OpenAI Publishes GPT-6 Astra Prompting Guide With Banned 'Slop Words' List

OpenAI has published model documentation for GPT-6 Astra detailing specific behavioral quirks and the prompts developers can use to work around them, including an explicit blocklist of AI-cliché phrases the company calls "slop words."

According to OpenAI, GPT-6 Astra asks clarifying questions more often than its predecessor, GPT-5.6 Sol, rather than making assumptions independently. OpenAI describes this as making the model a "more effective collaborator," but notes the trade-off: the model sometimes stops where users expect it to continue working.

To counter this, OpenAI recommends a prompt instructing the model to infer user intent from context and demonstrate a "bias towards action." The suggested prompt tells the model to treat phrases like "can you..." or "help me..." as calls to act rather than invitations for follow-up questions, and to work independently — creating isolated work trees, resolving merge conflicts, or drafting PRs — unless actions are "clearly destructive or irreversible." The model should wait for approval until it has "a concrete, reviewable result," per OpenAI's guidance.

OpenAI also warns that GPT-6 Astra, while better at following long instructions, is more sensitive to context. Unclear or contradictory guidance in skill files such as AGENTS.md can cause the model to block work or behave unexpectedly. OpenAI recommends auditing all accessible skill files and giving explicit priority to user instructions. A suggested debugging prompt forces the model to name the exact skill file and quote the instruction that caused it to pause, helping developers trace unexpected behavior to its source.

The slop word blocklist

A significant portion of the documentation addresses writing style. OpenAI says GPT-6 Astra defaults to lists, tables, and Markdown formatting, and reuses stock phrases across sessions. For prose output, OpenAI recommends explicitly instructing the model to write "concise paragraphs using plain language and active voice," with lists reserved for information that is genuinely parallel or sequential.

OpenAI's documentation lists specific banned phrases, including "delve into," "leverage," "it's worth noting," "what's important is," and closing constructions like "In short:" or "The simplest mental model is:." Contrastive phrasing such as "X, not Y" is also flagged, along with invented hyphenated compounds like "exact-head checks" or "editorial-row layouts." OpenAI instructs the model to state what it's doing directly rather than listing what it won't do.

On coding tasks, OpenAI notes GPT-6 Astra runs thorough test suites even for small changes, producing testing effort disproportionate to the task. The company recommends prompting the model to rerun tests only when new failures justify it. Separately, OpenAI says the model underuses sub-agent delegation for parallel work and requires explicit instructions on when and how much to delegate, noting that inter-agent messages can contain grammar or spacing errors.

Developers migrating existing projects can reportedly use Codex with the OpenAI Docs skill to apply these prompt changes automatically via the command $openai-docs migrate this project to GPT-6 Astra.

What this means

This is prompt engineering guidance for an existing model, not a new release. It signals that GPT-6 Astra's default behavior — heavy clarification-seeking, verbose testing, and formulaic phrasing — diverges enough from developer expectations that OpenAI felt compelled to document specific counter-prompts rather than adjust the model's defaults. The granularity here, down to banning specific hyphenated compound words, suggests OpenAI is treating prompt-level style control as a substitute for retraining, likely because behavioral fine-tuning at this scale is slower and riskier than shipping a documentation page. For developers, the practical takeaway is that upgrading to a new model version may silently change agentic behavior and output style, making systematic prompt audits a necessary step rather than an optional one.