This is a scannable reference for prompting Mistral Le Chat, powered by the unified Medium 3.5 flagship (128B, 256K context, chat + reasoning + code in one model). Le Chat runs at chat.mistral.ai, on mobile, and via API, with web search, Canvas, a code interpreter, image generation, and a no-code Agent Builder. Bookmark this and copy straight from the tables and prompt blocks below.
If you want fully written prompts instead of a reference, jump to our best Mistral prompts roundup, or grab reusable skeletons from the templates.
1. The prompt formula
Every strong Mistral prompt follows one shape: [Role] + [Task] + [Context] + [Constraints] + [Output format]. You do not need all five every time, but naming the output format is the single biggest quality lever — and Mistral rewards concision, so cut any sentence that is not doing work.
| Slot | What it sets | Example |
|---|---|---|
| Role | The expertise Mistral adopts | "You are a senior data analyst." |
| Task | The one action to perform | "Compare these three pricing plans." |
| Context | Background, inputs, audience | "For a bootstrapped SaaS founder." |
| Constraints | Limits, tone, length, must/avoid | "Under 200 words, no jargon." |
| Output format | Exact shape of the answer | "Return a markdown table." |
You are a senior financial analyst. Compare the three SaaS pricing plans I paste below for a bootstrapped founder with ~50 customers. Constraints: assume 5% monthly churn, ignore enterprise tiers, keep it under 250 words. Output: a markdown table with columns Plan, Monthly cost, Best for, One risk, followed by a single recommendation line.Best for: any analysis, writing, or planning task where you can picture the ideal answer's shape before you ask.
2. Mistral models table
Pick the model to the job. Le Chat defaults to the flagship, but the coding models and Work mode each unlock a different capability.
| Model / mode | What it is | Use it for |
|---|---|---|
| Mistral Medium 3.5 | Unified flagship — chat, reasoning, and code in one 128B model, 256K context, 77.6% on SWE-Bench Verified | Almost everything: writing, analysis, coding, long documents |
| Work mode | Routes to Medium 3.5 with extended reasoning and default-on connectors | Multi-step research, long drafts, structured analysis |
| Codestral 25.08 | Dedicated code model behind Le Chat's coding features | Completions, generation, fast code tasks |
| Devstral 2 & Vibe | Agentic coding stack — Vibe CLI plus async cloud agents in sandboxes | Autonomous fixes, multi-file tasks with GitHub/Slack |
| Open Small models | Smaller open-weight models | Cheap, high-volume extraction and classification |
| API JSON mode | Schema-enforced structured output at the decoding layer | Production JSON that must always parse |
Best for: deciding in one glance whether a task needs the flagship, a coding model, an autonomous agent, or a cheap open model.
3. Le Chat modes & features
Le Chat is more than a chatbox. These features change what a prompt can do — most are available on the free tier.
| Feature | What it does | How to invoke |
|---|---|---|
| Web search | Live web results with citations | "Search the web and cite every source with a link" |
| Canvas | Editable workspace for docs, code, and mockups with version tracking | "Draft this in a Canvas so I can edit in place" |
| Code interpreter | Sandboxed Python for analysis, charts, prototypes | "Use the code interpreter to run this on my data" |
| Document / OCR | Reads PDFs, spreadsheets, and scanned images | Upload the file, then ask for a summary or extraction |
| Image generation | Text-to-image via Black Forest Labs' Flux | "Generate an image: [subject, style, aspect ratio]" |
| Connectors | Gmail, Calendar, Drive, Slack, GitHub context | Enable in settings, then reference the source |
| Agent Builder | No-code reusable, shareable agents with tools | Build in Agent Builder with instructions + tools |
| Memory | Remembers preferences and context across chats | Tell it what to remember; manage in settings |
Best for: knowing which lever to pull — live data (web search), an editable draft (Canvas), real computation (code interpreter), or a reusable tool (Agent Builder).
