Latest Posts

The Prompt Is Dead: Why Loop Engineering Is Replacing Manual Prompting in 2026

Loop engineering AI workflows are systems you design once so that the AI gets prompted, checked and re-run without you typing each instruction. You don’t write the perfect prompt. You write the goal, the test that decides whether the goal is met, a file the AI uses to remember what it already tried, and a schedule that tells it when to start. The model goes round that loop until the check passes or a limit stops it.

That’s the whole idea. It sounds like a small change, but it moves your job from “person who types prompts” to “person who designs the system that types prompts”. This guide explains where the term came from in mid-2026, how a loop is built, what it costs in tokens, and when a plain prompt is still the better tool.

Last checked: November 2026. Tool features mentioned here change often, so confirm details in the official docs before you rely on them.

What is loop engineering, in plain words?

A normal chat with ChatGPT or Claude is one round trip. You ask, it answers, you judge the answer and decide what to ask next. You are the loop.

Loop engineering takes you out of that middle seat. In his June 2026 essay Loop Engineering, Addy Osmani (who, per his site bio, works on Claude Code at Anthropic) describes it as replacing yourself as the person who prompts the agent. He also describes a loop as a recursive goal: you define a purpose and the AI keeps iterating until it is complete.

Think of it like a pressure cooker with a whistle. You don’t stand there checking the dal every minute. You set the conditions, and the whistle tells you when something has happened. A loop is the cooker. The verification step is the whistle.

Where the idea came from (and what was actually said)

The term spread quickly in June 2026, mostly among people building coding agents. Here’s what can be traced to public sources, without adding words nobody said:

  • Addy Osmani wrote the essay that named and organised the idea, then followed it with Practical Loop Engineering in August 2026, a hands-on guide built around Claude Code’s /goal and /loop commands.
  • Peter Steinberger is quoted in Osmani’s essay describing the work as designing loops that prompt your agents.
  • Boris Cherny, described by Osmani as head of Claude Code at Anthropic, is quoted in the same essay saying “My job is to write loops.” In later interviews and talks, he has kept returning to one theme: verification.

You may also see blogs claiming that other famous tech leaders “declared prompting dead”. There are no verified quotes behind those claims, so they’re not repeated here. What is fair to say: public discussion in mid-2026 highlighted loops as the next step after prompt engineering and context engineering, especially among people building with coding agents.

Prompt engineering vs loop engineering

Prompt engineering hasn’t disappeared. A loop still runs on prompts. The basics in Anthropic’s prompt engineering overview and OpenAI’s prompt engineering guide, such as clear instructions, examples and a defined output format, matter even more inside a loop, because a vague prompt now repeats itself dozens of times. The difference is who writes and judges them, and how often.

Prompt engineering Loop engineering
Your job Write and refine each prompt Design the goal, check, memory and schedule
Who judges the output You, every time A test, a checklist or a second “checker” model, then you at the end
Memory The chat window (lost when you start fresh) A file or board outside the chat
When it runs When you open the app On a timer, a trigger or until a goal is met
Best for One-off tasks, creative work, taste Repeatable tasks with a clear “done”
Cost profile Low tokens, high attention Higher tokens, lower attention

The last row is the real trade-off, and it’s the one most hype posts skip. More on that below.

The four building blocks of loop engineering AI workflows

Osmani’s essay lists several parts for coding agents: automations, worktrees, skills, plugins and connectors, sub-agents, and a state file. Coding-specific bits like git worktrees don’t matter much for bloggers, so here they’re grouped into four plain-language blocks.

1. Goal definition: what does “done” look like?

A loop needs a finish line it can check. “Write a great blog post” isn’t one. “Produce a 1,500-word draft with every statistic linked to a source and no paragraph over 80 words” is.

In Practical Loop Engineering, Osmani says that if you don’t have a clear idea of what done or good means, a loop may not be the right pattern. That’s the most useful sentence in the whole debate for content people. Our work is full of fuzzy goals.

