Issue 004 · 6 min read
Most people using a desktop AI app still think of it like a notepad—private unless they share. This week’s changes, especially to Claude, mean what you type can be pulled into a company review, even if you were working locally. Meanwhile, tools that once sat quietly in the background now introduce bigger review challenges as workloads are fanned across many agents or run automatically on new data. Automation buys you time back on the grunt work, but also sweeps in new ways for subtle mistakes to pass downstream. The only protection is clear standards and knowing which step actually needs a human to look directly.
Rather listen?
Skip if you like
An AI assistant is a tool you talk to in plain language, like a fast, very literal colleague who has read a great deal but knows nothing about your job until you tell it. It can draft, summarise and suggest, but it does not decide, and it can be confidently wrong.
A prompt is simply what you type or say to the assistant: your request, plus any background and examples you give it. The clearer and more specific the prompt, the more useful the answer, so treat it like briefing a new starter, not typing a search.
A context window is how much the assistant can hold in mind at once: your instructions, the document you pasted, and the conversation so far. When a chat gets long it starts to forget the earliest parts, so for a big task give it the key facts again rather than assume it remembers.
Start here
Newly announced
If you move old prompts to Claude Opus 5 this week, expect the AI to answer at greater length and sometimes over-explain or duplicate work. Some prompts that worked before will now drag out tasks or second-guess themselves, which might clog up your drafts or handovers.
See the detail and the sources →Newly announced
With Grok Build, you can now throw a sprawling job at hundreds of AI agents in parallel, but this means more places for weak findings or misunderstandings to sneak through. You still need to decide what your review check is, or risk trusting a committee with no one answerable.
See the detail and the sources →Newly announced
Grok 4.5 is being presented as ready for real workplace tasks instead of just chat or web summaries. The reality is, it may fit those jobs, or not. You'll only really know after you run your own spreadsheet or Word file through it, not from a launch slide.
See the detail and the sources →
The one that mattersWhat's genuinely new
Claude’s local session content on the desktop is now retrievable by Enterprise admins via the Compliance API, bridging a gap most people assumed was private. Workflow scaling and recurring automation are now not just possible but normal in Grok.
What the makers claim
Grok’s pitch about being ready for code and document-heavy workflow is their marketing spin; real-world fit has not been established. The length and apparent completeness of Claude’s transcript may oversell its evidentiary value.
Where it helps
If you use desktop AI apps for drafting sensitive client emails or early reports and assumed only you saw them, that assumption now needs checking.
Where it doesn't
If all your AI use happens through tightly managed web tools with controlled input and clear audit logs, the new compliance reach changes little for you.
What remains yours
Anthropic released Claude Opus 5 with a one million token context window, up to 128,000 output tokens, thinking enabled by default and standard API pricing of US$5 per million input tokens and US$25 per million output tokens.
Check if the transcript reflects your actual intent, agreement or next steps—not just what was typed.
Try this
Find out what’s really surfacing from your AI local work history.
Export a transcript of one of your Claude desktop app sessions using the new Compliance API (if it’s enabled for you).
Compare what the transcript shows against your own notes or email draft—does it actually contain everything a reviewer would need?
Pick a session in the Claude desktop app from last week that you believed was private.
If on an Enterprise plan, use the Compliance API to retrieve the session transcript, or ask an admin to extract it.
Lay the transcript next to your final file or decision to see what a third party would understand from it alone.
Check if the transcript reflects your actual intent, agreement or next steps—not just what was typed.
The Human Lens · what stays yours
A transcript or automation log feels official until it’s missing the crucial detail—the late change, the offhand clarification in a call, or the reason you chose that column in the spreadsheet. You still need to bring your own context and judgement. The facts might be pulled for review, but only you know what turned a rough draft or email chain into a final answer.
One question
See you next Monday.
Bring it back to your work
Bit starts from this issue, then asks you enough about your own work to make it worth acting on. If the honest answer is that none of this touches your job yet, it will tell you that too.