THIS Is the AI Setting Everyone Gets Wrong

Name: THIS Is the AI Setting Everyone Gets Wrong
Uploaded: 2026-07-13T22:00:59.000Z
Duration: 16 min 18 s
Channel: Mark Kashef
Description: Video investigates how 'effort' settings in AI models (GPT, Claude) actually function and demonstrates through experiments that maximum settings are often suboptimal for task execution.

Mark Kashef13 July 2026Watch on YouTube

Part of series

Ep. 5 · Beter webdesign met AI

Analyse en tutorials over hoe je AI-tools inzet voor hoogwaardiger, onderscheidend webdesign in plaats van generieke resultaten.

View the series

What you'll learn

Effort settings in AI models (GPT, Claude) don't work as a simple slider where higher always means better output
Maxing out effort levels can backfire and produce worse results for certain tasks instead of improved performance
A framework exists to determine the optimal effort level for any AI model based on your specific task requirements

Frequently asked questions

Why doesn't maximum effort always lead to better AI results?

Maximum effort can cause overcomplication of the task and result in less relevant answers. Experiments with Claude Code and Codex show that moderate effort levels often outperform ultra-high settings.

How do you determine which effort level to use for a specific task?

By first selecting your AI model, then experimenting with different effort levels on your specific prompt, and comparing the results. The video presents a framework and free decoder guide to streamline this process.

Do effort levels work differently across AI providers like OpenAI and Anthropic?

Yes, effort levels function differently per provider since each company implements its own settings. The video demonstrates that the same effort level doesn't have identical effects on Claude Code compared to Codex.

Topics

prompt engineering llm optimization ai models claude gpt

Description from the channel

Master Agentic Workflows: https://www.skool.com/earlyaidopters/about Get the effort decoder FREE: https://markkashef.gumroad.com/l/effort-decoder Work With Us: https://www.promptadvisers.com Every month a new model family drops, and every one ships with an effort dial. GPT 5.6 alone gives you six levels across three tiers, which is 18 possible combinations for a single task. Most people treat that dial like a slot machine and assume more effort means a smarter model. It doesn't. In this video I break down what effort actually is, why maxing it out can quietly backfire, and the framework I use to reverse engineer the right level for any model, even ones that haven't come out yet. To prove it, I ran the identical prompt across 12 effort levels on Claude Code and Codex and walk through every result side by side. The differences will surprise you. By the end you'll know when low is more than enough, when high actually earns its tokens, and why max, ultra, and extra high are almost never the answer. --- 0:00 - the effort trap 1:02 - the slot machine habit 2:07 - what level to use when (short version) 3:11 - when high effort backfires 4:00 - pick the model first, then the effort 5:23 - the brain in a jar (harnesses) 6:37 - why effort levels differ across providers 7:18 - every effort level, visually 9:04 - the experiment, 12 runs, one prompt 10:33 - claude code results, low to max 11:46 - codex results, low to max 14:01 - the framework for any new model 15:37 - the free effort decoder guide #ai #claudecode #codex #claude #anthropic #openai #gpt56 #aitools #aiagents #promptengineering #aiproductivity #llm #vibecoding #aiforbusiness #sol #terra