Français

Applied AI — Professional prompting

Intermediate Level · Session 2 · Yann Isola

Interactive playground — works offline

Workshop: write a prompt, the simulator evaluates it

Write a prompt below. The simulator analyzes your specification (the 6 blocks + guardrails), assigns a score, then simulated the response of a model: the more precise your prompt, the better the simulated output.

Reminder : a prompt is not a question, it's a specification . Role · Context · Task · Constraints · Format · Examples — plus the anti-invention guardrail and data compartmentalization.

Temperature simulator

The temperature regulates the chance in the choice of the next token. Same prompt, different setting → different outputs. Test.

Suggests a slogan for a home-to-work carpooling application.
0
determinist 1,2
very creative
T = 0.0
To observe: at T = 0, click “Generate” several times — the output does not change (almost) not. Turn up the heat: variety appears… and so do the slip-ups.

The trap to remember

Deterministic ≠ correct. Temperature 0 makes the output reproducible , not true . If the model errs at T = 0, it will err the same way each time it runs. Reliability comes from verification (see Matrix tab), never from chance adjustment.

Usage: extraction, classification, strict formats → low temperature. Brainstorming, creative variations → higher temperature.

The 2×2 trust matrix

Cost of a mistake not detected × ease of verification . Drag each case to the correct quadrant (or tap a case, then tap a quadrant), then click "Correct."

Cost of undetected error → high (high) / low (low)

🟡 Leverage area

High cost · easy verification — AI produces, human validates

🔴 Prohibited area

High cost · difficult verification — expert required or abstention

🟢 Free zone

Low cost · easy verification — frictionless use

🟢/🟡 Acceptable area

Low cost · difficult verification — invisible error costs nothing
Verification: easy (left) ←→ difficult (right)
Key lesson: it's almost never there stain which determines the zone, it is the verification process which surrounds it. “Contractual clause” is in a prohibited area without legal proofreading, and in leverage zone with .

Comparison: fuzzy prompt vs specification

Choose a scenario. Look side by side at the fuzzy prompt and its specified version (highlighted blocks), then at the corresponding simulated outputs.

❌ Prompt blur

Typical output:

✅ Specification

role/context constraints / format anti-invention safeguard delimiters (data)

Typical output:

To remember: everything you don't write, the model invents for you. And for tone: examples beat descriptions — showing a typical output is better than three adjectives.