Stop Prompting Claude. Build Loops Instead.
In 7 days, four top AI engineers said the same thing. This article breaks down what they mean and gives you a no-code system to start today.
On June 7th, a developer named Peter Steinberger posted two sentences on X.
“Here’s your monthly reminder that you shouldn’t be prompting coding agents anymore. You should be designing loops that prompt your agents.”
8 million people read it.
Then the head of Claude Code at Anthropic confirmed it from the inside:
“I don’t prompt Claude anymore. I have loops running that prompt Claude. My job is to write loops.”
Then, Addy Osmani, Director of Google Cloud AI, wrote 1,300 words breaking it down. 1.3 million views in 24 hours.
Four days before any of this, an engineer at Anthropic had already shipped the feature that made it all possible.
Four different people. One signal.
The way you work with Claude right now is already outdated.
You’re the reason this newsletter exists. What AI tool or workflow should I test next? Drop your topic in the comments - the most-requested one becomes the next issue.
By the end of this issue, you'll get a copy-paste Claude Loop Starter Kit, 5 prompts that build your first personal Claude system, zero code required.
I’ll be honest.
When I first saw that two-sentence post hit 8 million views, my reaction was: “That’s for engineers. Not for me.”
Then I sat with it for two hours.
Then I built something small, a Claude Project that knows my newsletter context, runs my weekly content research, and hands me a first draft before I even open a blank page.
I’ve been prompting Claude like a vending machine.
Put something in, get something out, repeat.
The people who built Claude are running it like a factory floor. The machine runs. You supervise.
That reframe changed how I work this week.
This didn’t start with a tweet. It started with Anthropic shipping something.
June 3: Thariq Shihipar, technical staff at Anthropic, Claude Code team, publishes the Dynamic Workflows feature. 2.9 million views. The post opens with example prompts like:
“Use a workflow to dig through #incidents in Slack for the past six months and find recurring root causes where nobody has filed a ticket.”
Not a coding task. A thinking task. Automated.
June 7: Peter Steinberger posts his now-famous two sentences.
He calls it a “monthly reminder”, meaning he’s been saying this for a while, and people still weren’t listening. 8 million views. 14K saves.
June 9: Boris Cherny, Head of Claude Code at Anthropic, confirms it from inside:
“I don’t prompt Claude anymore. I have loops running that prompt Claude and figure out what to do. My job is to write loops.”
The same day, Addy Osmani publishes his 1,300-word breakdown. 1.3 million views.
Four sources. Seven days. One consensus: the leverage point has moved.
"Four people at the top of the AI world said the same thing in one week. That's not a trend. That's a signal."
If this is useful, share it with one person who's still copy-pasting prompts every morning. Takes 10 seconds.
WHY SOLO PROMPTING BREAKS DOWN
Before we talk about the fix, we need to understand why prompting alone fails at scale.
Three failure modes happen when you rely on a single Claude session for complex work. Once you see them, you can’t unsee them.
Failure Mode 1: Agentic Laziness
Claude stops before finishing and declares the job done. You asked for a 50-item security review. It did 20, then told you it was complete. You didn’t notice.
Failure Mode 2: Self-Preferential Bias
Claude grades its own homework. You asked it to verify its output. It agreed with itself. Of course it did.
Failure Mode 3: Goal Drift
Across a long session, the original goal quietly shifts. Edge cases get dropped. “Never do X” constraints disappear after a few summarization steps. Claude is still working, just not on what you originally asked.
These aren’t bugs.
They’re structural limits of a single context window, trying to both plan and execute at the same time.
A loop fixes this by giving each task its own Claude with its own context window and a single focused goal.
The maker doesn’t grade the maker’s work. A second Claude does.
If this is useful, share it with one person who needs it. Takes 10 seconds.
OLD WAY → NEW WAY CONTRAST
The old way:
You open Claude. You think of what you need. You write a prompt. You read the output. You copy something. You close the tab. Tomorrow, you repeat everything from scratch. Claude has no memory of your project, your goals, or your voice.
