Prompts Are Dead, Build Loops: What AI Loops Are And How To Start Today
Anthropic and OpenAI have independently converged on the same advice: stop writing single prompts, start writing loops. A loop is a cycle where the model does the work, scores its own output against
June 11, 2026
Prompts Are Dead, Build Loops: What AI Loops Are And How To Start Today
TL;DR: Anthropic and OpenAI have independently converged on the same advice: stop writing single prompts, start writing loops. A loop is a cycle where the model does the work, scores its own output against a rubric, critiques it, rewrites the weak parts, and repeats until the result stops improving. It beats one-shot prompting because scoring and rewriting are separate cognitive acts, which turns blind editing into directed search. Fable 5’s long autonomous runs make loops practical now. The fastest way to start: install the free plan-optimizer skill, or paste a simple “score, critique, rewrite, repeat” instruction after any task.
What is an AI loop?
A prompt is one shot. Input goes in, output comes out, and you take whatever you get. A loop is a cycle: the model produces output, evaluates its own work against a standard, fixes the weak parts, and runs again until the result is right.
As Addy Osmani puts it in his writeup on loop engineering, you are “replacing yourself as the person who prompts the agent. You design the system that does it instead.” That is the whole shift. You stop hand-holding the model through every step and instead build a small system that prompts it for you.
The reason loops beat single prompts is mechanical, not magic. Scoring a piece of work and rewriting it use different cognitive muscles. When you force the model to critique its own output before it rewrites, you convert blind editing, which tends to drift sideways, into directed search that climbs toward a target. The loop keeps the best version it has seen and stops when the score plateaus.
Why loops matter right now
Two things happened at once.
First, the two most important AI labs converged. As one widely shared breakdown (1.49M views) put it: “Anthropic and OpenAI are both telling engineers to write loops. Not prompts. Not agents. Loops.” Both Claude Code and OpenAI Codex landed on the same agentic-loop pattern independently. When rivals that rarely agree arrive at the same answer separately, that is a signal worth respecting.
Second, the models got good enough to actually run loops. Claude Fable 5, released in June 2026, can run autonomously for hours without falling apart and check its own work along the way. Loops were a nice idea before. Now they work in practice. One creator who called loops “overhyped” two days earlier publicly reversed after Fable 5 one-shot a large full-stack build, and said everyone should be building loops.
How to start using loops today
You do not need to be an engineer. Here are three ways to start, from easiest to most powerful.
1. Install the free plan-optimizer skill
The fastest way to feel a loop is a free Claude Code skill called plan-optimizer (the idea is credited to @goodalexander, packaged and shared by @seangeng). It generates a plan, scores it 0 to 100 against a rubric, critiques it, rewrites it to fix the weaknesses, and repeats until the score plateaus, keeping the best version it finds.
Install it with one command:
curl -fsSL https://seangeng.com/plan-optimizer-skill.zip -o /tmp/plan-optimizer-skill.zip && unzip -o /tmp/plan-optimizer-skill.zip -d ~/.claude/skills/ && rm /tmp/plan-optimizer-skill.zipRestart Claude, paste any draft plan, and say “improve this until it can’t get better.” It runs the score, critique, rewrite loop for you and shows the trajectory.
2. Paste a one-paragraph loop
Do not want to install anything? Paste this after any task and Claude will loop on its own output. It works on writing, plans, code, anything with a quality bar:
After you finish, score your own work 0 to 100 against a clear rubric you write first. List the weakest parts and why they cost points. Rewrite to fix the top weaknesses while keeping what scored well. Repeat until the score stops improving by a real margin, then give me the best version and the score trajectory.
3. Run a spec, build, review loop
For bigger builds, run three small skills in a loop instead of one giant prompt. This is the pattern operators are using to set Fable 5 off on full projects for hours:
- A
/specskill that asks you questions until it fully understands the feature, then writes a detailed spec and saves it to a notes file or second brain. - A
/buildskill that builds strictly from that spec. - A
/reviewskill that checks the build against the spec, lists what is wrong, and sends it back to/build.
Loop build and review until the review comes back clean. Put Fable 5 on high effort, kick it off, and step away.
FAQ
What is the difference between a prompt and a loop? A prompt is a single call: input to output, once. A loop runs that call repeatedly, with the model grading and fixing its own output between rounds, until the result stops improving.
Do I need to know how to code to use loops? No. The simplest loop is a single paragraph you paste after any task asking the model to score, critique, and rewrite its own work until it plateaus.
Why are loops suddenly everywhere? Two reasons: Anthropic and OpenAI both independently pointed people toward loops over single prompts, and Fable 5 can now run autonomously for hours, which makes loops practical instead of theoretical.
What is the fastest way to try a loop? Install the free plan-optimizer skill, hand it any plan, and watch it score, critique, and rewrite itself.
Want the three loops on one page plus the install commands? Grab the free Loops Starter Guide.