16 · Sources & verification log
Every official Anthropic source consulted, what was taken from each, and — importantly — what could not be verified. All retrieved 2 August 2026.
Sourcing policy applied
Official Anthropic properties only: claude.com (docs, product, blog), code.claude.com/docs, support.claude.com, platform.claude.com, trust.anthropic.com, anthropic.com. No third-party reviews, blogs, benchmarks or community sources were used anywhere in this pack.
Cowork — primary sources
| Source | Used for |
|---|---|
| claude.com/docs/cowork/overview | Core definition ("same agentic architecture that powers Claude Code"); four key capabilities; extensibility via connectors/skills/plugins/monitoring; the ~/.claude non-reading statement; Customize as the management surface. |
| claude.com/docs/cowork/guide/projects | Project contents table; three creation paths; drag-and-drop behaviour; 50 MB file read limit; Cowork vs claude.ai project comparison; archiving semantics; "Projects live on your computer. They aren't synced to the cloud or shared with other people." (two sentences; §05 condenses them with an ellipsis) |
| claude.com/docs/cowork/guide/dispatch | Dispatch definition and prerequisites (Pro/Max, latest Desktop, macOS/Windows); routing table (Code vs Cowork); six task states; the 10-minute permission auto-deny; no grandchild tasks; mobile briefing flow. |
| claude.com/docs/cowork/guide/plugins | "Plugins are available in Cowork and Code"; four plugin component types; marketplace and git-repo-as-marketplace; org-required plugins; full limits table. |
| claude.com/docs/cowork/monitoring | OTel setup fields; Team/Enterprise gating; version requirements (v1.1.4173+, v1.17377+); all six event types and their attributes; prompt.id correlation; standard and resource attributes; content-capture defaults; the user.email-always-included note; "Cowork VM" and egress behaviour. |
| claude.com/docs/cowork/changelog | Release cadence evidence; confirmation that plugin hooks exist in Cowork; general volatility assessment. |
| claude.com/pricing | Authoritative plan pricing: Free $0; Pro $17 annual / $20 monthly; Max from $100 (5× or 20×); Team standard $20 and premium $100 per seat annual, 2–150 seats; Enterprise seat price plus usage at API rates. Plus Anthropic's "prices and plans are subject to change" footnote. |
| claude.com/product/cowork | Which tiers include Cowork; the Max 20× $200 figure; target audiences; "say what, not how"; unattended operation; parallel processing; cross-platform claims; the stated limitations list including "consumes limits faster than Chat", beta/research-preview status, and the audit-log/Compliance API gap. |
| claude.com/blog/the-claude-cowork-product-guide | "Knowledge work agent" framing; "describe the goal, desired outcome and cadence"; local file access, cross-app integration, scheduled tasks, subagents. |
| claude.com/docs/office-agents/fsi-plugins | Existence and scope of the official financial-services plugin set for Cowork. |
| claude.com/blog/bringing-claude-code-and-claude-cowork-to-government | Evidence that the two products are positioned and sold as a portfolio. |
Cowork — Help Center
| Source | Used for |
|---|---|
| Get started with Claude Cowork | System requirements; supported platforms; plan availability and rollout order; the four-step enablement flow; the three permission modes and the Auto-mode cost note; file access and supported file types; usage-limit explanation; known limitations (chat memory doesn't carry over; no session sharing; desktop-only features). |
| Use Claude Cowork safely | Remote execution model and the "processed on Anthropic's servers" statement; the two-condition prompt-injection rule; built-in protections list; user responsibilities; oversight-to-stakes guidance. |
| Cowork on Team and Enterprise plans | Org-wide toggle (on by default); cloud-session defaults per plan; the connector "Always allow" setting; four-tier plugin marketplace governance; local vs cloud data handling and Compliance API coverage; network egress carve-outs; custom roles and groups. |
| Cowork on web, desktop, and mobile | The desktop/web/mobile feature matrix; remote-session semantics; the exact list of what requires the desktop app open. |