Old workflow:
Open → Think → Type → Read → Copy → Close → Forget → Repeat.Every session starts cold. You are the system. You are also the bottleneck.
The new way:
A system holds your context, triggers the right prompts at the right time, and hands you the output. One Claude does the work. A second Claude checks it. You review what passes.
New workflow:
Build once → System prompts → Claude works → Second Claude verifies → You review.
You go from being the person who prompts to being the person who designed what gets prompted, and who checks it.
WHAT A LOOP ACTUALLY IS (NO CODE VERSION)
There are many difficult terms like worktrees, JavaScript files, and sub-agent spawning. Ignore all of that for now.
Here’s what a loop is in plain language:
A set of instructions that runs automatically, does useful work, and hands you the result.
That’s it.
For engineers, it’s code. For you and me, people using Claude daily without writing a single line of JavaScript, a loop is a Claude Project that knows your context, runs your recurring tasks, and uses a second pass to check its own work.
Think of it like this:
You don’t build a new kitchen every time you cook.
You build it once. Then you cook in it.
A Claude loop is your kitchen.
If this is useful, share it with one person who needs it. Takes 10 seconds.
THE 6 LOOP PATTERNS (TRANSLATED FOR NON-ENGINEERS)
Pattern 1: Fan-Out and Synthesize
Split a big task into smaller pieces. Run Claude on each piece separately. Combine the results.
Real example for you: Ask Claude to research 5 different content angles separately, then synthesize the best one. Don’t ask it to do all 5 in one prompt; quality degrades.
Pattern 2: Adversarial Verification
After Claude produces something, a second prompt reviews it from the opposite direction, looking for flaws, not validating it.
Real example for you: After Claude writes your newsletter draft, run this: “You are a skeptical editor. Find every weak claim, vague sentence, and missed opportunity in this draft.” Different role, different output.
Pattern 3: Tournament
Generate multiple versions of something. Have Claude judge them against a rubric. Pick the winner.
Real example for you: Ask Claude to write 5 hooks for your next post. Then ask: “Judge these 5 hooks against this criteria: curiosity gap, identity signal, share potential. Rank them 1–5 with one line of reasoning each.”
Pattern 4: Loop Until Done
Instead of a fixed number of tries, keep going until a real condition is met.
Real example for you: “Keep improving this draft until there are zero filler sentences and every claim has a specific number or example backing it. Don’t stop until that condition is true.”
Pattern 5: Generate and Filter
Produce a large batch of ideas. Filter ruthlessly by a rubric. Only keep the best.
Real example for you: “Generate 20 newsletter headline options. Then filter: keep only the ones that use accusation, desire, or paradigm-shift framing. Delete the rest.”
Pattern 6: Deep Verification
Have one Claude identify all the claims in your content. Have a second, Claude, verify each one.
Real example for you: “Read this draft. List every factual claim I’m making. Then verify each one, flag anything vague, unsubstantiated, or potentially wrong.”
If this is useful, share it with one person who needs it. Takes 10 seconds.
Before you hit the prompts, stop here for a second.
I know what you’re thinking. “This is for engineers. I’m not building worktrees and spawning sub-agents.”
You’re right that you’re not doing that. But here’s what you ARE doing every week: opening Claude fresh, typing the same kind of prompt you typed last week, getting a similar output, and moving on. No memory. No improvement. No system.
Your version doesn’t need JavaScript. It needs a Claude Project, a memory file, and three recurring prompts with a review step. That’s the loop. That’s enough to start.
Build that this week. Everything else compounds from there.
STEP-BY-STEP: BUILD YOUR FIRST CLAUDE LOOP
Step 1: Build Your Memory File
This is the brain of your loop. It tells Claude who you are, so every session starts informed, not cold.
