OpenAI’s New Advice: Delete Half Your Prompt

Three years of advice told everyone to write longer, more elaborate prompts. Assign a role, specify a format, add constraints, give it a persona. OpenAI Developers has now published guidance for GPT-6 Astra pointing the other way, and the instruction quoted from it is blunt: revisit each instruction and ask whether the task still needs it.

What the guidance actually says

The audience is developers working with Codex, skill files and AGENTS.md configuration, which is worth stating before anyone applies it to their marketing prompts. Four themes run through it.

Descriptions should be as short as possible while still making clear when to use them. Scope a tool description to the specific trigger rather than the general category.

Point to detail rather than including it. A root configuration file should route to supporting documents rather than containing everything, because reading it costs context and pushes the session closer to compaction.

Grant permission instead of writing defensive rules. Blanket instructions like review everything first or always run tests get replaced with explicit permission for specific known-safe workflows. The guidance asks directly whether an ask-first rule exists because an action is genuinely high-stakes, or because an older model once misbehaved.

Define what done means. Astra is described as more tentative than its predecessor about continuing without review, so a task that includes running the result and fixing what fails needs to say so, or the model stops earlier than you wanted.

Why old instructions turn into liabilities

Almost every elaborate prompt in circulation was written to compensate for something a model got wrong at the time. Tell it to think step by step, because it used to skip steps. Tell it not to invent citations, because it used to. Tell it to check its arithmetic. Each line was a patch for a specific failure.

The failures moved. The patches stayed, because nobody deletes a line from a prompt that appears to be working. Now they cost something twice over: the context they consume, and the behaviour they suppress when they tell a capable model to stop and ask about something it would have handled.

This is the maintenance problem we described in a new AI model every three weeks, showing up in a place most businesses never think to maintain. Software gets updated. Prompts get saved once and reused for years.

What transfers to ordinary business use

Being careful here matters, because the leap from developer guidance to everyday advice is exactly where this kind of story gets oversold. OpenAI did not publish a claim that shorter prompts produce better emails.

What plainly carries over is the audit habit. If your team has saved prompts, they have dates, and the ones written more than a couple of model generations ago are worth rereading with one question in mind: is this line telling the model something about my business, or is it telling the model how to be a model.

The left column is information the model cannot get from the request. The right column is instruction about how to think, which is where the older advice concentrated. Test rather than assume, because occasionally a line is doing more work than it looks like.

A twenty-minute audit

1. Find your most-used saved prompt. The one somebody pastes several times a week. That is where any waste compounds.

2. Make a copy and cut every line about how to think. Keep the audience, the constraints, the format and the examples. Delete the role-play, the encouragement, and the instructions not to hallucinate.

3. Run both versions on five real tasks. Not invented test cases. Actual work from last week, where you know what a good answer looks like.

4. Keep whichever wins, and date it. If the trimmed version holds up, replace the original and note the date and model. If a specific line turns out to be load-bearing, put it back and now you know why it is there.

Repeat once a model generation, which currently means every few months. Our collections in essential business prompts and ChatGPT prompts for business are worth running through the same test rather than treating as fixed.

The agent version is more consequential

For anyone who has an agent configured against their systems, the permission point deserves attention. Replacing blanket restrictions with explicit permission for known-safe workflows is the same reasoning as scoping what an agent may reach rather than listing what it may not, which we set out in three boxes for an AI agent.

One caution before anyone deletes their approval gates on the strength of a prompting guide. Removing an ask-first rule because the model is now competent is different from removing it because the action is reversible. Keep the gate on anything that moves money, sends externally, signs, or deletes, regardless of how good the model has become. The capability side of Astra is covered in OpenAI's new model is built to use your computer.

Frequently Asked Questions