So musst du KI jetzt prompten: Was ist Loop Engineering (Deutsch)

Overview: “Loop Engineering” in AI prompting (Deutsch)

The video explains Loop Engineering as an emerging concept in prompt design. It argues that it’s not just “better prompts,” but about how to structure iterative AI behavior when outputs need to be corrected until they meet defined goals.

Background / where the term comes from

• The term Loop Engineering is described as coming from the developer community.
• It is associated with:
  • Boris Journey (inventor of Claude Code)
  • Peter Steinberger (developer of Open Claw)
• Andrew Ng is cited as having written an article about it (published in his newsletter The Batch on June 26), framing it as more than a trend.

The 3-stage progression (conceptual framework)

The speaker presents “Loop Engineering” as Stage 3 in an evolution:

1. Stage 1: Prompt Engineering

   • Focus: how you write prompts (role, task, format, examples).
   • Historically important because models were more rigid: wording changes could drastically alter results.
   • Newer models are more robust, shifting attention from pure wording to what you’re effectively asking the model to do.

2. Stage 2: Context Engineering

   • Focus: what the model uses (files, personal instructions, projects, memory).
   • Emphasizes that good prompts should include context-building blocks.
   • Since models may now retrieve or read data themselves, failures can occur between rounds, not only in the initial prompt.

3. Stage 3: Loop Engineering (main topic)

   • Focus: designing the process the AI follows across multiple rounds.
   • Key questions the loop must define:
     • How many times can the AI iterate?
     • How does it measure whether the result is good?
     • How does it know when it is finished?
   • Important: Stage 1 and 2 are not removed—looping is an additional layer on top.

Core instruction: “criteria” instead of “control sentences”

A major practical point is to avoid prompts like:

“Check your answer.”

Instead, define measurable quality criteria the AI can validate against.

Why this matters

• Newer models can self-check automatically, so telling them to “check the check” adds steps without improving outcomes.
• The loop should clearly specify:
  • what is measured
  • and what condition ends the loop

Example tutorial embedded in the explanation

The video demonstrates the approach with an e-commerce text task.

Task: Create product descriptions for an online shop with four objective criteria:

1. Max 300 characters
2. Material must be mentioned
3. No superlatives (no “best,” “perfect,” “unique”)
4. Care instructions at the end

What “loop” prompting looks like (conceptually)

• The prompt specifies:
  • the four criteria,
  • a stopping condition (“you’re finished when all four criteria are met”),
  • and a limit on attempts (e.g., stop after five attempts and report which criterion is missing).

How the loop behaves

• Round 1 may fail length / remove superlatives / miss care instructions.
• The AI identifies which criterion(s) are not satisfied and rewrites.
• Subsequent rounds continue until all criteria are met.

When Loop Engineering works best

The speaker gives a key rule:

• Use loop criteria only when they are objectively measurable.

Examples from the video: character count, presence of material, absence of superlatives, format rules.

If something requires human judgment or missing knowledge, it should stay with the human.

• Example given: brand tone or whether an item is truly made of linen.

Why keep loops short (tokens + context-window reasons)

Long multi-round processes create two main problems:

1. Tokens & time cost

   • Each round uses tokens and adds waiting time.
   • Over-iterating when fewer rounds would work wastes quota.

2. Context window degradation

   • The model has limited “working memory” (context window).
   • Multi-round drafts and corrections quickly fill it.
   • As the chat grows, performance can degrade (“answers get weaker”).
   • Practical advice:
     • keep the process as short as possible (e.g., 3 rounds instead of 20),
     • if the chat becomes too long, start a new chat and carry over only the final result.

The speaker also mentions their “context compressor” tool (from a previous “Skills” video) as designed for this scenario.

Main speakers / sources (as identified in the video)

• Boris Journey (inventor of Claude Code)
• Peter Steinberger (developer of Open Claw)
• Andrew Ng (author in The Batch newsletter)
• Anthropic (referenced via their Prompting Guide)
• The video’s speaker (host; not named in the subtitles)

---

Summary

Loop Engineering is presented as a Stage 3 evolution of prompting: after prompt and context engineering, you design a multi-round AI process with explicit measurable quality criteria and a clear finish condition. The video emphasizes using criteria for measurement (not “check yourself” control phrasing), limiting iterations for cost, and keeping chat histories short to avoid context-window problems.