Quick setup:
Open Claude → Projects → Create new project
In Project Instructions, paste your context: your name, what you create, your audience, your tone, and your recurring goals
Add this line: “Every time I open this project, confirm you understand my context. Then ask: what are we working on today?”
"This one file is what separates a loop from a one-off prompt; the loop remembers, the prompt forgets.
Step 2: Write Your Trigger Prompt/Skill
A trigger starts every session. It activates your loop before you’ve typed a single word.
Quick setup:
Write a master prompt that opens with: “You are my [content/research/work] assistant. Every session begins by reviewing [X] and producing [Y].”
Save it pinned inside your Project Instructions
Test it: close the project, reopen it, see what Claude does first
"The trigger is the 'good morning' you teach Claude to say to itself, so you never start from scratch."
If this is useful, share it with one person who needs it. Takes 10 seconds.
Step 3: Add Adversarial Verification
This is the one step most people skip. It’s also the most important.
After your loop produces output, run a second prompt that checks it from the opposite direction.
Quick setup:
After any major output, paste this: “You are a skeptical editor. Your job is not to validate this. Your job is to find every weak claim, vague sentence, and missed opportunity.”
Use a different role framing than the first prompt, which forces a different perspective
Combine the outputs: keep what survived both passes
"The maker Claude and the checker Claude should never be on the same prompt. One writes. One attacks."
Step 4: Build Your Prompt Library
Your three to five most recurring tasks become saved prompts.
This is your loop’s engine.
Quick setup:
List your top recurring Claude tasks (content research, draft writing, weekly review, email responses, idea generation)
Write one optimized prompt or update your skill for each, with a real context inside, no placeholders
Save them in a Notion page or Claude Project doc labeled “Loop Prompts.”
Note: Use the best outputs (that performed so well as a reference/ context file for your project or skill)
If this is useful, share it with one person who needs it. Takes 10 seconds.
Step 5: Run a Weekly Loop Review
The loop gets smarter every week, but only if you spend 10 minutes reviewing it.
Quick setup:
Every Sunday, open Claude and paste: “Review what we worked on this week. What produced strong output? What was weak? Suggest one update to my Project Instructions based on what you noticed.”
Apply the suggested update
Delete any loop prompts that didn’t save you time this week
"Skip this step, and the loop stagnates. Do it and the loop compounds. Ten minutes a week."
Your gift - as promised:
The 5 prompts below are your Claude Loop Starter Kit. Copy them into a Claude Project today. Replace the bracketed context with your own work. Your loop starts running this week.
Start with the Memory Installer. Then add the Adversarial Editor. Those two alone will change your output quality before the weekend.
The Tournament Hook Generator is your weekly secret weapon. Run it every time you need an opening line and can’t find one.
Prompt 1 - The Memory Installer
Paste this into Project Instructions. This is your loop’s brain.
You are my personal AI assistant. Here's what you always know about me:
Name: Hamza
What I do: I write the AI In Public newsletter on Substack. I create Claude guides,
tutorials, and AI workflow systems for ambitious builders and career-changers.
My goals right now: Grow newsletter to 10K subscribers. Build a content system that
runs with less manual effort each week.
My tone when writing: Direct. Honest. Practical. Never hype. No fluff.
Things I never do: Generic advice without a specific example. Vague prompts with
placeholder text. Long intros before the value.
Every time I open this project, start by confirming you understand my context.
Then ask: "What are we working on today?""The last two lines are the loop trigger; they make Claude active, not passive, from the first second."
Prompt 2 - The Adversarial Editor
Run this after any first draft. The most important prompt in the kit.
You are a skeptical editor reviewing a draft for AI In Public newsletter.
Your job is NOT to validate this draft. Your job is to find problems.
For every section, tell me:
1. What claim is weakest or most vague?
2. What sentence could be cut without losing value?
3. What's missing that a skeptical reader would ask for?
Do not soften your feedback. Be specific. Diagnose, don't summarize.
