Stop prompting, start looping

You're at A: you prompt an AI, read what comes back, fix it, and prompt again, and you're the one checking every answer. For a sales manager at about $106 an hour, every weekly hour spent like that costs about $5,512 a year. By the end, you'll be at B: you'll be able to set up a simple loop where something other than you does the checking, so you only look at finished work. You won't need any code.

Here's the whole issue in one sentence: the expensive part of AI isn't the model, it's the person checking the model's work. So the fix isn't a better prompt. It's moving the checking off the person.

Here's how most teams use AI. Someone types a prompt, reads what comes back, spots what's wrong, fixes the prompt, and goes again. That loop is expensive because every round needs a person to read and judge, and an hour of a person costs far more than an hour of tokens (the small chunks of text AI companies bill by).

Picture a sales manager who spends Friday afternoon turning CRM notes (from the customer database sales teams use) into a pipeline summary for leadership. With a chatbot, the job gets a little faster, but it's still theirs: paste, read, correct, paste again. That person costs about $106 an hour, fully loaded, based on Bureau of Labor Statistics figures for sales managers. So every hour a week they spend in that loop costs about $106 × 52 = $5,512 a year, because it happens every week of the year.

Here's the same point from another angle. The chatbot didn't remove the work. It turned writing into proofreading, and you're still paying a manager's rate for every round of it.

Anthropic built this move into Claude Code, its AI assistant that can work through a multi-step job on its own. The feature is called a goal. Jordan Crawford's walkthrough is what put it on our radar. You describe the finish line, and after every round a separate AI model checks whether it's been met. Anthropic's documentation says the decision comes from "a fresh model rather than the one doing the work." You can cap it with a limit like "or stop after 20 turns," and Claude Code stops the loop if Claude stops making progress. The checker matters because judging no longer waits on a person. The cap matters because a loop with no limit can keep running and keep spending.

Claude Code's lead designer, Meaghan Choi, shows something similar on her team. Several Claudes work in parallel, each one is told to loop until done, and Claude checks its own interface changes in the browser before a person looks.

Different tools, same move: the person stops being the checker and becomes the approver.

On October 2, I set up a content team of AI agents to run this newsletter, with research, curation, writing, and editing each owned by one agent. The research agent pulled two weeks of AI launches. By YouTube's public view counts, which it pulled on Oct 2, Claude's top new demo videos were drawing between 35,000 and 87,000 views a day, roughly 10 times what that channel's older videos typically get. The curation agent scored a test batch of my saved posts and surfaced the ideas behind this issue. The editor picked the topic and briefed the writer.

Then the issue itself ran through the loop. The writer agent drafted it, and separate agents checked it against the sources. The research agent flagged four data fixes, including a view-count range that mixed in the wrong videos and a cost claim the numbers didn't support. The curation agent caught places where the draft credited Jordan Crawford and Meaghan Choi with more than they actually said. The editor sent it back until every check passed. What you're reading is the fourth draft.

Granted, a checker can't judge everything. Meaghan Choi, Claude Code's lead designer, says Claude still isn't good at design, so people keep the taste calls. However, that doesn't put the person back in the checking seat. It moves them to the approval step, where taste calls belong. None of the fact-checking on this issue was mine. My part was approving it, and that's the whole idea one more time: something else checks, and the person approves.

Anthropic's Opus 5.5, released September 22, costs $4 per million input tokens and $20 per million output, down from $5 and $25 for Opus 5. That's 20% cheaper per token.

Cheaper per token doesn't automatically mean cheaper per task, though. Opus 5.5 always thinks before it answers, and a loop runs more rounds than a single prompt, so one task can burn more tokens than it used to. I'm logging what our loops actually cost, and I'll publish the real number once it's tallied, even if it looks bad.

You're at B if you can now take one recurring task, write its finish line as three yes/no questions, and run it through a second chat until every answer is yes.

Next week: why the next big software companies will sell the finished work, not the tool.

P.S. If you'd like a loop like this set up for your revenue team, just reply to this email and tell me which task eats your Fridays.