| Schedule recurring tasks in Claude Cowork | Remote execution of scheduled tasks; all-paid-plans availability and beta status; cadence options; guided vs manual setup; management actions; the "can't be tied to a folder on your computer" constraint; documented use cases. |
| Let Claude use your computer in Cowork | Pro/Max-only availability; macOS/Windows via Desktop; research-preview status; the connector → browser → screen priority order; per-application permission; default and custom blocklists; stated limitations and the sensitive-application recommendation. |
| Assign tasks from anywhere in Claude Cowork | Mobile/Dispatch pairing; the connector-first resolution order; local file reachability rules for remote sessions. |
| Claude Cowork collection index | Enumerating the complete set of published Cowork help articles. |
Claude Code — primary sources
| Source | Used for |
|---|---|
| Overview | Core definition; install commands for every method; the five surfaces; the eight "what you can do" categories; the "use Claude Code everywhere" routing table; shared-engine statement. |
| How Claude Code works | The agentic loop and its three phases; "agentic harness"; the five tool categories; what Claude can access; the three execution environments; sessions, forking, resuming; context window and compaction; checkpoints; the four permission modes; all five "work effectively" principles. |
| Platforms and integrations | Platform selection table; integration table; the "work away from your terminal" comparison; CLI-only features; cross-surface config sharing. |
| Feature availability | Everything in the provider matrix; the features-requiring-a-subscription list; the per-plan availability table including the Pro/Max-only and Team/Enterprise-only rows; all footnotes. |
| Extend Claude Code | Feature-to-goal matching table; build-your-setup-over-time triggers; all six "compare similar features" tabs; layering and precedence rules; context-cost-by-feature table; the "put guardrails in hooks" statement. |
| How Claude remembers your project | CLAUDE.md scopes and locations; load order and concatenation; imports and the external-import dialog; the 200-line guidance; .claude/rules/ and path scoping; auto memory storage, limits and behaviour; AGENTS.md interop; managed CLAUDE.md and claudeMd; claudeMdExcludes; the settings-vs-CLAUDE.md enforcement table. |
| Security | Permission-based architecture; built-in protections; working-directory boundary; prompt-injection safeguards; MCP security stance; cloud execution security controls; Remote Control security model; best practices; HackerOne reporting. |
| Data usage | Training policy by account type; Development Partner Program; all retention periods; encryption at rest by provider; telemetry services and opt-outs; default behaviours by API provider; WebFetch domain safety check. |
| Run agents in parallel | The subagents / agent view / agent teams / dynamic workflows comparison and their maturity labels; worktrees and /batch; the token-multiplication note. |
| Explore the .claude directory | Structure of project .claude/ and ~/.claude; what each file governs. |
| Documentation index (llms.txt) | Enumerating the full Claude Code documentation set (200+ pages); establishing the existence and one-line description of features cited but not individually fetched (routines, channels, workflows, sandboxing, analytics, costs, artifacts, output styles, gateway, worktrees, code review, ultraplan, ultrareview, goal, computer use, mobile, slack, chrome, github-actions, gitlab-ci-cd, managed-mcp, server-managed-settings, zero-data-retention, corporate-launcher, plugin-marketplaces, champion-kit, communications-kit, desktop-scheduled-tasks, and others). |
| Claude platform documentation index | Enumerating the complete Cowork documentation set and locating the skills, plugins and connectors sections. |
Learning material & adoption frameworks (§12, §13)
These sources were gathered in a second research pass on 2 August 2026. The two highest-value items — the deployment guide and the course catalogue — were re-fetched and verified directly rather than accepted at second hand.
