Claude Cowork: Your Best Session This Week Will Be Filed as Waste
Claude now writes reports on your work for your manager. The sessions where you did your hardest thinking are the ones most likely to land under “inefficiencies”. That was 1 of 7 changes around your agent in a single week.
On Tuesday you spend 40 minutes with Claude on a client proposal. You test 3 pricing structures. The first undercuts your margin. The second is too complex to explain in a meeting. The third fails the moment you imagine the client’s finance director reading it.
You close the session with nothing to save, open a blank document and write the proposal yourself in 25 minutes. It is the best proposal you have sent this quarter.
The 40 minutes were the reason the 25 minutes went well. You know that. The report your manager reads on Friday does not.
What your manager sees
On October 28, 2026, Anthropic made smart reports available to every Claude Enterprise organization. Claude reads a sample of a team’s sessions in chat, Claude Code and Cowork, and writes a report: workstreams, deliverables, spend, pain points.
And a section called “Inefficiencies”. It counts sessions Claude classified as personal or off-topic, and sessions that produced nothing usable. Both appear only as totals with a cost. Names are hidden by default, and the feature is off until an Owner turns it on.
The off-topic counter got the headlines. The other line matters more. Your Tuesday session ended with a judgment, “none of these will survive the finance director”, and that judgment never appeared on screen. From inside the transcript, it looks like 40 minutes of attempts followed by silence.
Principle 29, Invisible Work, says work done with an agent disappears from view. Smart reports bring it back into view through a classifier, and the classifier reads deliverables. Exploration has none.
There is a second detail. The report is written from a sample, by a model. A workstream that took 6 sessions of hard thinking and one that took 6 sessions of formatting can get the same paragraph. You did not write the description of your month, you did not see it before it was read, and you cannot correct it inside the report. Principle 14, Responsibility Without Control, now has a second layer.
The real risk comes next. Once people know a classifier is reading, they start writing for it. Exploratory sessions get shorter. Every session ends with a deliverable nobody needed. Principle 4, the Feedback Loop, depends on being allowed to try things that fail. A team that stops exploring will produce cleaner reports and weaker judgment.
The fix takes 30 seconds, and it is the free prompt at the end of this article.
6 more changes, 6 open questions
The same week, between October 28 and November 3, 6 other things changed around your agent. Each one moved a decision that used to be yours.
Security moved inside the agent’s machine. Anthropic opened EDR sensors inside Cowork’s virtual machine, and CrowdStrike, SentinelOne and Palo Alto Networks announced integrations. They catch the dangerous action. The open question: which mistakes in your work look completely normal to a security tool, and why do people approve faster once they know security is watching?
Get Mihailo Zoin’s stories in your inbox
Join Medium for free to get updates from this writer.
The engineer who leads Cowork said delete the expert persona. Felix Rieseberg called “you are an expert in X” a 2024 habit, and said some skills work better removed. The open question: which of the rules in your project instructions are fixes for a model you no longer use?
The same engineer dropped coding interviews for his team. Claude now beats the small tests. He looks for taste, empathy and principled thinking. The open question: if the agent can pass your profession’s test, how would you prove your own skill in your next review?
A security startup now protects agent rules from the people they protect. Neo restores hooks that users or agents delete. The open question: who switched off your last safety rule, and why is it probably still off?
Your absence now has a timer. Message threads make long sessions faster while you are present and expire after about 5 minutes idle. The open question: why does the decision you made at 10:20 disappear when you come back at 14:10?
The money went to watching agents. In February the market feared agents would replace software. By October, cybersecurity stocks had more than doubled from their April low. The open question: has your own job made the same move, and does anyone know?
Each question has a full answer in the library this week: a working scene you will recognize, the principles behind it, and prompts you can run in your next session.
Key insight: Execution keeps getting cheaper and more watched. The judgment around it, what counts as done, what the agent may touch, which rules hold at 11 p.m., keeps getting more valuable and stays hard to see. Make yours visible before a report describes it for you.
Cowork task
Use this at the end of any session that tested ideas and produced no file. At the end of your request, add:
“Before we close, write a session record of no more than 5 lines. Line 1: the question I was trying to answer. Line 2: the options we tried. Line 3: what I decided, including ‘none of these’ if that is the answer. Line 4: the reason, in my words from this conversation. Line 5: where the decision will be used. Label it ‘Exploration’ and save it as session-record-[date]-[topic].md in [folder].”
For the Tuesday session, the record would read: “Question: pricing structure for the client proposal. Tried: 3 structures. Decided: none of them. Reason: the finance director rejects anything that hides the margin. Used in: the proposal I wrote by hand.” 5 lines, and a session that looked like waste now shows exactly what it produced.
After a month, read the records in a row. You will see where your judgment actually lives.
For individuals: Write a session record for every exploratory session this week. It protects your thinking whether or not your organization ever turns reports on.
For teams: Agree on which session types look unproductive and are not: pricing tests, draft comparisons, red-team questions. Bring that list to your manager before the first report arrives.
For organizations: Decide in writing how the inefficiency section will be read, before anyone reads it. Measure exploration separately from waste.
This week in the library: 7 new examples, one for each change above, with 14 prompts, including the monthly self-report that turns your session records into an answer for the inefficiency line, the checkpoint list for actions security cannot judge, and the 3-step test that shows which of your instructions to delete.
This article is part of Claude Cowork: The Tool That Shapes the Master, a living library for professionals who work with AI agents daily. 30 principles. 100+ examples. 50 workflows and automations. 99 commands. More every week. Founding members get all 30 principles and the full example library, with new examples added every week from current practice.
