Most people open Le Chat, ask one question, read the answer, close the tab. That is maybe a tenth of what the assistant does now. Mistral’s chat product has grown into a small workspace: Projects that remember your context, a Canvas you can edit inside, a Python sandbox, web search with citations, image generation. Used in the right order, those pieces cut a surprising amount of busywork.
Here is the workflow I actually run, step by step, with the settings and prompts doing the heavy lifting. Setup takes about fifteen minutes. After that, most tasks are a twenty-second job.
Pick your model before you pick your words
The model selector sits next to the prompt box, and it changes the answer more than any wording trick you’ll find on a listicle. Three options cover almost everything:
- Mistral’s general models for writing, summarising, brainstorming and rewriting. Fast, and good enough for most tasks.
- Magistral, the reasoning model for anything with multiple steps: debugging, comparing options against criteria, working through a budget.
- Codestral when the prompt is mostly code and you want language-aware answers rather than prose.
An example. “Summarise this six-page supplier contract into eight bullets” runs in about two seconds on a general model. “Which of these three suppliers is cheapest once you factor in the 4% early-payment discount and the $1,200 setup fee?” belongs on a reasoning model, because the arithmetic and the comparison both have to hold together. For everyday work, Magistral is more than enough; you don’t need to reach for the heaviest option out of habit.
Build a Project instead of re-explaining yourself
Projects are the feature people miss. A Project stores custom instructions and uploaded files, and every chat inside it inherits both. You stop pasting your brand voice, client brief or product spec into every new conversation.
What to put in one
- Custom instructions: your role, the tone, the output format, and hard rules such as “always lead with the answer” or “never use exclamation marks”.
- Reference files: a style guide, the last two quarters of customer emails, a spreadsheet of product specs.
- One line describing what a finished answer looks like. It sounds fussy. It works.
A support team I know built one Project per product line, dropped in real customer threads, and added the instruction “Reply in our help-centre voice, 120 words maximum, finish with the exact next step.” Drafting a first reply went from roughly six minutes to under one, and the edits were mostly names and numbers.
Turn on the right toggle for the job
Le Chat’s toggles aren’t decoration. Flipping the wrong one is the fastest way to get a worse answer.
- Web search for anything with a date attached. It’s a big part of why Le Chat feels so quick at answering current questions, and it cites sources you can check. Switch it off for creative work; it drags the tone towards press-release blandness.
- Canvas when you’ll edit the output more than once. Drafts and scripts live in a side panel where you request targeted changes instead of regenerating the whole thing.
- Code interpreter for CSV cleanup, quick charts and date arithmetic. It runs Python in a sandbox and shows you the result, which beats trusting a language model to add up a column.
- Image generation for moodboards and social mockups. Skip it for anything with real text in it, because the text comes out mangled.
Write prompts like a brief, not a wish
The gap between a passable answer and a usable one is usually one extra sentence of context. Compare these two.
Weak: “Write a post about our new pricing.”
Usable: “Write a 900-word post announcing our new $29 Pro plan. Audience: free users who keep hitting the daily message limit. Open with the problem, then three bullets on what changes, then a short paragraph on how to upgrade. Avoid hype words. Give me three headline options at the end.”
The second prompt takes an extra forty seconds to type and saves two rounds of rewriting. A decent rule of thumb: name four things every time. Audience, length or format, what to avoid, and what the finished piece should let the reader do.
Ask for structure, not just content
Vague formats produce vague documents. Say “return a markdown table with columns for supplier, unit cost and lead time” and you get something you can paste into a doc without reformatting. If the output feeds another tool, ask for JSON with named keys. Le Chat follows that reliably, and it removes a layer of manual tidying from your afternoon.
Use Canvas for the second draft
First drafts belong in the chat. Second drafts belong in Canvas. Open a Canvas, generate or paste the draft, then make small requests: “tighten paragraph three, keep the example”, “turn the second list into a table”, “cut 200 words without losing the pricing detail”. Each change lands in place, so you finish with one clean document rather than five chat messages you have to stitch together by hand.
The same applies to code. Ask for a Python script that reads a folder of CSVs and writes one combined file with a source column, send it to the code interpreter, run it, and paste the error back if it breaks. Three rounds of that beats guessing at a fix in the dark.
Know when to leave Le Chat
Le Chat is hosted, which is fine for everyday work and awkward for anything under NDA or sitting in a regulated folder. If files can’t leave your machine, a local setup covers similar ground. AnythingLLM lets you chat with your own documents without sending them to the cloud, and Jan AI runs a ChatGPT-style assistant entirely on your laptop. If you’d rather keep a familiar interface pointed at models you control, Open WebUI is a self-hosted front end that behaves like the commercial tools. For coding specifically, an editor-native helper such as Continue.dev keeps suggestions inside your IDE, where the surrounding files are visible. Swapping tools mid-project is rarely worth it, but knowing the exit route matters when legal asks where your data went.
A daily loop that makes this stick
Feature lists don’t change habits; routines do. This is the pattern that survived contact with my own calendar:
- Morning, five minutes. Open your main Project and dump the day’s messy notes in. Ask for a prioritised list with rough time estimates.
- Through the day. One chat per task, in the right Project, with the right toggle on. When an answer is nearly right, open Canvas rather than regenerating from scratch.
- Evening, three minutes. Paste the day’s decisions back into the same Project and ask for five bullets for tomorrow. Those bullets become your opening prompt in the morning.
Two things separate people who get real value here from people who don’t: Projects, so context is never retyped, and Canvas, so a good draft never dies in a bad regeneration. Set up one Project today with three files and two rules in the instructions. You’ll notice the difference by tomorrow afternoon, and you’ll stop opening a blank chat window out of reflex.