Practice 05: Drafted, Not Sent
Why this gets violated constantly
Practice five sounds obvious and gets violated constantly, because automation is seductive: if the agent can draft the email, why not let it send? Here's why not.
The agent emails a client report directly. One wrong number reaches the client. The five minutes of review that got saved gets spent many times over on damage control. There's no measured multiplier behind that; it's the plain shape of the trade: minutes saved on one end, hours of correcting and re-explaining on the other. Worse, trust (the client's and the organization's own confidence in the whole approach) resets to zero.
The fix: drafted, not sent
The fix costs almost nothing: design every workflow so its output arrives drafted-but-unsent. The report gets built overnight; the email sits in drafts by morning, reviewed over coffee, then sent. Nearly all of the work is automated; every bit of the accountability is kept exactly where it belongs. That's a description of the shape of the trade, not a measured split, since no one is timing what fraction of "the work" a report represents.
Where a human always signs
The agent does everything up to the irreversible step, then stops. Three kinds of output cross that line and always need a human signature before they leave:
Any email, report, or deliverable that leaves the building. Once it's out, it can't be unsent.
Payments, journal entries: anything that moves money or restates a number.
Tax, legal, regulatory, compliance: anything with an external authority on the other end.
Anthropic's own finance team is reported to put the underlying discipline bluntly: "you need to verify everything that comes out, and take ownership of it." Like the "roughly 150 skills" figure in the previous lesson, this circulates as a company anecdote rather than a dated, citable document. Hold it the same way: repeatable as a paraphrase of the discipline, not as a pinned, sourced quote. Automate the work. Never automate the accountability.
Demo: the monthly report that writes itself
The third worked example puts practices 04 and 05 together. The setup from Lesson 08 is already in place: a CLAUDE.md handbook with the company's standing rules, and a monthly-report skill defining the inputs, format, and checks for that recurring job.
With both of those already written and saved in the project folder, the entire prompt for this month's report, typed straight into the terminal exactly as shown, is one line:
claude
Then, at the prompt that opens:
Run the monthly report on the June files.
Everything else (which files, what format, which totals to check) lives in the handbook and the skill, not in the sentence typed that morning. That one sentence is the entire brief, because practice 04's files already carry the rest.
What comes back
What comes back is the report itself, variance commentary, and a stakeholder email, visibly marked drafted, not sent. The signature is still the last step, every month, no matter how short the prompt that got there. A paragraph of instructions the first time became a sentence the second time, because the instructions turned into infrastructure. That's what practices 04 and 05 look like once they're both running.
All five practices are in place. Next: why a working demo proves less than it seems, and what closes the gap to production.
Reply here and it goes straight to Rod. Same as replying to one of his emails.