| Source | Used for |
|---|---|
| Deploying Claude Across Your Organization (PDF, 29 Apr 2026, ungated) Direct PDF · announcement post |
Verified directly. The five-level maturity model and its "Chat is everyone's Level 0" framing; the six-month Evaluate/Pilot/Scale roadmap; the Chat vs Cowork vs Code surface table; the four use-case selection criteria; starting points and metrics by function; the adoption quotes; the Thomson Reuters, Jamf, Zapier, Airtree and Anthropic-internal customer examples. |
| claude.com/resources/courses | Verified directly. The complete course catalogue with durations, lecture and quiz counts, and the Professional / Engineering / Education / Nonprofits category taxonomy. Confirms that Introduction to Claude Cowork publishes no duration. |
| anthropic.com/learn and its three sub-hubs (claude-for-work, build-with-claude, claude-for-you) | "Anthropic Academy" branding and subhead; the seven-item inventory; the three-tile persona split; section headings on each sub-hub; the role guides; the observation that the business page lists nonprofit and educator courses. |
| claude.com/resources/tutorials | Role-specific Cowork walkthroughs (legal, sales, marketing ops, small business); the "Delegating your first task" 5-minute video; the AI-literacy set; the Category / Features / Products filter taxonomy. |
| Claude Cowork Enterprise Administrator Guide | The four-phase rollout including a phase titled "Change Management & Launch"; the three-question adoption measurement model; the "In Chat, users collaborate… In Claude Cowork, users delegate" and "An ideal Claude Cowork task produces a deliverable" lines; the inclusion of change-management leaders in the stated audience. |
| Scaling workflows with Claude Cowork | The three governance postures (Admin Curated / Guided Creation / Fully Open); the one-champion-per-25–50-users ratio; the >70% weekly-active-users week-4 target. |
| Making Claude Cowork ready for enterprise | Enterprise control surface (RBAC, group spend limits, usage analytics, OpenTelemetry, per-tool connector controls); the Zapier quote. |
| Scaling Agentic Coding Across Your Organization | The three-stage Claude Code rollout (super users → hackathon → internal expertise); the ROI metric list; "Agentic coding shifts software development from writing every line to guiding implementation." |
| Claude Code communications kit · docs version | Launch templates, the 20-message drip campaign, the Executive Sponsor Send variant, the C-suite sponsor checklist item, and the kit's own "draft copy, not finished copy" caveat. |
| The AI Fluency Framework (one-page PDF) | The four competencies and their verbatim definitions; the three modes (Automation / Augmentation / Agency); the "effective, efficient, ethical and safe" framing; the CC BY-NC-SA 4.0 licence and 2025 Dakan / Feller / Anthropic copyright. |
| Four role-based Claude certifications | The four certification tracks and their target roles; Partner Academy / Pearson / Credly mechanics; the foundation-to-professional progression statement. |
| Claude Partner Network | The $100M commitment to the Claude Partner Network for 2026. (The 36,000-consultant, 1,300-organization and 400,000-completion figures come from the certifications blog post, not this page — corrected 2 Aug 2026 after review.) |
| novo-nordisk · palo-alto-networks · ig-group · cox-automotive | Each fetched and verified directly, 2 Aug 2026, after a review flagged them as unchecked. Every figure now in §12 is quoted from its own page, and each row carries the product the case study is actually about — only Novo Nordisk is Claude Code. A "review cycles down 50%" figure previously attributed to Novo Nordisk was not on the page and has been removed. The Cox Communications 7× ROI figure is a different company and remains unverified. |
| anthropic.skilljar.com | Course delivery platform; module and lesson breakdowns for Claude 101, Introduction to Claude Cowork, Claude Code 101 and Claude Code in Action, including their stated audiences and prerequisites. |
Second-pass methodology
The §12 and §13 material was gathered by a dedicated research agent working under the same official-sources-only rule, then spot-verified: the deployment guide and the course catalogue were re-fetched independently before anything from them was written up. Where the agent flagged a claim as unconfirmed, that flag is preserved on the page itself rather than quietly dropped.
Cited but not individually fetched
The following were referenced through the documentation indexes and cross-links rather than fetched page by page. Claims drawn from them are limited to their published one-line descriptions, which are quoted or closely paraphrased:
routines · channels · workflows · sandboxing · sandbox-environments · analytics · costs · artifacts · output-styles · worktrees · code-review · ultraplan · ultrareview · goal · computer-use · mobile · slack · chrome · github-actions · gitlab-ci-cd · managed-mcp · server-managed-settings · zero-data-retention · corporate-launcher · plugin-marketplaces · champion-kit · communications-kit · desktop-scheduled-tasks · claude-apps-gateway-spend-limits · advisor · fast-mode · model-config · prompt-caching · desktop · agent-sdk/* · docs/cowork/3p/extensions · docs/skills/* · docs/plugins/*
What could not be verified
| Open question | Status | How to resolve |
|---|---|---|
| A complete Cowork tool reference (the equivalent of Claude Code's tools-reference) | Does not appear to exist publicly. Cowork capabilities are described narratively, not enumerated per tool. | Ask your Anthropic account team. Do not present an inferred list as documented. |
