_edited.png){kind=small}
Prompting Is Dead: How to Build an AI Loop That Works Without You
- Ben Angel
- 4 minutes ago
- 13 min read
Most solopreneurs use AI the same way they used Google: type something into a box, wait for an answer, decide what to ask next.
Learning how to build an AI loop changes that relationship. Instead of deciding every next step, checking every answer and explaining every mistake, you give AI a goal, a quality test, the context it can use and a boundary that tells it when to stop or return to you.
The result is not a magical business that runs itself. It is something far more useful: recurring work that can research, draft, check and improve without trapping you inside every tiny correction.
Quick answer: What is an AI loop? An AI loop is a repeatable workflow in which AI completes a task, checks the result against a defined standard, improves weak work and repeats until it meets the goal, reaches a limit or needs human approval. Prompting asks AI for one response; looping gives it a job with a definition of done.
That distinction matters because speed is not the same as leverage. A fast answer still leaves you operating the machine. A well-designed loop makes your standards reusable while keeping consequential decisions—publishing, sending, spending, deleting and committing—under human control.
In This Article
What Is an AI Loop—and What Does It Achieve?
An AI loop performs a task, observes the result, compares it with a success standard and decides whether to improve, repeat, stop or ask a person for help. The easiest way to remember the difference is this:
Prompting gives AI a task. Looping gives AI a job.
Imagine you want five strong email subject lines. With prompting, you ask for ideas, reject the generic ones, explain why they are weak and request another batch. You may repeat that cycle four or five times before finding something worth testing.
With a loop, you define the assignment differently:
Generate 20 subject lines.
Score each against curiosity, clarity and click-intent criteria.
Reject anything generic, misleading or too similar to previous campaigns.
Rewrite the weak options.
Repeat for no more than five rounds.
Return only the five strongest choices with the reasoning behind each one.
The AI is still using prompts internally. What disappears is the need for you to hand-crank every revision. The work becomes more complete because it stops against a definition of done. Approved examples, exclusions and scorecards make your standards repeatable. Your time shifts from correcting individual outputs to improving the system that produces them.
Looping does not automatically make AI smarter or more truthful. It makes the workflow persistent. If the objective, sources or quality test are weak, the loop can produce bad work more efficiently. The value comes from combining iteration with judgment, not from repetition alone.
Where Did AI Loops Come From?
No single person invented the AI loop. It developed through several strands of AI research and practical experimentation.
In 2022, the ReAct research paper demonstrated a pattern in which a language model alternates between reasoning, taking an action and observing what happened. That action-and-observation cycle became part of the foundation for modern AI agents.
In 2023, Reflexion explored how an agent could turn feedback and failures into written reflections that influence later attempts. Self-Refine tested a similar generate, critique and revise pattern, reporting stronger results than one-step generation across its evaluated tasks.
In July 2025, Geoffrey Huntley made the pattern tangible with the Ralph Wiggum technique, a simple system that repeatedly sent an AI coding agent back to a persistent project plan. It showed what could happen when an agent continued working toward a completion condition instead of waiting for the next human prompt.
The idea entered the wider AI conversation when Boris Cherny, creator and head of Claude Code, described moving from writing code, to prompting agents, to running loops that prompt the agents. His line—“My job is to write loops”—captured the shift in five words. The original statement was shared publicly on Cherny’s X account, while The New Stack’s reporting on loop engineering documented how the idea spread.
HubSpot co-founder Dharmesh Shah then translated the concept for non-developers. In Learning to Write Your First AI Loop, he described three essential ingredients: an objective, a metric and a boundary.
That is the version that matters to a solopreneur. You do not need to understand the software architecture beneath an agent. You need to define the outcome, teach the system to recognize acceptable work and decide where it must return to you.
How to Build an AI Loop With the GOAL Framework
To make Shah’s three-part model easier to apply inside a small business, I expand it into four questions. Together, they form the GOAL framework.
G — Goal: What completed outcome should exist?
“Research potential customers” is an activity. “Create a reviewed list of 50 US companies that match my ideal-customer profile” is an outcome.
