gpt-oss Prompt Generator

The gpt-oss Prompt Generator writes harmony-format system and developer messages for OpenAI’s open-weight gpt-oss-120b and gpt-oss-20b models. It is free, runs in your browser and needs no sign-up. It creates prompts for models you run locally; it does not call a model.
Models last verified:

Generate your prompt

Access 40,000+ AI Prompts for ChatGPT, Claude & Codex — Free!

Subscribe to get instant access to our complete Notion Prompt Library — the largest curated collection of prompts for ChatGPT, Claude, OpenAI Codex, and other leading AI models. Optimized for real-world workflows across coding, research, content creation, and business.

Access Free Prompt Library

Which gpt-oss model should I choose?

Compare gpt-oss-120b and gpt-oss-20b by parameter count, memory requirements and context window.

ModelBest use and hardwareContext windowLicence
gpt-oss-120bProduction/general-purpose/high-reasoning workloads that fit a single 80 GB GPU (117B total parameters; 5.1B active).131,072 tokensApache 2.0
gpt-oss-20bLow-latency, local, or specialized workloads (21B total parameters; 3.6B active); MXFP4 fits within 16 GB memory.131,072 tokensApache 2.0
From OpenAI’s gpt-oss announcement and model cards, checked 17 October 2026. Open-weight models: you run them yourself, so there is no OpenAI token price.

How do I use my prompt?

  1. Choose gpt-oss-120b or gpt-oss-20b.
  2. Describe your task and set the reasoning level.
  3. Add any tools and response format requirements, then generate the prompt.
  4. Paste the generated text into the system and developer messages of a local runtime such as Ollama, LM Studio or vLLM.

Prompting tips for gpt-oss

  • Use the Harmony response format; OpenAI’s model cards say gpt-oss will not work correctly without it. gpt-oss-120b model card
  • Keep the required system identity, knowledge cutoff, current date, reasoning line, and valid channels close to the Harmony template. Harmony response format
  • Set Reasoning in the system message to low, medium, or high rather than asking for hidden chain-of-thought. Harmony response format
  • Use analysis, commentary, and final channels, but never expose the model’s analysis/chain-of-thought to end users. Harmony response format
  • Define functions in a dedicated # Tools section in the developer message and require function calls to use commentary. Harmony response format
  • After a final answer, omit earlier chain-of-thought on the next sample; retain it only when it is needed for a tool-call continuation. Harmony response format

Worked examples

Contracts assistant with a tool

Before: Answer questions about our contracts using the supplied documents and cite each clause.

After (made with this generator):

[System message]
You are ChatGPT, a large language model trained by OpenAI.
Knowledge cutoff: 2024-06
Current date: [TODAY]

Reasoning: high

# Valid channels: analysis, commentary, final. Channel must be included for every message.
Calls to these tools must go to the commentary channel: 'functions'.

[Developer message]
# Instructions
You are a contracts assistant for [COMPANY]. Answer only from the documents provided, and quote the clause number for every answer.

# Tools
## functions
namespace functions {
// Look up a contract clause by its number.
type get_clause = (_: {
clause_number: string,
}) => any;
} // namespace functions

# Response Formats
## clause_answer
{"type":"object","properties":{"answer":{"type":"string"},"clause":{"type":"string"}},"required":["answer","clause"]}

gpt-oss-120b: temperature 1.0, top_p 1.0 (OpenAI's recommendation).

What the generator added: The prompt adds high reasoning, a clause lookup function routed through the commentary channel, and instructions to answer only from supplied documents. It also defines a JSON schema requiring answer and clause fields.

Fast classifier (20b)

Before: Classify each message as billing, technical or other.

After (made with this generator):

[System message]
You are ChatGPT, a large language model trained by OpenAI.
Knowledge cutoff: 2024-06
Current date: 2026-10-17

Reasoning: low

# Valid channels: analysis, commentary, final. Channel must be included for every message.

[Developer message]
# Instructions
Label each message as billing, technical or other. Reply with the label only.

# Response Formats
## label
{"type":"object","properties":{"label":{"type":"string","enum":["billing","technical","other"]}},"required":["label"]}

gpt-oss-20b: temperature 1.0, top_p 1.0 (OpenAI's recommendation).

What the generator added: The prompt adds low reasoning and an instruction to reply with the label only. It also defines a JSON schema with a required label field restricted to billing, technical or other.

Frequently asked questions

Is the gpt-oss Prompt Generator free?

Yes. The generator is free, runs in your browser and needs no sign-up. It writes prompts only and never calls a model.

Which gpt-oss model should I choose?

Choose gpt-oss-120b for general-purpose or high-reasoning workloads that fit one 80 GB GPU. Choose gpt-oss-20b for low-latency, local or specialised workloads; its MXFP4 version fits within 16 GB of memory. Both support a 131,072-token context window.

Do I need to paste raw harmony formatting?

Usually, paste only the message text. Most local runtimes, including Ollama, LM Studio and vLLM, apply harmony formatting for you. The generator also provides a raw harmony view when you need the full format.

Does the JSON schema guarantee valid output?

No. The JSON schema in the developer message guides the model’s response but does not guarantee that the output will follow it.

Should I show the model’s reasoning to users?

No. OpenAI’s rules say never to show the analysis channel to end users. Drop earlier reasoning after a final answer, except during tool calls.

Which temperature and top_p settings should I use?

Use temperature 1.0 and top_p 1.0, OpenAI’s recommended settings. The generator includes these in a settings line.