| Whether Cowork supports Zero Data Retention | Not documented. ZDR is described only for Claude Code / API. | Account team, in writing, before making any commitment. |
| Cowork's model-selection controls outside scheduled tasks | Partially documented. A model can be chosen when configuring a scheduled task; per-session selection is not documented to the same depth as Claude Code's /model. | Verify in-product on the current desktop build. |
| Whether Cowork can be billed through Bedrock / Google Cloud / Foundry | Not documented. Only Claude Code has a published third-party-provider story. | Account team. Note that "Claude Desktop on 3P" exists and runs the Code tab on those providers — that is not the same as Cowork. |
| Cowork context-window size and compaction behaviour | Not documented. | Treat Claude Code's documented behaviour as indicative only; do not state it as fact for Cowork. |
| Exact Cowork usage-limit consumption ratios vs Chat | Directional only. Anthropic says "faster" and "Auto mode consumes more"; no multipliers are published. | Measure with the api_request OTel event in your own environment. |
| Cowork checkpoint / undo behaviour | Not documented. Deletion requires explicit permission, but no rewind mechanism equivalent to Claude Code checkpoints is described. | Verify in-product before claiming either way. |
| Certification pricing and full syllabi | Gated. claude.com/resources/certifications and claude.com/partners/certifications both 404 — the first is a broken link on Anthropic's own courses page. claude.com/certifications redirects to a login-gated page. | Register at claude.com/partners, or ask the account team. Do not quote a price. |
| Contents of the two form-gated eBooks — The Enterprise AI Transformation Guide and Scaling Agentic Coding Across Your Organization | Partially verified. Summaries in §12 come from the ungated landing pages; the forms were not submitted. | Submit the form if you want the full text. The landing-page summaries are enough for curriculum planning. |
| The "4 Ds behavioural indicators" content | Not extracted. The tutorial page renders but the indicators appear to be embedded or interactive. The page references an "Anthropic Education Report: The AI Fluency Index" that was not located. | Open the tutorial in a browser. Do not reproduce indicators you have not read. |
| Cox Communications "7× ROI on first year of AI investment" | Unconfirmed. Surfaced in search results referencing a claude.com page, not confirmed against the case study itself. | Verify at claude.com/customers before citing. Flagged in-place on §12. |
| Recording of the "Claude Cowork in an Hour: Where do I start?" webinar | Not yet available. The page exists; the recording does not. | Re-check periodically — a Cowork-specific facilitated session would fill a real gap. |
| The AI Fluency loop names ("Description-Discernment", "Delegation-Diligence") | Login-gated. Present in the Skilljar course module listing; absent from the public one-pager PDF, which shows the four competencies as a flat set. | §14's nested-loop reading depends on them. Flagged in place on §13. Enrol in the free course to confirm. |
| Deployment PDF pages 17–22 | Not read. Pages 1–16 were verified line by line; the Anthropic-internal Legal and Finance detail in §12 comes from the unread section. | Re-read before relying on those two rows. |
| Where a Dispatch-spawned Cowork child task actually executes | Unresolved — sources conflict. Platforms says Dispatch runs on "Your machine (Desktop)"; the Dispatch guide says "The work runs on your desktop"; Cowork safety guidance says Cowork work runs on Anthropic's servers, separate from your computer and network. | The pack's reading (machine = host and file bridge; Cowork VM = executor) is an inference, flagged as such in §04. Ask the account team. Do not assert either reading in a room. |
| Whether Cowork activity reaches the Compliance API | Unresolved — sources conflict. claude.com/product/cowork: "Cowork activity is not yet captured in audit logs or Compliance API." Team/Enterprise help article: "Cowork via mobile and web is captured in the Compliance API." | The pack's local-vs-cloud reconciliation is ours. Get a written position before any regulated deployment. Flagged in §08. |
| The phrase "This is early-stage technology with imperfect safeguards" | Could not be located. Previously quoted in §08 as an Anthropic caveat on computer use; not found in either computer-use document on re-check. | Removed from §08 and replaced with sourced caveats. Do not reinstate without locating it. |
| Contents of the "Claude Cowork Security Overview" on the Trust Center | Listed but gated. Referenced at trust.anthropic.com; not retrievable without authentication. | Request through the Trust Center for security-review conversations. |
Methodology note
Content was retrieved from Anthropic's machine-readable documentation endpoints (llms.txt indexes and .md page variants) wherever available, which returns documentation source rather than rendered marketing HTML and therefore preserves tables, footnotes, version annotations and exact wording. Help Center articles and product pages were retrieved as published. Where a figure or claim appears only on a marketing page rather than in documentation — chiefly pricing and audience segments — the page is identified as the source and the claim is flagged as more volatile.
If you extend this pack
Keep the sourcing rule and keep this page honest. The value of the material to your executive audience depends almost entirely on their believing that you know the difference between what's documented, what's inferred, and what you don't know.