Every factual claim in The Working Guide, with the Anthropic page it came from and the date that page was read. Where two of Anthropic's own pages disagree, both are listed below rather than one being quietly picked.
Most of Anthropic's help-centre articles show a relative timestamp rather than a date, so they are marked undated and carry our read date instead. Developer documentation moved from docs.claude.com to platform.claude.com during 2026; the links here use the current addresses. This page is rewritten each time the guide is re-verified.
What has changed since the first printing →
Six places where two official sources say different things. None of these are our findings about Claude; they are findings about the documentation. They are listed so you can check for yourself rather than take our word.
Are Skills on the Free plan?
The help centre says skills are available on Free, Pro, Max, Team and Enterprise. The developer documentation says custom skill upload is Pro and above. The two can be partly reconciled — one is about skills generally, the other about uploading your own — but as written they disagree, and neither acknowledges the other. The same pair also disagrees on where you upload: Settings then Features, versus Customize then Skills.
Does the skill recorder run on Windows?
The help centre says recording is macOS-only and not available on Windows. Six weeks later, the desktop changelog records a fix for “Record a skill opening an unresponsive chooser window on Windows.” Anthropic shipped a Windows fix for a feature its help centre says does not exist on Windows. The changelog is newer, so check in-app before relying on either.
Do Cowork projects sync between devices?
The help centre says new Cowork projects are saved to your Claude account so you can pick them up on other devices. The Cowork documentation says flatly that projects live on your computer and are not synced to the cloud. This one matters for anyone deciding where to keep their work.
Is Dispatch in beta, and what does it need?
Three disagreements in one pair. The help centre calls it a limited beta, available on only some Pro and Max plans, requiring both the desktop and mobile apps. The Cowork documentation gives no beta label, no mobile requirement and no plan qualifier. The app itself shows a Beta tag next to Dispatch in the left menu, which is the tiebreaker we used.
Is role prompting recommended?
An Anthropic blog post says modern models are sophisticated enough that heavy-handed role prompting is often unnecessary. The current prompting documentation still has a section headed “Give Claude a role” and says a single sentence makes a difference. The documentation is the newer artefact, since it covers models released in 2026. The honest reading: one line of role still helps; the elaborate multi-paragraph persona is what the blog is calling unnecessary.
Which platforms is Cowork on?
The help centre lists macOS and Windows. The product marketing page adds ChromeOS and Linux, and the changelog corroborates Linux with an apt auto-update entry. We follow the help centre in the book and note the wider list here.
ConfirmedCowork uses the same agentic architecture that powers Claude Code, with no terminal required.
ConfirmedIn Chat, Claude cannot access your files directly. In Cowork it can read, edit and create files in folders you specify.
ConfirmedCowork sessions run on Anthropic's servers by default. Local file access is limited to folders you have connected on the desktop.
ConfirmedCowork is on desktop (macOS and Windows), web and mobile in beta, and the Chrome side panel. Paid plans only.
ConfirmedCowork became generally available on macOS and Windows on 9 April 2026.
ConfirmedThree permission modes: manually approve, automatically approve, and skip all approvals. Anthropic advises the third only when you completely trust every action.
ConfirmedConnecting a folder: Projects, then +, then Use an existing folder.
Not publishedA click path for connecting a folder outside project creation.
ConfirmedThe lineup: Claude Fable 5.1, Opus 5, Sonnet 5 and Haiku 4.5.
ConfirmedContext windows: 1M tokens for Fable 5.1, Opus 5 and Sonnet 5; 200K for Haiku 4.5.
ConfirmedEffort levels are low, medium, high, xhigh and max. The API default is high, and setting high behaves identically to omitting the setting.
ConfirmedHaiku 4.5 does not support effort at all.
ConfirmedThinking cannot be turned off on Fable 5.1 or Opus 5.
Not publishedWhich effort level the app treats as default for each model. Anthropic marks one in the menu but never publishes which it is.
ConfirmedFast mode gives up to 2.5x more output tokens per second on Opus, at premium pricing, using the same model weights.
ContradictedAnthropic documents fast mode as an API-only research preview. The toggle in the app's model menu is not documented anywhere. It is on the screen regardless — see the walkthrough in chapter two.
