Boris Cherny is the man behind Claude Code.

At Acquired Unplugged earlier this month, he described his workflow, and 1 line that I want to talk about today:

"I don't prompt Claude anymore... my job is to write loops."

A year ago he was writing code by hand with autocomplete, and soon after he was running 5-10 claude sessions at once, prompting each one himself.

Now he prompts none of them, and a few hundred agents watch his GitHub, Slack and Twitter, work out what's worth building, and hand the prompting to the loops.

The phrase caught on fast. Addy Osmani, an engineering director at Google, named it loop engineering, and Peter Steinberger too pushed the same idea.

For 2 years everyone raved about prompt engineering. You typed, read the diff, typed again, and held the tool the whole time.

The leverage point moved. Now it's the system that decides what the agent works on, when, with what gate, and what survives between runs.

You'll have seen the stat that Anthropic engineers merge 8x as much code per day as in 2024. Anthropic itself calls that almost certainly an overstatement.

The number is noise but the mechanism under it is not.

None of which means you should go build one. That is what the hype skips, and where this edition starts.

Before We Begin

Hive: An open-source, multi-agent pipeline that carries a one-line idea through brainstorm, plan, execute, and review stages and ships a merge-ready PR. It runs tasks in parallel in the background on agent CLIs like claude and codex, and the only thing it asks of you is answers to its questions in a doc.

Baroque: An AI UI design generator that produces interface designs from curated, master-grade templates rather than a blank prompt box. The templates are the starting point, so the output leans on proven layouts instead of whatever a model improvises from scratch.

Conan: A native macOS app that wraps Claude Code in a live HUD, streaming every prompt, tool call, skill, and token as the session runs. It surfaces the context window filling up and the cost of each move in real time, and it's a one-time $29 purchase with no subscription.

Loops earn their cost under 4 conditions, and missing 1 means the loop costs more than it returns.

The task repeats at least weekly, and verification is automated, meaning a test, type check, or build can fail bad output without you in the room.

Your token budget can absorb retries and re-reads, and the agent can run its own code and see what breaks.

Fail one and stop here. Keeping it a manual prompt is the right answer, because most developers don't need a loop yet.

The demo below builds a dependency-bump loop that passes all four: dependencies update weekly, the test suite is the gate, the work is mechanical, and the agent runs the suite itself.

That makes it one of the few loops safe to build first.

Do not automate anything yet. Open Claude Code in your repo and run the task once by hand.

Watch it work. If the manual run is shaky, the loop will be shaky with nobody watching. Fix the prompt until one clean run happens end to end.

Caveat: pin it to non-major bumps. Major versions carry breaking changes that a passing test suite won't always catch, and that's judgment work a loop shouldn't own.

The agent re-derives your project context from zero every cycle unless you write it down. Create the skill folder.

Then .claude/skills/dep-bump/SKILL.md:

Intent now compounds. Every run reads the same rules instead of you re-typing them.

The agent forgets, but the repo does not. Create STATE.md at the root:

Now tomorrow's run resumes instead of restarting. The agent reads what it already handled and what it must leave alone.

This is what decides whether the loop helps or just spends. The gate is objective: a test that passes or fails, not a second agent with an opinion.

For this loop the gate is already there: npm test returns zero or non-zero. Make it strict.

If your suite has flaky tests, the loop will revert good bumps on false failures. Fix flakes before you automate, or the gate lies to you.

Add a human approval gate too. The loop opens a PR but does not merge.

Only now do you automate. In Claude Code:

/goal runs until your condition is true, and a separate small model checks completion so the agent that wrote the change isn't the one grading it.

To make it recurring, move it to a Desktop scheduled task or run it on Replicas so it survives your laptop closing.

The order: manual run > skill > state > gate > loop > schedule.

This is the Ralph Wiggum loop, named by Geoffrey Huntley. An agent meant to signal "done" only when finished signals it early, and the loop exits on half-done work.

Without a hard gate it fails and keeps spending.

The metric that tells you the loop is working is cost per accepted change.

If fewer than half the PRs it opens get merged, you're doing review work the loop was supposed to remove, and the loop is losing.

I almost built a CI triage loop last month, a nightly job to draft and fix PRs for the easy failures. The idea looked clean until I ran the four-condition test on myself and failed two.

My token budget can't absorb a loop re-reading context on every retry, and half my repos don't have a test suite good enough to fail bad output.

It would have been two optimists agreeing on repeat while the bill climbed.

Reading the research changed my mind.

It holds two messages at once: that you can build fleets of agents, and that most people should keep prompting by hand for now.

Only one of them goes viral.

What scares me more than the token bill is comprehension debt. 

The faster a loop ships code you didn't write, the wider the gap between what your repo contains and what you understand. 

The bill that hurts arrives the day you debug a system nobody on the team has read.

So if you pass the test, build the smallest thing. Dependency bumps before CI triage, and anything with real judgment in it comes last.

Get the manual run reliable first.

Until next time,
Vaibhav 🤝🏻

If you read till here, you might find this interesting

#Partner 1

TLDR is the free daily email that does the reading for you. Each issue summarizes the most interesting stories in startups, tech, and programming, curated by ex-Google and Anthropic engineers, so you get the signal without the noise.

It's one email instead of 12 open tabs. A 5-minute read with your morning coffee, and you walk in already caught up.

Tech is just the start. We also cover AI, marketing, dev, and more, so you can follow the topics that matter to your work.

Free, daily, and read by 7M+ subscribers. Subscribe and get your mornings back.

Subscribe for Free

#Partner 2

Rolling out your first HR tool? Get a step-by-step guide to avoid common mistakes, drive adoption, and build a scalable HR foundation.

Download the guide