I've been experimenting with Le Chat (Mistral's Large) extensively over the last few weeks, primarily using it to draft campaign logic and troubleshoot API sync issues between platforms. For the most part, it's fantastic—especially for brainstorming segmentation rules or parsing documentation. However, I've hit a consistent and frankly puzzling snag that's impacting my trust in using it for precise automation design.
It seems to occasionally, and quite randomly, ignore direct, unambiguous instructions given in the same prompt. For example, I'll structure a request with very clear formatting rules:
"Please outline the steps for a lead scoring workflow. Use a numbered list for the main steps, and for each step, **do not** provide sub-bullets or examples—just the concise step. Also, format the final output in a table with two columns: 'Step' and 'Owner.'"
What I'll often get back is a beautifully formatted table... but with sub-bullets and examples under every single step, which I explicitly forbade. It's like it processed the "make a table" part perfectly but completely dropped the "do not provide sub-bullets" constraint. This isn't a one-off. I've seen it with:
* Ignoring "do not use technical jargon" when explaining a process.
* Providing Python code when I said "provide pseudocode only, no actual syntax."
* Adding a summary section after I explicitly stated "do not include a summary or conclusion."
This feels less like a comprehension problem and more like a context management or priority bug. The model appears to understand the core task (build a workflow, explain a concept) but sometimes fails to adhere to the *meta-instructions* that govern *how* to execute that task. It's as if those constraints get lost in a later layer of processing.
I'm curious if others in the marketing automation or CRM integration space have run into similar issues. Are you seeing this primarily with formatting instructions, or also with substantive constraints (like "do not include any mention of [specific tool]")? I'm trying to determine if there's a pattern—like it happening more with longer prompts or when the instructions are placed at the beginning versus the end.
My workaround has been to reiterate the critical constraint in a separate sentence at the very end of the prompt, which helps but shouldn't be necessary. It makes me hesitant to use it for generating final draft documentation or specs that need to follow strict guidelines.
— Emma
If it's not measurable, it's not marketing.
Oh, I've run into something like this too, and it's maddening when you're trying to get a clean output for documentation. For me, it often happens when the instruction is phrased as a "do not."
I've had better luck flipping it to a positive, direct command. Instead of "do not provide sub-bullets," I'll try something like "Provide each step as a single, concise sentence only." It seems to latch onto the affirmative action more reliably.
There's also a chance it's getting tripped up by conflicting formatting cues in the same prompt. Asking for a table is a strong structural request, and sometimes the model appears to prioritize that structure over the stylistic "don'ts" buried in the text. Have you tried breaking the prompt into two separate messages? One to set the strict rules, then a follow-up for the actual task?
Keep it simple.
You're spot on about flipping "do not" into positive commands. In my experience, models are trained on patterns of "do this," making them far more reliable at following affirmative patterns.
> it's getting tripped up by conflicting formatting cues
That's a great observation and likely the core of it. The model has to reconcile multiple, sometimes competing, instructions within a single prompt window. The stronger structural cue (like a table) often wins. Breaking the prompt into two messages is a solid workaround we use for complex CI/CD pipeline generation - one message defines the strict YAML structure and output rules, the second contains the specific logic to implement. It significantly improves compliance.
I've found the order of instructions matters too. Putting the most critical "must do" item last, right before the task description, sometimes helps it stick.