Controlled generation
Emails, summaries, rewrites.
What is a prompt? The instruction you give a model
A prompt is the text (or multimodal message) given to a generative model to steer its answer: role, task, constraints, examples, and context. With LLMs, the prompt partly replaces procedural code. Google and OpenAI document design strategies; OWASP notes that poorly isolated prompts enable injection.
In one sentence
The prompt is the instruction—and context—you give the model.
Key points
- It combines goal, output format, constraints, and optional examples.
- System / user / tool messages structure the API dialogue.
- A good prompt reduces ambiguity; it does not replace reliable data (RAG).
- User input must be treated as untrusted (injection).
Term at a glance
- Prompt
- LLM instruction · Model prompt
- English term
- Prompt
- Domain
- Artificial intelligence
- Category
- Generative AI
- Level
- Beginner
What does “prompt” mean exactly?
With instruction-tuned models, the prompt is the main interface: “Act as…”, “Answer in JSON…”, “Cite only the context…”.
We often split system prompts (durable policy) and user messages (request). RAG context is injected as material, not unchecked absolute truth.
For an SME, standardizing a few prompts (ticket summary, email draft, extraction) stops every employee inventing risky instructions.
How do you write an effective prompt?
- 01
Clarify the task
One action, one audience, one success criterion.
- 02
Specify the format
JSON, bullets, length, language.
- 03
Add constraints
Forbidden topics, allowed sources, tone, confidence.
- 04
Test and version
Example suites, A/B formulations.
A concrete prompt example
A support team uses: “You are a tier-1 agent. From the ticket and provided FAQ, draft a Canadian French reply, max 120 words, and list points not covered by the FAQ. Do not invent any policy.” Human edit rates drop because the model knows when to abstain.
What are prompts used for?
Extraction
Structured fields from free text.
Classification
Route a request to the right queue.
Agents
Decide which tool to call next.
Benefits and limits of prompting
- Fast iteration without heavy redeploy
- Accessible to guided non-developers
- Adaptable per use case
- Complements RAG and tools
- Sensitive to wording
- Can be bypassed (injection)
- Not 100% deterministic
- Hard to maintain without versioning
How does prompting differ from fine-tuning?
| Prompt | Fine-tuning | |
|---|---|---|
| Change | Instructions at inference | Model weights adapted |
| Upfront cost | Low | Higher (data + training) |
| Flexibility | Immediate change | Retrain cycle |
| Best fit | Varied tasks, textual policies | Very stable style/format at scale |
Why prompts matter for Quebec SMEs
They are the fastest lever to get LLM value—and the first error surface. Shared, tested, secured prompts prevent leaks and invented answers in front of customers.
Frequently asked questions
Is a longer prompt better?
Not necessarily. Clear, structured, with useful context beats a redundant wall of text.
Do we need few-shot examples?
Often yes for format and edge cases; 2–5 well-chosen examples frequently suffice.
How do we avoid injection?
Separate instructions from user data, delimit context, restrict tools.
Who writes prompts at work?
Ideally product + business + tech, reviewed like code.
Related terms
Sources and references
Unstable or risky model answers? We structure, test, and secure prompts for your business cases.
Improve your prompts