OpenAI recently published a post titled "Rethinking skills and prompts for GPT-6 Astra," rethinking the skills and prompts for GPT-6 Astra.
Original link: https://developers.openai.com/blog/rethinking-skills-and-prompts-for-gpt-6-astra
Those who are interested can read the original; below are my own thoughts.
From the earliest Prompts to Workflow to Skill, as model capabilities improve, complex rule content does not make the model perform better, and can even become a burden.
For example, whether a factory hires employees with an elementary, middle school, high school, or college education, if it puts them all on the same assembly line and has them work according to the same rules, the final result is obvious.
To tell whether your Skill is concise and clear enough, first look at the description. If the description seems to want to cram the entire contents of the Skill into it, then you might consider streamlining/discarding this Skill.
In my understanding, the description only needs to focus on describing "when to use it and what it produces"; usually two sentences are enough to make it clear. Then comes the content. Everyone has their own formatting style; there is no unified rule for this. But the wording must be concise and refined, using declarative sentences.
There is a very interesting detail: many people want to write a Skill that removes the AI flavor, but find writing it themselves troublesome, so they ask AI to help write it. The wording of the Skill text written by AI itself has a strong AI flavor; using a Skill that inherently has an AI flavor to remove the AI flavor is simply impossible.
How to slim down a Skill? Here is a simple example:
- Before rewriting: Prohibit the use of weak password policies during login, registration, password recovery, password change, mobile phone number binding, mobile phone number unbinding, email update, account deletion, and account cancellation.
- After rewriting: Sensitive operations must always use strong password validation and secondary confirmation.
For better models, you do not need to exhaustively enumerate specific scenarios; you cannot enumerate them all, and it only increases the volume of rules. Here are my practical suggestions:
- The frontmatter description should only describe "when to use it and what it produces"
- Follow the higher-level principle; do not exhaustively enumerate special cases
- No outdated rules, no duplicate rules of the same kind, no lower-level fragmented scenarios
- If it can be said clearly in one sentence, never pile on explanations
After creating a new Skill, I always review the Skill against the above rules.