The Evaluation Loop: How AI Improves Its Own Output

AI is better at evaluating content than generating content.

This makes sense, when you stop to think about it. For a large language model to interpret your instructions, and create output, all in the same step, it dilutes the same token budget over two requests.

But if you have ask AI to evaluate its own outputs, it can focus all its attention on identifying how to improve. Then it can rewrite, separately.

Here’s a simple copy/paste process you can use to rapidly improve your LLM responses:

Ask an LLM to create some content for you
In a second tab, have another LLM evaluate the output.
Ask the second LLM for a clear brief documenting improvements.
Copy that brief, and paste it into the first LLM tab with the prefix, “Analyse this feedback.”
Ask for a V2.
Repeat until satisfied
Compare with diffchecker.com if you want to see the changes

This simple process will help you stir your prompt around in a cauldron, like the wizard of the future that you are.

⚖️ “With a leveraged worker, judgment is far more important than how much time they put in or how hard they work.” – Naval

🎥 Replay: Create a Coaching Relationship with AI

If you missed my webinar last week, you can watch the replay, and use 30 copy/paste prompts to deepen your relationship with an AI thinking partner.

Watch Now

📰 New AI News This Week

Anthropic released Claude Sonnet 4.6 which gives Opus-like reasoning for a fraction of the cost
Google launched Gemini 3.1 Pro which has twice the reasoning capacity of Gemini 3
Digital Ocean reports the biggest blocker in implementing AI agents in enterprise is the cost of inference