1.4Common prompt mistakes
The essentials in 30 seconds
- Five mistakes explain disappointing answers: vague request, multiple request, missing context, unspecified format, leading question.
- The model is sycophantic by default: “my plan is good, right?” gets a yes — demand critique and counter-arguments.
- Everyone has THEIR recurring mistake: audit your recent prompts to find your point of vigilance.
Threaded mini-case — Amine audits his prompts before the steering committee
Amine is preparing a steering committee and rereads his week's prompts: a vague request (“tell me about risk management”), four questions crammed into one message, a missing client context that the AI filled with assumptions, and above all a leading question (“my new schedule is solid, right?”) that earned an enthusiastic yes. He fixes each one with the level's method, then asks: “Critique this schedule like a sceptical client: 3 major weaknesses”. Final episode: his art of prompting is complete.
Final episode of cases 1.1-1.3 — Amine hunts his recurring mistakes.
1. Five recurring mistakes
Five mistakes explain the vast majority of disappointing answers: the too-vague request (“tell me about management”), the multiple request (five questions in one), the missing context (the AI fills the gaps with assumptions), the unspecified format (essay received, bullets expected) and the leading question that dictates its answer to a naturally sycophantic model.
Amine finds all five in his own week: diagnosis starts by naming the mistake, like a doctor names the symptom.
2. The fixes
For each mistake, a fix: vague → specify task, angle and use; multiple → split, one question per message; missing context → give the situation and the audience; absent format → impose length, structure, tone; leading question → phrase neutrally or ask for both pros AND cons.
Amine splits his four questions into four messages and rephrases his vague request as “5 techniques usable by an overwhelmed project manager, one implementation sentence each”.
3. Anti-sycophancy
The model is sycophantic by default (it goes along with you): “my plan is good, right?” gets a yes. Demand critique for sound decisions: “Critique this idea like a sceptical investor: 3 major weaknesses”.
Amine submits his schedule to this stress test before the committee: the three revealed weaknesses are worth more than a polite yes followed by failure in the meeting.
🔎 Concrete example — 🔎 Concrete example — Anti-sycophancy
The most important technique in business. Test both wordings.
“My new schedule is solid, right?” → the AI says yes (sycophancy).
“Critique this schedule like a sceptical client: 3 major weaknesses, then suggest one fix per weakness.”
→ you finally get analysis useful for the decision.Reading: the leading wording dictates its answer to the model; the critical wording frees it.
Lesson: audit your own prompts — everyone has a recurring mistake.
Section vocabulary
- Vague request
- No goal or angle: the AI produces a lukewarm encyclopedia. Fix: specify task + intended use.
Simple example: Amine's “tell me about risk management” before correction.
What it is not: “5 techniques usable by an overwhelmed project manager, one sentence each”.
Reflex: always task + angle + use.
- Multiple request
- Several questions in one prompt: superficial answers to each. Fix: split.
Simple example: Amine's four questions crammed into one message.
What it is not: one question per message, handled one by one.
Reflex: one question = one message.
- Sycophancy
- The model's tendency to go along with you: “it's a good idea, right?” gets a yes.
Simple example: The enthusiastic yes earned by Amine's schedule before the stress test.
What it is not: explicit critique requested (“3 major weaknesses”).
Reflex: demand critique for every decision.
- Leading question
- Wording that contains the expected answer: it biases the model.
Simple example: “My new schedule is solid, right?” which dictates the yes.
What it is not: phrase neutrally, or ask for both pros AND cons.
Reflex: neutral + systematic counter-arguments.
In practice — In practice — Failed-prompt clinic
- Test the vague one: “Tell me about time management” — then the fix: “Give me 5 time-management techniques usable by an overwhelmed manager, one implementation sentence each”.
- Test the multiple one: ask 4 questions in one message, note the superficiality — then ask them one by one.
- Test sycophancy: describe an average idea and ask “it's a good idea, right?” — then: “Critique this idea like a sceptical investor: 3 major weaknesses”.
- Audit your 5 most recent real prompts: which mistake recurs most for you? That is your point of vigilance.
Ready-to-say script — share anti-sycophancy in a meeting (30 s)
“Careful: if we ask the AI whether our plan is good, it will say yes by default. Let's ask it instead: critique this plan like a sceptical client, three major weaknesses with one fix each. Then we will decide with full knowledge.”
Say it before any AI-assisted decision; note the weaknesses on the board.
Copy-paste model — prompt audit check-list
MY PROMPT: _______________
VAGUE? [ ] → task + angle + use: _______________
MULTIPLE? [ ] → split into: _______________
MISSING CONTEXT? [ ] → situation + audience: _______________
ABSENT FORMAT? [ ] → length + structure + tone: _______________
LEADING? [ ] → neutral version + “critique: 3 weaknesses”: _______________Decision-maker in a hurry
Top priority: anti-sycophancy. Before any decision, demand critique and counter-arguments — it is the fix that avoids costly mistakes.
Project manager in Algiers
Amine posts the check-list in the project space: every AI-assisted deliverable passes the five-mistake audit before the committee.
Off-target answer
If the answer misses the target with no visible mistake: add “If my request is ambiguous, ask me your questions first before answering” — many assistants then play along.
Magic-formula seeker
No universal perfect prompt: what is universal is the method — RCTF (1.1), iteration (1.2), examples (1.3), critique (here).
✓ Best practices
- Specify and split.
- Context, or it gets invented.
- Demand critique and counter-arguments.
✗ Common pitfalls
- The vague request.
- Several questions in one.
- The leading question.
Key takeaways
- Vague → specify task, angle and use; multiple → split.
- The AI fills gaps with assumptions: give the context or it will invent it.
- The model is sycophantic by default: demand critique and counter-arguments.
- Audit your own prompts: everyone has THEIR recurring mistake.
48h challenge — apply tomorrow
Audit your 5 most recent real prompts with the check-list above, note your recurring mistake and fix one of them with its fix. 20 minutes is enough.
- I can name the 5 mistakes and their fixes.
- I know how to practise anti-sycophancy.
- I audited my prompts + identified my recurring mistake.
Frequently asked questions
Why doesn't the AI tell me my request is ambiguous?
Models are trained to answer rather than to question. Tip: add “If my request is ambiguous, ask me your questions first before answering” — many assistants then play along.
Is there a universal perfect prompt?
No: the right prompt depends on the task and the context. What is universal is the method — RCTF, examples, iteration, critique — which you have just acquired in this level.
Which technique matters most against AI sycophancy?