarily. Files can be deleted or configured to expire between one hour and 90 days, but expiry is not a guaranteed-deletion control. Check the current retention policy and use workspace, access and audit controls appropriate to the data.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">Is the whole release available on every Claude platform?<\/h3>\n\n\n\n<p>No such universal claim should be made. Anthropic&#8217;s announcement and current docs describe different availability by surface. Verify the current model, API, cloud, workspace and Managed Agents compatibility before implementation.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">About the author and editorial process<\/h2>\n\n\n\n<p>Tayeeb Khan publishes DMT&#8217;s research-focused guides on AI and digital marketing. For this article, the DMT editorial workflow used Anthropic&#8217;s official announcement and product documentation as the evidence base, separated vendor case studies from independently verified facts, and reviewed the final copy for search ownership, factual limits, internal links and publication integrity. No hands-on product test or performance result is claimed.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Sources<\/h2>\n\n\n\n<p>The following first-party sources were accessed on 24 August 2026:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><a href=\"https:\/\/claude.com\/blog\/computer-use-skills-api-files-api\">Anthropic: Build production agents with computer use, the Skills API, and the Files API<\/a> \u2014 dated 20 August 2026; GA status, platform boundaries and release scope.<\/li>\n<li><a href=\"https:\/\/platform.claude.com\/docs\/en\/agents-and-tools\/tool-use\/computer-use-tool\">Computer use tool documentation<\/a> \u2014 client execution, action loop and security guidance.<\/li>\n<li><a href=\"https:\/\/platform.claude.com\/docs\/en\/agents-and-tools\/tool-use\/browser-use-tool\">Browser use tool documentation<\/a> \u2014 browser toolset, page structure and security guidance.<\/li>\n<li><a href=\"https:\/\/platform.claude.com\/docs\/en\/agents-and-tools\/agent-skills\/overview\">Agent Skills overview<\/a> \u2014 modular Skills and on-demand loading.<\/li>\n<li><a href=\"https:\/\/platform.claude.com\/docs\/en\/build-with-claude\/skills-guide\">Using Agent Skills with the API<\/a> \u2014 API integration, versioning, code execution and the Skills ZDR statement.<\/li>\n<li><a href=\"https:\/\/platform.claude.com\/docs\/en\/build-with-claude\/files\">Files API documentation<\/a> \u2014 file IDs, workspace scope, lifecycle, limits and the Files \u201cZDR Not eligible\u201d label.<\/li>\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\">Bottom line<\/h2>\n\n\n\n<p>Anthropic&#8217;s 20 August release makes a practical agent pattern easier to assemble: a Skill supplies the method, Files supply reusable inputs and outputs, browser use handles web applications, and computer use covers desktop software. That combination can make a marketing workflow more repeatable\u2014but it does not remove the need for evidence, access control or editorial judgment.<\/p>\n\n\n\n<p>The best first use is a bounded, read-only job that returns a reviewable artifact. Pin the procedure, expire the files, log the tool path and require a person to approve any action that can publish, spend, send, delete or change access. That is how a marketing team turns a capable API release into an accountable operating workflow.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Source and visual credits<\/h2>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Central facts and code-shape references: Anthropic announcement and Claude Platform docs linked in the source section above; accessed 24 August 2026.<\/li>\n<li>Supporting context: AIToolsReview and AI Codex pages linked above; discovery\/context only.<\/li>\n<li>Community signals: r\/AskMarketing, r\/ClaudeAI and r\/ClaudeWorkflows links above; questions and terminology only.<\/li>\n<li>Visual plan: DMT-created workflow table\/diagram, credited in caption to Anthropic&#8217;s official announcement and docs. No external image copied or generated in this worker handoff.<\/li>\n<\/ul>\n\n\n<\/article>","protected":false},"excerpt":{"rendered":"<p>Anthropic&#8217;s Claude agent APIs are GA. Learn when marketers should use browser, Skills and Files, plus the approval and data controls a safe pilot needs.<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[183,178,177],"tags":[193,196,194,247,293,195,422],"class_list":["post-2831","post","type-post","status-publish","format-standard","hentry","category-ai-in-marketing","category-artificial-intelligence","category-digital-marketing","tag-agentic-ai","tag-ai-agents","tag-ai-marketing","tag-anthropic","tag-claude-api-2026","tag-marketing-automation","tag-marketing-operations","no-featured-image"],"_links":{"self":[{"href":"https:\/\/dmarketertayeeb.com\/blog\/wp-json\/wp\/v2\/posts\/2831","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/dmarketertayeeb.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/dmarketertayeeb.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/dmarketertayeeb.com\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/dmarketertayeeb.com\/blog\/wp-json\/wp\/v2\/comments?post=2831"}],"version-history":[{"count":0,"href":"https:\/\/dmarketertayeeb.com\/blog\/wp-json\/wp\/v2\/posts\/2831\/revisions"}],"wp:attachment":[{"href":"https:\/\/dmarketertayeeb.com\/blog\/wp-json\/wp\/v2\/media?parent=2831"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/dmarketertayeeb.com\/blog\/wp-json\/wp\/v2\/categories?post=2831"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/dmarketertayeeb.com\/blog\/wp-json\/wp\/v2\/tags?post=2831"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}