A strong goal names the deliverable, the audience and the business purpose: “Prepare a weekly decision brief showing which landing page deserves optimization first.”
O — Objective Test: How will AI recognize acceptable work?
For lead research, you might require every company to have 10–50 employees, operate in the United States, serve a named industry and include a website, founder and visible reason it matches your offer.
If you cannot describe what “good” means, AI cannot reliably apply it. Your test can be a checklist, scoring rubric, approved example, required data field or measurable threshold. The more subjective the task, the more examples matter.
A — Assets: What context may AI use?
Useful assets include:
your ideal-customer profile
examples of previous winning customers or campaigns
brand, editorial and compliance rules
approved websites or primary research sources
spreadsheets, presentations and project files
connected tools such as Google Drive, Slack or your CRM
L — Limits: When must the loop stop or return to you?
Set limits around attempts, time, cost, uncertainty and external actions. For example:
Make no more than five research passes.
Stop if fewer than 50 companies can be verified.
Flag uncertain entries instead of guessing.
Do not contact anyone or update the CRM.
Return the spreadsheet for approval.
The loop can do repetitive work, but the limits determine its authority.
A five-step implementation rule
Use this sequence when you build your first loop:
Choose a recurring task you already understand.
Write the GOAL instructions and define what “done” means.
Run it manually and inspect the failure points.
Add examples, exclusions and stronger stop rules.
Schedule it only after the output is consistently reviewable.
Begin where the inputs are visible, the result is easy to check and a mistake is reversible. The goal of the first loop is learning where your judgment must be made explicit—not achieving maximum automation.
A Real AI Loop for Weekly Content Opportunities
Suppose you publish weekly content about a fast-moving market. Your routine may involve opening dozens of tabs, checking dates, rejecting recycled news and choosing a topic when you are already tired of looking.
That is a strong candidate for a loop because the work is repetitive, source-based and easy to review before anything goes public. After writing eight books and building a content business across articles, email, video and courses, I know the expensive part is rarely the first draft. It is choosing the argument, protecting the voice and deciding what is worth saying. In my own content operation, I use this separation deliberately: AI can search, compare, score and prepare the decision, but the final editorial angle still comes back to me.
The assignment could read like this:
Every Monday, review reliable X news published during the previous seven days. Find stories with a material consequence for solopreneurs. Reject stories that are speculative, repetitive, more than seven days old or unrelated to business owners. Score the remaining stories from 1–10 for urgency, practical consequence, source quality and fit with my expertise. Continue researching until five ideas score at least 8/10 or you have checked all approved sources. For each winning idea, prepare a headline, a one-paragraph angle and links to the primary sources. Do not draft or publish the article. Return the five opportunities for my approval.
The AI does not merely find five stories and stop. It evaluates whether each deserves your attention, removes weak ideas, follows a completion rule and respects a publishing boundary.
How to judge whether a task deserves a loop
Build a loop when most of these conditions are true:
The task recurs weekly, daily or whenever a predictable event occurs.
The inputs can be found in approved files, tools or sources.
Acceptable work can be described or scored.
Several attempts may be needed before the result is useful.
The output can be reviewed before it causes an external consequence.
Repeating the task currently drains your attention more than your expertise.
Avoid looping a task merely because it is annoying. First ask whether it is the bottleneck. My guide to the five AI automation tasks a small business should tackle first provides a practical audit for separating visible busywork from work that is actually delaying revenue or decisions.
The payoff is decision-ready work. Instead of arriving at Monday morning with 40 tabs and a vague sense that something changed, you receive a short, sourced list that is ready for your judgment.
How to Set Up an AI Loop in ChatGPT Work or Claude Cowork
You can create a useful loop without programming. ChatGPT Work and Claude Cowork both support multi-step assignments, project context and recurring work, although their availability and handling of local files differ.