2. Verification: who checks the work?

This is the part that separates a loop from a runaway script. Osmani calls an agent’s claim that it’s finished “a claim and not a proof”. His fix is a maker/checker split: one agent drafts, a separate one verifies.

For a blogger, verification can be:

  • A script that counts words, checks every link returns a live page and flags banned phrases.
  • A second AI pass with a strict checklist (“List every number in this draft. For each, give the source URL or write MISSING.”).
  • A hard rule that stops the loop, like “stop after 5 attempts” or “stop if two attempts produce the same result”.

Cheap, objective checks work best. A word count is cheap. “Is this funny?” isn’t something a loop can judge for you.

3. Persistent memory: where does the loop remember things?

Models forget between sessions. Osmani’s line is “The agent forgets, the repo doesn’t.” The loop keeps a state file (a Markdown note, a spreadsheet or a task board) recording what was tried, what passed and what’s still open.

For a research loop, that file might hold the sources already read, claims already verified and topics already rejected. Without it, every run starts from zero and pays again for the same work.

4. Scheduling and discovery: when does it start?

A loop needs a trigger. Claude Code’s /loop reruns a prompt on an interval. Osmani notes that it runs on your machine and stops when the session or machine stops, and that recurring loops expire seven days after creation. For anything longer, he points to scheduled cloud routines (/schedule). The official Claude Code scheduled tasks documentation confirms the split: /loop tasks are session-scoped, while cloud and desktop scheduled tasks keep running without an open session.

“Discovery” means the loop finds its own work: new comments to answer, new search queries showing up in Google Search Console, new items in a reading list. If nothing new turns up, a well-designed loop reports that and stops.

A simple research loop for bloggers (no coding required to understand it)

Here’s how those four blocks fit together for a weekly content-research routine. It’s a design pattern, not a product. You can build parts of it with Claude Code, ChatGPT’s scheduled tasks or a no-code automation tool, depending on what you already use.

Goal: Every Monday, produce a one-page brief on what changed in my niche last week, with every claim linked to a primary source.

Discovery: Read the five sources listed in sources.md (official blogs, newsrooms, documentation changelogs).

Maker step: Summarise only items published in the last seven days.

Checker step: A second pass with this prompt:

You are a strict fact-checker. For each bullet in the brief:
1. Is there a source URL? If not, mark NO-SOURCE.
2. Does the source page support the exact claim? If unsure, mark CHECK.
3. Is the item older than 7 days? Mark OLD.
Return only the bullets marked OK, then a list of rejected bullets with reasons.

Memory: Append every approved item and every rejected URL to state.md, so next week’s run skips them.

Stop rules: Maximum two maker/checker rounds. If fewer than two items pass, output “Quiet week, nothing worth covering” and stop.

Notice the human step that remains: you read the brief and decide what’s worth writing about. The loop saves you the gathering and checking, not the judgement.

The honest part: token cost

Loops trade your attention for tokens. Every round re-reads instructions, files and earlier output. Osmani warns in his essay that you have to be careful about token costs and that sub-agents burn more tokens.

The horror stories mostly come from loops with no stop rule. A Towards AI write-up describes a developer who reported about $6,000 in usage after one unattended loop command. That’s a single reported case from secondary coverage, not a typical bill, but the mechanism is real: when a session’s history grows, each new call carries more of it.

Practical guardrails:

  • Set a spending cap at the account or API level before you schedule anything.
  • Give every loop a stop rule: maximum rounds, maximum time, or “stop if nothing changed”.
  • Keep the state file short. Summarise old entries instead of letting it grow forever.
  • Start manually. Run the loop by hand three or four times before you put it on a timer.

If you’re on a fixed monthly plan rather than pay-per-use API billing, the risk is hitting usage limits rather than a surprise bill. That’s still lost working time.

What users report: single prompts vs multi-step loops

This section summarises what people building these systems have said publicly, not a personal test. Treat it as patterns from public user feedback, not lab results.