4. Output-format modifiers
Add one line to control how Mistral returns the answer. These stack cleanly on top of any prompt above.
| Modifier | Add this line |
|---|---|
| Table | "Return a markdown table with columns X, Y, Z." |
| Bullet list | "Answer as a bulleted list, one idea per bullet." |
| JSON | "Output valid JSON only, matching this schema: {…}. No prose, no fences." |
| Step-by-step | "Give numbered steps I can follow in order." |
| Cited | "Cite every claim with a link and date. If you can't cite it, say so." |
| Language / register | "Answer in [French], formal register (vous)." |
| Length | "Under 150 words" / "exactly 5 sentences" / "one paragraph." |
Best for: reusing a good prompt across contexts. Keep the core prompt, swap the last line to change the shape of the output. Full walkthrough of the JSON case in the structured output guide.
5. Copy-paste starter prompts
Six ready-to-run prompts covering the common jobs. Edit the bracketed parts and go. For a coding-specific set, see Mistral prompts for coding.
Research (web search)
You are a research analyst. Search the web and give me a briefing on [TOPIC] as of this month. Cover: current state, the 3 biggest developments in the last 6 months, key players, and open questions. Cite every claim with a link and date. Output a bulleted brief under 400 words, then a one-line bottom line.Best for: current, sourced facts instead of an unsourced guess.
Coding
You are a senior [LANGUAGE] engineer. Here is my code: [PASTE]. Task: find the bug causing [SYMPTOM], explain the root cause in 2 sentences, then give the corrected code. Think step by step before answering. Output the fix as a single code block, then a short note on how to prevent it.Structured extraction (JSON)
You are an extraction function. From the text, return valid JSON only — no prose, no fences — using exactly: {"name": string, "email": string, "intent": one of ["demo","pricing","support","other"]}. Use null for missing fields, do not invent values, and if the text is empty return {}. Text: [PASTE TEXT]Translation
Translate the text below into [TARGET LANGUAGE] for a [AUDIENCE]. Localize idioms, units, and tone so it reads natively — do not translate word-for-word. Keep names and product terms exact. Return the translation, then a one-line note on any choice you made. Text: [PASTE TEXT]Writing
You are a writer for [AUDIENCE]. Write a [LENGTH]-word [FORMAT] on [TOPIC] with this angle: [ANGLE]. Lead each section with the takeaway, use short paragraphs and concrete examples, and avoid clichés. Draft it in Canvas so I can edit in place. Output clean Markdown with H2 sections.Business analysis
You are a pragmatic strategy advisor for a [COMPANY TYPE]. We are deciding whether to [DECISION]. Context: [KEY FACTS]. Use Work mode to reason it through, lay out the options and trade-offs, name the biggest risk, and recommend a move. Be blunt about weaknesses. Output a one-page brief: recommendation, rationale, risks, next step.Best for: turning a vague decision into a clear, defensible recommendation.
Frequently Asked Questions
What is the basic Mistral prompt formula?
Role + Task + Context + Constraints + Output format. Tell Mistral who to be, what to do, the background it needs, any limits, and exactly how to return the answer. Mistral rewards concision, so make every sentence earn its place, and naming the output format is the single biggest quality lever.
Which Mistral model should I use?
Mistral Medium 3.5 is the unified flagship for chat, reasoning, and code, with a 256K context. Use Codestral and Devstral 2 for coding-specific work, the open Small models for cheap high-volume tasks, and switch Le Chat to Work mode when a task needs multi-step reasoning.
What is Work mode in Le Chat?
Work mode is Le Chat's setting for complex, multi-step tasks. It routes you to Medium 3.5 with extended reasoning enabled and default-on connectors, so it plans before answering. Use it for research, long document drafting, and structured analysis.
How do I force JSON output?
Name the exact schema with keys and types, say "return valid JSON only — no prose, no fences", use null for missing fields, and give a fallback for empty input. In the API, also enable JSON mode or pass a schema so the format is enforced at decoding.
Does Le Chat have image generation and web search?
Yes. Le Chat includes web search with citations, image generation powered by Black Forest Labs' Flux, a Canvas workspace, a sandboxed Python code interpreter, document OCR, memory, and a no-code Agent Builder — much of it on the free tier.
What is Vibe?
Vibe is Mistral's coding-agent stack: a Vibe CLI plus asynchronous cloud agents that handle routine tasks like bug fixes in isolated sandboxes, with GitHub and Slack integrations. You give an agent a task spec and it works, then reports back.
Is Mistral good for non-English work?
Very. Multilingual coverage is a core Mistral strength, especially across European languages, and the models can match a native register rather than a translated-English tone. Name the language and register in the prompt for best results.