ConfirmedProjects are self-contained workspaces with their own chat histories and knowledge bases.
ConfirmedProjects are available on Free accounts, to a maximum of five. This resolves a question left open in the first printing.
ConfirmedUploaded project documents are cached, and only new or uncached portions count against your limits on later use.
Not publishedWhether claude.ai projects sync between your devices. Never addressed; only that they live in your Claude account.
ConfirmedClaude creates an artifact when content is significant and self-contained, typically over fifteen lines.
ConfirmedPublishing makes an artifact available to anyone with the link. Team and Enterprise sharing keeps it inside your organisation.
ConfirmedOnce you unpublish an artifact you cannot publish that same artifact again.
ConfirmedSharing an artifact also gives viewers access to attachments and files in the conversation that created it. Anthropic states this under organisational sharing; no equivalent warning appears under public publishing.
ConfirmedPersistent storage is Pro and above, 20 MB per artifact, published artifacts only, and unpublishing deletes it permanently.
ConfirmedScheduled tasks run remotely, on their cadence, even when your computer is asleep or the desktop app is closed.
ConfirmedCadences offered are hourly, daily, weekly, weekdays or manual. All paid plans.
ConfirmedScheduled tasks were introduced on 25 February 2026.
ConfirmedDispatch runs on your desktop, so your computer must be awake with the app open. This is the opposite of a cloud session.
ConfirmedIf a Dispatch permission prompt goes unanswered for ten minutes, the request is automatically denied and the task carries on without that action.
Not publishedAny cap on the number of scheduled tasks, or an interval shorter than hourly.
ConfirmedSkills package instructions, metadata and optional resources that Claude uses automatically when relevant.
ConfirmedA skill folder is packaged as a ZIP with SKILL.md at its root and uploaded through Customize, then Skills.
ConfirmedA screen recording for building a skill can run about ten minutes.
ConfirmedUse skills only from sources you trust: your own, or Anthropic's.
ConfirmedConnectors are MCP servers, catalogued in the Connectors Directory across all Claude products.
ConfirmedVerification means Anthropic reviewed a connector more closely than a community one. It is explicitly not a security audit, and the developer can change its tools after review.
ConfirmedOnce connected, a community connector has the same capabilities and access as any other connector you grant.
ConfirmedMalicious MCP servers may carry hidden instructions. Review permission scopes at authentication and use Allow always only where you trust the server and the tool.
ConfirmedPlugins bundle skills, connectors and sub-agents. Some include local MCP servers that run with the same permissions as any other program you install.
Not publishedA count of first-party Anthropic plugins. No total is published; the directory changes. The first printing gave a number and should not have.
ConfirmedA token is approximately 3.5 English characters, varying by language. This is the figure the calculator on this site uses.
Confirmed1M tokens is roughly 555,000 words on the current tokenizer, introduced with Opus 4.7. Older guidance of about 750,000 words is now wrong.
ConfirmedUsage limits reset on a rolling five-hour window, and paid plans add a weekly limit on top.
ConfirmedLonger conversations that trigger automatic context management consume more of your usage limit. This resolves a question left open in the first printing: yes, it counts.
ConfirmedAs token count grows, accuracy and recall degrade. Anthropic calls this context rot, and it is why what is in the window matters more than how much room is left.
Not publishedAny absolute number of messages or hours per plan. Anthropic publishes multipliers only and reserves the right to change them.
ConfirmedThe best prompt is not the longest. It is the one that achieves your goal with the minimum necessary structure. Start simple and add complexity only when needed.
ConfirmedOpus 5 runs longer by default than earlier models, and changing effort does not reliably change response length. Ask for brevity explicitly instead.
ConfirmedClaude Academy is free AI education from Anthropic, with tracks covering the products, AI fluency, specialised audiences and building with the API.
Not publishedAn official course count. None is published; the catalogue showed 26 courses on the day this page was written.
ConfirmedAnthropic's trademark guidelines require prior written permission for use of its marks and contain no carve-out for independent educational content. This guide is unaffiliated and says so on every page.
Compiled 10 September 2026. Sources are Anthropic's own documentation, help centre, release notes and changelog except where a page is marked as a blog post. Nothing here is legal advice.