Set up the loop in ChatGPT Work
OpenAI’s current Work guidance describes Work as the mode for longer research and finished materials, while Codex remains focused on software development. OpenAI says the feature is rolling out gradually, so it may not yet appear for every account. Its current pricing comparison lists limited desktop Work access on Free and expanded desktop, web and mobile access on Plus and Pro.
Use this practical setup:
Open Work and create a project or task for the recurring workflow.
Add only the files, examples, instructions and connected sources it needs.
Paste the GOAL assignment and ask Work to explain its proposed process before running.
Run the task manually and check the evidence, scoring and stop condition.
Correct vague criteria and add one or two approved examples.
Schedule the stable workflow, choosing whether each run should be independent or return to the existing task.
OpenAI says a scheduled task can start a fresh task for each independent run or return to an existing task when the workflow needs its accumulated context. Its documentation also recommends testing the prompt manually and reviewing the first few runs before relying on the schedule.
If the task requires a folder on your computer, use the desktop app and keep the computer and app available when the schedule runs. Web tasks can use uploaded material and connected tools, but they cannot directly work inside a local folder.
For a deeper look at what happens when ChatGPT moves beyond responding and begins taking multi-step action, see my guide to using ChatGPT Agent for real solopreneur workflows.
Set up the loop in Claude Cowork
Anthropic’s Cowork documentation says Cowork handles complex multi-step tasks without requiring a terminal. It is available on paid Claude plans and across desktop, web and mobile, although Anthropic describes the web and mobile rollout as beta.
The setup follows the same operating logic: create a Cowork project, add only the permitted sources and connectors, describe the outcome with GOAL, run it manually, strengthen weak evidence rules, then use `/schedule` or the Scheduled page once the work is reliable.
Anthropic says remote scheduled tasks can continue when your computer is asleep. However, if the task requires a local file, browser or app on your computer, the desktop app must be open and able to reach it. That distinction is easy to miss and worth checking before you depend on a recurring run.
Which platform should a solopreneur choose?
Choose the environment where the work and context already live, then run the same low-risk loop in both platforms and compare:
Does it follow the source restrictions?
Does it apply the scorecard consistently?
Can you understand why it rejected an option?
Does it stop at the approval boundary?
How much supervision and paid usage does a complete run consume?
The winner is the platform that produces dependable, reviewable work inside your real business—not the one with the longest feature list.
I guarantee you, my personal recommendation may surprise you.
What Should Never Be Left to an AI Loop?
The purpose of a loop is to remove repetitive supervision, not human accountability.
Good candidates include researching approved sources, sorting information, drafting, checking against a scorecard and preparing recommendations. These activities are reversible and reviewable.
Keep explicit approval in front of actions such as:
publishing public content
sending emails or direct messages
purchasing software, inventory or advertising
deleting or replacing important files
changing customer records or CRM stages
making legal, financial, medical or employment decisions
Less obvious risks matter too. A loop can optimize the wrong metric, repeat a bad assumption or consume far more paid usage than expected. Asking the same model to create and judge its own work can improve consistency, but it is not independent proof. Important claims still need primary sources, customer evidence, analytics or a human reviewer.
Use a red-yellow-green authority rule
A simple authority map prevents accidental overreach:
Green: AI may research, sort, draft, compare and recommend.
Yellow: AI may prepare a change, but a person must approve it before the system updates, sends or publishes anything.
Red: AI may not perform the action, even after completing preparatory work, unless a specifically authorized human takes over.
This is the principle I return to in The Wolf Is at the Door: as AI takes over more execution, human judgment becomes more valuable, not less. The owners who benefit will not be those who surrender the most control. They will be the ones who make their judgment clear enough to scale while knowing which decisions should never be delegated.
The safe rule is simple: loop the repetition; keep the consequences human.
Frequently Asked Questions
Are AI loops the same as automation?
No. Traditional automation follows predetermined rules, while an AI loop can interpret information, evaluate a result and choose the next step. That flexibility makes its success criteria, source rules and limits especially important.
What is the difference between prompting and looping?
Prompting asks AI for a response; looping gives AI a repeatable job with a quality test and stopping condition. A prompt may be one step inside a loop, but the owner no longer needs to manually request every revision.
