Workshop: Content Creation
On this page
Workshop Objective
This final workshop explores the creative side of Mistral models: brainstorming, creating training scenarios, and generating marketing content. The paradox of AI-assisted creation is that it demands more structure, not less: high temperature without guardrails produces unusable chaos, while high temperature within a well-constrained prompt yields ideas that are both original and actionable. This combination—controlled freedom—is what you will practice in the three exercises below.
Exercise 1: Structured Brainstorming
Let’s start with a product case: you are preparing the launch of a new mobile app and need ideas for differentiating features. Raw brainstorming would yield a list of vague slogans; to be actionable in a product meeting, each idea must arrive pre-evaluated. Notice how the system prompt imposes an evaluation grid (effort, impact, differentiation) on each idea, turning the model into a product sheet generator rather than a suggestion machine.
messages = [
{
"role": "system",
"content": """You are a creative product manager specializing in B2C mobile apps.
For each idea, provide:
- Feature name (max 5 words)
- Description (2 sentences)
- Development effort: S (< 1 week) / M (1-4 weeks) / L (> 1 month)
- Estimated user impact: low / medium / high
- Competitive differentiation: low / medium / high
Propose varied ideas: some pragmatic, some bold."""
},
{
"role": "user",
"content": """Product context:
- Personal budget management app for 25-35 year olds in France
- Current features: expense tracking, automatic categorization,
savings goals, bank synchronization
- Main competitors: Bankin, Linxo, Finary
- Identified problem: 60% of users abandon after 3 weeks
Generate 8 feature ideas to improve retention.
Rank them by decreasing impact."""
}
]
response = client.chat.complete(
model="mistral-large-latest",
messages=messages,
temperature=0.8,
presence_penalty=0.6
)The model should produce 8 truly varied ideas: social features like friend challenges, gamification with streaks, badges, and levels, predictive intelligence like overdraft alerts, or original integrations such as group expense sharing. This diversity is no accident: the combination temperature=0.8 and presence_penalty=0.6 echoes exactly the lessons from the previous section—temperature opens the field of possibilities, presence penalty prevents the model from spinning out the same idea eight times.
Extend the exercise. Lower the temperature to 0.4 and compare: ideas become more conventional but more realistic. Add the constraint “At least 2 ideas must leverage generative AI” and observe how the model complies. Finally, change the persona: “You are a UX designer” will steer proposals toward experience rather than features.
Exercise 2: Creative Scenarios for Training
Change of scenery for the second exercise: you are creating scenarios for a management training course. A good training scenario must be realistic, engaging, and—this is the hardest part—allow for multiple legitimate approaches: if there’s an obvious answer, group discussion fizzles out in five minutes. The system prompt encodes these requirements one by one, down to the figures that make the dilemma concrete.
messages = [
{
"role": "system",
"content": """You are an instructional designer specializing in management training.
Each scenario must:
- Be grounded in a realistic French professional context (2026)
- Present a dilemma with no obvious solution
- Involve at least 3 stakeholders with divergent interests
- Include figures to make the decision concrete
- End with 3 reflection questions for the group
Tone: professional, immersive, engaging."""
},
{
"role": "user",
"content": """Create a training scenario on the theme:
"Managing a conflict between individual performance and collective well-being"
Context: a 12-person tech team in a Paris-based scale-up.
The scenario should last about 30 minutes of group discussion.
Also provide a facilitator sheet with key points to surface."""
}
]
response = client.chat.complete(
model="mistral-large-latest",
messages=messages,
temperature=0.7
)The expected result is a complete scenario: a situation described in detail with its figures (revenue, deadlines, turnover), at least three characters with their own motivations, a central dilemma with defensible arguments on both sides, open reflection questions with no single answer, and the facilitator sheet with key points to surface, pitfalls to avoid, and an evaluation grid. A temperature of 0.7 gives the model freedom to invent characters and situations, while the imposed structure ensures the scenario remains usable in the classroom.
Exercise 3: Multi-Format Marketing Content
The final exercise puts you in the shoes of a marketing manager: you are launching a new service and must produce content for several channels simultaneously, from a single brief. Each channel has its codes—a LinkedIn post isn’t written like an email subject line—and the system prompt enforces them, backed by numerical constraints.
messages = [
{
"role": "system",
"content": """You are a content director producing multi-channel marketing
material from a single brief.