Multi-step beats single-shot when the check is objective. Across developer write-ups and the Claude Code team’s own commentary, the consistent claim is that quality improves most when a separate step verifies the work. Cherny’s public emphasis on verification is the most cited example, and Osmani’s own loop engineering essay builds its design around a separate checker. Writers who move from one big “write my post” prompt to an outline → draft → fact-check chain describe similar gains: fewer made-up numbers and less padding, because each step has one narrow job.

Loops can spin. Osmani describes the same command being tried again and again with no change in result. Users report the same in forums: a loop that can’t pass its check keeps paying to fail. The fix is a “no progress, stop” rule.

Unattended means mistakes are unattended too. In Osmani’s words, a loop running unattended is also a loop making mistakes unattended. He admits he nearly pushed changes he hadn’t reviewed closely, and calls the whole approach “still early”.

Creative work resists loops. Osmani lists tasks that need human taste, subjective design or open-ended creative exploration as a poor fit. That matches what many bloggers find: AI loops are good at gathering, checking and formatting, and weak at deciding what your readers will care about.

When you should NOT build a loop

Skip the loop and use a normal prompt (or a short prompt chain) when:

  • You can’t write the “done” test in one sentence. If you can’t define done, the loop can’t either.
  • The task happens once. Designing a loop for a one-off post costs more time than it saves.
  • Taste is the product. Headlines, personal stories and opinion pieces need you.
  • Mistakes are expensive. Anything touching money, passwords, legal claims or publishing live to your site needs a human gate. Osmani says he watches loops closely when they touch authentication, security or finance.
  • You haven’t capped spending. No cap, no schedule.

A sensible path for most bloggers: start with prompt chaining, where you run each step yourself. Automate only the step with the cheapest, clearest check, usually link checking or source collection. Then add a schedule once that step has run cleanly a few times.

Related reading

Frequently asked questions

Is prompt engineering dead in 2026?

No. Loops still run on prompts. What changed is that experienced users increasingly write prompts for systems (the maker, the checker, the stop rule) rather than typing a fresh prompt for every task. The headline is a provocation, not a literal claim.

Who coined the term loop engineering?

Addy Osmani’s June 7, 2026 essay named and structured the idea, drawing on public comments by Peter Steinberger and Boris Cherny. The underlying pattern of agents iterating until a check passes is older. The 2026 shift was treating the loop itself as the thing you design.

Do I need to code to use loop engineering AI workflows?

Not for the thinking part. Defining a goal, a checklist and a memory file is planning work. Running loops on a schedule is easiest with developer tools like Claude Code, but some chat apps and no-code automation tools offer scheduled tasks too. Check what your plan includes.

Are AI loops expensive?

They can be. Each round uses tokens, and sub-agents use more. With stop rules, short memory files and a spending cap, a small weekly research loop stays modest. Without them, costs and usage limits can climb fast, as several user reports show.

What’s the difference between prompt chaining and loop engineering?

In prompt chaining, you run a fixed sequence of prompts and judge each output yourself. In loop engineering, the system repeats steps, checks its own work against a test and decides whether to continue, with you reviewing the end result.

What is the most important part of a loop?

Verification. Both Osmani and Cherny return to it repeatedly. A loop without a reliable check is just a faster way to produce unchecked output.

The bottom line

Loop engineering AI workflows aren’t magic, and the prompt isn’t really dead. What has changed is where the skill sits. In 2026, the people getting the most out of AI design small systems with a clear goal, a strict check, a memory file and a sensible stop rule, then step back.

For bloggers, the win is narrow but real. Let loops do the gathering and checking. Keep the judgement, the voice and the final “publish” click for yourself.

The AI Side-Income Starter Kit – free PDF, 13 pages
Free starter kit
The AI Side-Income Starter Kit

5 realistic ways to earn your first $100 online with AI, 10 copy-paste prompts and a 7-day plan. Free PDF.

Latest Posts

spot_imgspot_img

Don't Miss

Stay in touch

Get new travel guides, AI workflows and blogging tips in your inbox. No spam, unsubscribe anytime.