Is an AI loop the same as an AI agent?
No. An AI agent is the system that can reason, use tools and take actions; a loop is the cycle that tells the system to act, observe, evaluate and repeat. An agent can run inside a loop, while a simple loop may use only one model and a checklist.
Is prompt engineering really dead?
No. “Prompting is dead” describes a change in operating style, not the literal disappearance of prompts. Good instructions still matter, but your attention moves from crafting one clever request to designing the complete workflow, evidence rules and approval boundary.
Do I need to know how to code?
No. ChatGPT Work and Claude Cowork let you describe multi-step and recurring workflows in ordinary language. Coding becomes useful when you need custom integrations or advanced control, but it is not required for the research, reporting and content examples in this guide.
How much does it cost to run an AI loop?
The cost depends on your subscription, task length, frequency and how many attempts the loop makes. OpenAI says Work uses the same general usage structure as Codex and that consumption varies by task. Anthropic currently includes Cowork in paid Claude plans; its US Pro price is listed at $20 monthly or $17 per month with annual billing, while heavier Max plans begin at $100 per month. Pricing and availability can change, so confirm the current ChatGPT plans or Claude plans before choosing.
Is ChatGPT Work available to everyone?
Not in the same form. OpenAI says ChatGPT Work is still rolling out, while its pricing page currently lists limited desktop access on Free and expanded desktop, web and mobile access on Plus and Pro. If Work does not appear in your mode selector, your account, plan, workspace or rollout region may not have access yet.
Is Claude Cowork available on the free plan?
No. Anthropic currently lists Claude Cowork for paid Pro, Max, Team and Enterprise plans. Availability can also vary by surface and rollout stage, so check Anthropic’s current Cowork documentation for your account.
What is the best first AI loop for a solopreneur?
Start with a repetitive, low-risk task you already understand, such as weekly research, content screening, meeting follow-up, reporting or first-pass lead qualification. The output should be easy to inspect before it affects a customer, your reputation or your money.
How does an AI loop learn?
Most business loops do not retrain the underlying model. They carry feedback, approved examples, results, saved instructions or project context into the next attempt so the workflow can apply what happened previously.
How many times should an AI loop repeat?
There is no universal number. Three to five passes may suit content, while research might stop after verifying a defined number of records or exhausting approved sources. Every loop should have a cap so it cannot repeat indefinitely or burn through paid usage.
Can an AI loop hallucinate or make mistakes?
Yes. Repetition does not eliminate hallucinations and can reinforce an incorrect assumption. Require primary sources for factual claims, flag uncertainty, use external evidence for important decisions and keep a human approval gate before consequential actions.
Is it safe to connect an AI loop to email, files or a CRM?
It can be, but access should be limited to what the task genuinely needs. Begin with read-only or preparation work, test the workflow manually, restrict the files and connectors it can reach and require approval before sending messages, changing records or deleting anything.
When should I stop using a loop and do the work myself?
Take over when the outcome is highly subjective, the evidence is weak, the task keeps failing the same quality test or the decision carries legal, financial, reputational or personal consequences. A loop should reduce cognitive drag; if supervising it costs more attention than completing the task, the design—or the use case—is wrong.
What is the biggest mistake when building an AI loop?
The biggest mistake is giving AI an activity without a checkable outcome. “Make this better” creates endless subjective revision. “Meet these five standards in no more than four passes” creates a usable operating rule.
The promise of AI loops is not that your business runs without you. It is that your expertise stops being trapped inside thousands of tiny corrections. AI can handle more searching, sorting, drafting and checking while you remain responsible for direction and consequences.
That is the real shift from prompting to looping: your judgment becomes a system that can keep working even when your attention has moved somewhere else.
And for a solopreneur, attention is the resource worth protecting.
If you want to turn this approach into a practical operating habit, the 28-Day AI Mastery Challenge provides a guided implementation path. It is designed to move you beyond collecting prompts and toward building repeatable AI workflows around the work your business actually needs.