Here is the draft: [paste draft]“Diagnose, don’t summarize’, this one phrase forces Claude out of cheerleader mode and into editor mode.
Prompt 3 - The Tournament Hook Generator
Run this when you need a hook for any post, thread, or newsletter.
Write 5 different hooks for this content. Each hook must use a different format:
1. Accusation hook ("You're doing X wrong")
2. Paradigm shift hook ("The way X works just changed")
3. Personal proof hook ("I tested X for 30 days. Here's what happened")
4. Desire hook ("Imagine being able to X in under Y minutes")
5. Shock hook ("X% of people using Claude don't know this exists")
Topic: [your topic]
Audience: Ambitious builders and content creators who use Claude daily
After writing all 5, judge them on: curiosity gap, identity signal, share potential.
Rank 1–5. Explain the top pick in one sentence.“The ranking step at the end is what turns this from an idea list into a decision, don’t skip it.”
Prompt 4 - The Loop Until Done Refiner
Use this when a draft is close but not there yet.
You are a ruthless editor with one job: eliminate all filler.
Read this draft. Keep going until every single sentence meets ALL three criteria:
1. Contains a specific number, example, or named source — OR
2. Moves the reader to the next idea — OR
3. Creates a genuine emotional reaction
If a sentence fails all three tests, cut it. No exceptions.
Do not stop until the condition is true for every line.
Here is the draft: [paste draft]“’ Do not stop until the condition is true for every line. This is the loop trigger. It forces multiple passes.”
Prompt 5 - The Sunday Loop Review
Run every Sunday. Updates your system automatically.
Review our work together this week in this project.
For each major task:
1. What produced strong, usable output?
2. What was weak, vague, or required heavy editing?
3. What was the root cause — a bad prompt, missing context, or wrong framing?
Based on this review, suggest ONE specific update to my Project Instructions
that would improve next week's output.
Write the update as a ready-to-paste instruction, not a suggestion.
Be specific. If you can't point to a concrete improvement, say so.“’ Write the update as a ready-to-paste instruction’, this closes the loop. You paste it in, and the system evolves.”
If this is useful, share it with one person who needs it. Takes 10 seconds.
RADICAL HONESTY
Here’s the thing nobody says in the hype posts about this.
A loop running unattended is also a loop making mistakes unattended.
Loops are powerful, and they have a cost, token cost, comprehension cost, and trust cost if you stop reading what they produce. The faster the system works, the easier it is to stop understanding what it’s building.
Your job doesn’t disappear when you build a loop. The repetitive parts disappear. The judgment parts stay yours.
Build the loop. Stay the editor.
In 60 seconds:
Four engineers confirmed it in 7 days: the leverage point moved from prompting to building systems
Solo prompting fails from agentic laziness, self-preferential bias, and goal drift
A loop = Memory File + Trigger Prompt + Adversarial Check + Prompt Library + Weekly Review
Six loop patterns work without code: Fan-Out, Adversarial Verify, Tournament, Loop Until Done, Generate + Filter, Deep Verify
Start with Prompt 1 (Memory Installer) this week, and everything else builds from there
Save this. Screenshot it. Come back to it.
Here are a few issues that you might have missed (go check them out):
Clone Any Faceless YouTube Channel with Claude (Step-by-Step)
How to fix your Obsidian graveyard with Claude under 20 mins
I gave Claude Fable 5 a Month-Long Project - Here is what happened
15 Insane Things People Built With Claude Mythos in 48 Hours
We tested it in public. Now go use it in private.
If this issue helped you, forward it to one person who needs it. It’s free, and it takes 10 seconds.
Sharing is always caring :)
And for shorter takes between issues, follow me on X → Hamza Khalid
PS: Which of the 6 loop patterns fits your work the most right now?
Hit reply - I want to know which one you’re starting with.
- Hamza 💙
This is practical wisdom.
The best kind.