Home/News/OpenAI publishes GPT-6 Astra prompting guide with a slop word blocklist

AI

OpenAI publishes GPT-6 Astra prompting guide with a slop word blocklist

2026-09-05 AIHOT 精选 AI
OpenAI publishes GPT-6 Astra prompting guide with a slop word blocklist
图:AIHOT 精选

OpenAI documented GPT-6 Astra's behavioural differences alongside prompting guidance, including a blocklist of "slop" words. That list is more useful to content people than to developers.

Key facts

  • Timing: Documentation is published and usable now
  • Who it affects: Stores using AI for product copy, blogs and email; teams running automation
  • What to do: Worth acting on: paste the blocklist straight into your writing prompt template

What happened

On behaviour, OpenAI notes Astra asks clarifying questions more often than GPT-5.6 Sol and is more context-sensitive, with the side effect that it sometimes stops when the user expects it to continue. The recommended fix is to frame instructions so the model infers user intent and shows a "bias towards action", making explicit that phrases like "can you" and "help me" should trigger task completion rather than more questions. It also advises having the model prepare concrete, reviewable results before requesting approval, and stripping out unsolicited warnings and safety checklists built on hypothetical risks.

The slop blocklist names "delve into", "leverage", "it's worth noting", "what's important is", "promote", and constructed compounds like "exact-head checks", recommending you state intended actions directly rather than listing what will not happen. On skill files such as AGENTS.md: unclear or contradictory instructions cause unexpected behaviour, so audit every skill file the model can read, prioritise user instructions explicitly, and use a debugging prompt that forces the model to cite the exact instruction causing a pause. On sub-agents: the model delegates less often than expected, so specify when and how extensively it should.

What it means for independent store owners

The blocklist converts to money fastest. "Delve into", "leverage" and "it's worth noting" are the clearest fingerprints of generated text. Readers may not name what is wrong, but the paragraph reads hollow to them. Adding these as banned words in your product description, blog and email prompt templates costs nothing and immediately improves how the copy lands. If you already keep a brand banned-words list, merge them in.

"State what you will do rather than listing what you will not" applies equally to customer-facing copy. A product page reading "no sulphates, not tested on animals, no added fragrance" works worse than one saying what the product contains and what it fixes. The first reads like a disclaimer; the second sells.

If you run automation — scheduled content generation, automated support replies — take the skill file audit seriously. Automations that drift out of control usually have not degraded; they have accumulated contradictory rules across prompts and config files. OpenAI's debugging method is genuinely practical: make the model name the exact instruction that caused it to stop or change behaviour, which beats guessing line by line. And note the delegation point: if your workflow assumes the model will split work into sub-tasks on its own, do not assume. Write it out.

Someone should be watching these changes for you

I run AI automation for independent stores — content, SEO/GEO, product listing, social, and keeping up with policy shifts like this one.

See how it works