# MiniMax prompt guide: coding and agent prompts for MiniMax-M3

> How to write prompts for MiniMax-M3 and MiniMax-M3.1 Flash: acceptance criteria, fixed sampling, JSON without a JSON mode and tool loops, from MiniMax's API docs.

Source: Promptea — https://www.promptea.me/en/guides/minimax-prompt-guide · Language: English · Last updated: 2026-11-14

## What the docs tell you

- MiniMax lists both models as frontier coding models with a 1M-token context and text, image and video input.
- Temperature and top_p are ignored: the models always run at 1 and 0.95. Precision has to come from the prompt.
- MiniMax-M3 reasons by default and can turn it off (thinking: disabled). MiniMax-M3.1 Flash always reasons; tune it with reasoning_effort (default max).
- In tool calling, send the model's full response back on every turn, including reasoning_content, or the loop loses its reasoning.
- MiniMax publishes no general prompting guide, so the advice below is about what makes a coding prompt checkable.

## Patterns that work

- State the goal in one sentence, then the acceptance criteria as observable behavior, then the command that verifies them.
- Name the files that matter and what must not change; agents widen the scope when nothing bounds it.
- For JSON there is no documented JSON mode: give the exact shape, one example and 'return the JSON only'.
- Ask for a closing report: what changed, what was verified and what was not.

## Templates

### Bug fix with a failing test first

```text
Goal: fix [bug, one sentence].
Expected: [behavior]. Actual: [behavior + error message].
Relevant files: [paths]
Do not change: [public API, other modules]
Steps:
1) Write a test that reproduces the bug and fails.
2) Fix the code until it passes.
Verify with: [test command]
Report: the cause, the diff summary and anything left unverified.
```

[Open in Promptea](https://www.promptea.me/en/analyzer?prompt=Goal%3A+fix+%5Bbug%2C+one+sentence%5D.%0AExpected%3A+%5Bbehavior%5D.+Actual%3A+%5Bbehavior+%2B+error+message%5D.%0ARelevant+files%3A+%5Bpaths%5D%0ADo+not+change%3A+%5Bpublic+API%2C+other+modules%5D%0ASteps%3A%0A1%29+Write+a+test+that+reproduces+the+bug+and+fails.%0A2%29+Fix+the+code+until+it+passes.%0AVerify+with%3A+%5Btest+command%5D%0AReport%3A+the+cause%2C+the+diff+summary+and+anything+left+unverified.&purpose=code&target=minimax)

## Frequently asked questions

### Can I make MiniMax more deterministic with temperature 0?

No: MiniMax ignores temperature and top_p on these models. Make the output stable with an exact format, an example and explicit acceptance criteria.

### What happens to MiniMax-M2.7?

MiniMax-M2.5 and M2.7 take no new users since 2026-10-19 and retire on 2026-11-18. Move to MiniMax-M3.