For each channel, respect specific constraints:
- LinkedIn: 1200-1500 characters, expert tone, 3-5 hashtags, a punchy hook
- Email: subject (max 50 chars), preview text (max 90 chars), body (200 words max)
- Web page: headline (max 10 words), subtitle (max 25 words), 3 USP blocks,
main CTA
Style: professional French, formal 'you', data-driven where possible.
NEVER use empty superlatives ("market leader", "revolutionary solution")."""
},
{
"role": "user",
"content": """Product brief:
Name: SecureDocs AI
Type: Document management SaaS with integrated AI
Price: starting at 19 euros/user/month
Target: legal and compliance departments (50-500 employees)
USP: Automatic document classification by AI, risky clause detection,
integrated GDPR/NIS2 compliance
Launch: April 15, 2026
Launch offer: 3 months free for the first 100 signups
Produce content for the 3 channels."""
}
]
response = client.chat.complete(
model="mistral-large-latest",
messages=messages,
temperature=0.6
)You should get three distinct blocks, each in its channel’s codes. For LinkedIn: a professional post opened by an engaging hook—a question or a statistic—that states the problem, presents the solution, and ends with a CTA to the site, with relevant hashtags. For email: a short, intriguing subject, complementary preview text (not redundant with the subject), and a body structured problem/solution/offer/CTA. For the web page: a benefit-oriented headline, a subtitle specifying the target, three titled USP blocks, and a clear CTA.
Here too, variations are instructive: generate the same content with n=3 and temperature=0.7 to compare three versions and choose the best, add a channel (“Twitter/X Thread in 5 posts”), or change the target—“industrial SMEs” instead of “legal departments”—to see the entire marketing angle pivot.
The 3-Step Creative Pipeline
For high-quality content, best practice is not to ask for everything in a single call, but to chain steps adapting the temperature to each:
# Step 1: Brainstorming (high temperature)
idees = client.chat.complete(
model="mistral-large-latest",
messages=[{"role": "user", "content": "10 angles for an article on [subject]"}],
temperature=0.9
)
# Step 2: Selection and structuring (low temperature)
plan = client.chat.complete(
model="mistral-large-latest",
messages=[
{"role": "user", "content": f"Among these angles:\n{idees}\nChoose the best and create a detailed outline."}
],
temperature=0.3
)
# Step 3: Writing (medium temperature)
article = client.chat.complete(
model="mistral-large-latest",
messages=[
{"role": "system", "content": "Write according to this outline, tone [your tone]."},
{"role": "user", "content": plan}
],
temperature=0.5
)This pipeline separates divergent phases—brainstorming, where you want boldness—from convergent phases—structuring and writing, where you want rigor. This is exactly the approach of a human creative team: you don’t evaluate ideas while generating them, and you don’t write while structuring.
Key Takeaways
- High temperature (0.7-0.9) for brainstorming, medium (0.4-0.6) for writing
- Presence penalty encourages idea diversity
- Structure your creative prompts with precise constraints (format, length, tone)
- Use a multi-step pipeline and generate multiple versions (
n > 1) to select the best - Adapt the persona to the channel: expert for LinkedIn, sales-oriented for email, concise for web
Test Your Knowledge
After the four workshops, a final check of the techniques.
1. What is the purpose of the system role compared to the user role?
Answer: The system sets the enduring framework (identity, rules, format) for the entire conversation; the user carries the current requests—application instructions do not mix with messages.
2. When is few-shot preferable to instructions?
Answer: When showing is better than telling: precise formats, subtle tones, edge cases—one or two examples calibrate the output better than a paragraph of rules.
3. What does chain of thought change on a complex task?
Answer: Asking for reasoning before the conclusion makes steps unfold instead of jumping to a plausible answer—fewer errors on calculation, logic, and trade-offs.
4. Temperature and Top P: what setting for what need?
Answer: Low temperature for factual and extraction (determinism), high for creativity; Top P restricts sampling to the most probable tokens—you adjust one or the other, rarely both.
5. Name two prompting anti-patterns to avoid.
Answer: Asking for everything in one sentence without structure or output format, and stacking contradictory instructions—workshop lessons show the opposite: clear objective, clear constraints, iterations.
The four workshops have turned these techniques into habits—reuse their frameworks as templates for your own cases.