OpenAI has released an updated guide on crafting prompts for its flagship model GPT-5.6 Sol and the entire GPT-5.6 family. The key message of the document is less text, more meaning. The development team strongly recommends that users formulate queries as concisely as possible, abandoning the lengthy step-by-step instructions that were characteristic of previous versions.
Internal tests on agentic programming tasks demonstrated impressive results from this approach. Compact prompts improved quality scores by 10-15%, reduced token consumption by 41-66%, and lowered task execution costs by 33-67%. This directly refutes the previous paradigm, where it was believed that the more detailed the query, the better the result.
The new approach is strikingly different from the documentation for GPT-5, published in August 2025. At that time, the emphasis was on XML blocks, templates for gathering context, and detailed tool invocation scenarios. Now, OpenAI urges removing repetitions, style instructions that do not affect model behavior, ineffective examples, and steps that the neural network already handles stably. Instead, the prompt should only contain key elements: the visible result, success criteria, stopping conditions, and hard constraints, such as safety requirements.
Important warning: GPT-5.6 Sol strictly follows the query conditions. Conflicting rules confuse the model more than a lack of details. The words "always" and "never" are recommended only for invariants, while conditional rules should be formulated for other cases.
The guide has also been expanded with two new sections. The text.verbosity parameter sets the base level of response detail (low, medium, or high), solving the problem of overly shortened responses due to old "answer briefly" instructions, since GPT-5.6 defaults to shorter answers than GPT-5.5. The second section is Programmatic Tool Calling — programmatic tool invocation for tasks with clear boundaries, where the code itself filters and groups results, returning a compressed summary to the model instead of raw data.
Mixed Reviews and First Incidents
The GPT-5.6 family — models Sol, Terra, and Luna — was introduced in late June as a limited preview. The full release took place on July 9. However, user feedback has been divided. Renowned AI investor Matt Schumer reported that an agent based on GPT-5.6 Sol accidentally deleted almost all files on his Mac. The cause was an error in processing the $HOME system variable, leading to recursive cleanup of the home directory.
Notably, a similar scenario was described in the GPT-5.6 system card before the public release. Codex lead Thibault Sottiaux confirmed the launch issues and promised improvements. BridgeMind project founder Matthew Miller faced a similar situation, where code from GPT-5.6 Sol canceled all subscriptions in Stripe. However, Miller himself admitted that he gave the agent full write access instead of a restricted API key.
There are also counterexamples. OpenAI employee Ethan Knight stated that GPT-5.6 Sol Ultra proved the Cycle Double Cover hypothesis — a graph theory problem that had remained unsolved for half a century. The model with 64 subagents took less than an hour. Additionally, on the HealthBench Professional test, GPT-5.6 Sol scored 60.5 points versus 59 for GPT-5.5, significantly outpacing the average doctor score of 43.7 points.
My analysis: The paradigm shift from "more context" to "less noise" is a logical step for models that have already learned to understand user intent. However, the file deletion incidents highlight that the safety of agentic systems still leaves much to be desired. Users should be extremely cautious with the access rights they grant to such agents.