Claude

My own guide to using Claude says Claude Code has three interaction modes. Anthropic's current documentation retires two of those names and shows five in the selector. Here is the current list, the sentence saying plan mode runs shell commands rather than only reading, the one mode that does not persist between sessions, and the three questions worth asking anybody you pay to run Claude Code against your business.

I published a sentence on my own Claude guide saying the 1M context window was an API number. It isn't anymore. Opus 5 and Sonnet 5 now carry 1M in chat on every paid plan. But Anthropic's own platform docs name the catch: accuracy and recall degrade as the window fills, something they call context rot. Here are the three gates that decide your real ceiling, and the questions to ask before you pay somebody to build on long inputs.

A client asked me what I had actually touched inside their Claude Team plan, and on their plan there was no place to look it up. Nobody is credentialed in this work yet, so screening on credentials gets you nothing. Anthropic publishes the roles table, the connector limits and the one sentence that says local Cowork history cannot be exported by admins. Six questions to ask about access before you hire anybody.

Somebody asked me why their Claude output token count was so much bigger than the text they got back. I gave a confident answer about setting a thinking budget, then went to pull up the docs and found the setting does not exist on the model they run. The manual budget is gone on every Claude 5 model. What replaced it, why reasoning tokens are billed whether or not you ever see them, and the three controls you have left.

Anthropic ranks computer use last in its own documentation, behind connectors, shell commands and the browser, and calls it the broadest and slowest tool Claude has. What screen control is actually reserved for, which apps get full control by default, and the sentence where Anthropic admits an approved app can open one you never approved.

Somebody asked me whether Claude in Chrome could get into their client portal and I gave a worse answer than I should have. The permission is scoped to a website, not to the task you asked for, and Anthropic says so in its own permissions guide. What always-allow really covers, what Claude screenshots, the actions it will never take, and the rules I wrote down before pointing a browser agent at my own job search.

Anthropic's own terms say you should not rely on any Output or Action without independently confirming its accuracy. Its safety doc says you remain responsible for everything Claude does on your behalf, and in the same article says you shouldn't expect to validate every individual command. What sits between those two sentences is the checking layer, and deciding to build it yourself or pay somebody is the actual do-it-yourself question.

Everybody quotes the line that Claude scheduled tasks keep running while your computer is asleep. A note in the same article says a task tied to a folder on your computer only runs locally. Which side your work falls on, what a skipped run actually does, and the guardrails that have to live in the prompt because nobody is watching.

Everybody repeats that Claude Cowork only touches folders you connect. Anthropic's own safety doc draws the line somewhere else: isolation limits where Claude's code runs, not what Claude reads or does. What a connected folder really hands over, the three channels that boundary does not cover, and the setting that governs more of your risk than the folder does.

Anthropic makes auto mode the default in Claude Code on August 14. In a 1,053-tester study humans caught 13.6% of dangerous commands while the classifier caught 89%. But it only blocks actions that are irreversible, destructive, or aimed outside your environment, and publishing something wrong is none of those. Where the other guardrails have to live, from a consultant who lost a finished page to a failure no classifier would ever flag.

AI engineers run $35 to $60 an hour on Upwork, but the build price is only one of three costs. The other two are the monthly run cost, which is yours forever, and the question almost nobody puts in writing: whose account does the system live in when the work is done. Real numbers from Anthropic and Upwork, plus published rates from a consultant who shows his.