If you've started using GPT-6 Astra, you've likely noticed it behaves fundamentally differently than its predecessors. While it's more powerful, many users find it frustratingly hesitant or prone to repetitive, list-heavy 'slop'.
The Simple Version
Prompting GPT-6 Astra is less like coaching a student and more like giving a highly literal executive assistant a precise checklist. Older models often needed to be told to "think step-by-step" or "be careful," but Astra is so literal that those instructions can actually paralyze it.
How It Actually Works
According to the video, Astra is stronger at general instruction following than previous models, but this precision creates new challenges. As Nick Puru notes, "the old habits actually start working against you."
The 'Stop-and-Ask' Problem
Astra is designed to avoid assumptions. While this reduces hallucinations, it means a prompt like "Can you help me write this?" often results in Astra responding, "Yes, I can help! Here is my plan..." instead of actually doing the work. To fix this, you must explicitly tell Astra to treat requests for help as direct instructions to execute the task.
Combating 'AI Slop'
Astra has a tendency toward long, formatted responses with repetitive phrases and nested lists. This is a common trait in high-parameter models that over-optimize for "helpfulness" by providing exhaustive structures. To counter this, you need to specify a writing style that favors concise paragraphs and active voice.
The Reasoning Effort Scale
Unlike previous versions, Astra's thinking settings (Light, Medium, High) are not 1:1 replacements for older models. The video highlights that Astra on its lowest setting often outperforms GPT-5.6 Soul on its highest. This means running every task on "High" is not only a waste of tokens but potentially unnecessary for the quality of the output.
Technical Deep Dive
From an architectural standpoint, Astra's increased sensitivity to instructions suggests a more refined attention mechanism regarding system prompts and attached files. For technical users, the key is managing the Context Window and Reasoning Budget efficiently.
The video mentions specific API pricing: $10 per million input tokens and $50 per million output tokens. Because Astra is expensive, the goal is to reduce the number of turns in a conversation. By defining "completion criteria" (the result, the limits, and the check), you move from an iterative prompting loop to a single-shot execution.
For context, this approach mirrors Test-Driven Development (TDD) in software engineering. Instead of writing code and then testing it, you define the test (the "done" state) first. By providing Astra with a specific check—such as "ensure every next step has an owner and a due date"—you are essentially providing a validation script that the model runs against its own output before delivering it to you.
Why This Matters
The shift toward more literal instruction following means that "prompt engineering" is evolving into "constraint engineering." We are moving away from vague adjectives ("be professional") and toward structural requirements ("use active voice, no nested lists").
This is critical for business automation. If an AI agent is integrated into a workflow and stops to ask for permission at every step, the efficiency gains of AI are lost. By implementing the "do not stop at acknowledging capability" instruction, developers can create more autonomous, reliable agents.
Finally, the discrepancy between Astra and GPT-5.6 Soul emphasizes the importance of benchmarking. As the video points out, "the settings don't carry over between model generations." Users must re-baseline their workflows every time a new model version is released to avoid overpaying for compute they don't need.