6.3Good practices & pitfalls
The essentials in 30 seconds
- Golden rule: AI multiplies those who can steer and check it — framing, checking, protecting and keeping the human beat speed.
- Do: tool by data, RCTF prompts, few-shot example, required criticism, checking to the level of the stakes.
- Never do: paste sensitive content into a public tool, publish without checking, let AI decide, imitate a protected style.
Amine thread — three habits that pay
After one year, Amine reviews: what counted most was not a model, but three habits — writing in RCTF with an example, requiring criticism against sycophancy, and checking before signing. And his biggest narrowly avoided pitfall: pasting a sensitive document into an unapproved tool on a rush day.
Pick a single habit to anchor in your ritual, then move to the next.
AI-at-work dos and don'ts
The difference between use that multiplies and use that exposes comes down to a few habits. Here are the ones that come up most.
✓ Do
- Pick the tool by data: company → approved tool.
- Write RCTF prompts (role, context, task, format) and iterate.
- Show an example (few-shot) for formats and style.
- Require criticism and counter-arguments (anti-sycophancy).
- Check proportionately to the stakes: whoever signs checks.
- Protect data: classify, anonymise, if in doubt do not paste.
- Keep the human on decisions and sending (not just preparation).
- Build up: prompt library, custom assistants, charter.
✗ Avoid
- Believing a false answer because it is well worded.
- Pasting sensitive data into an unapproved tool.
- Taking AI sycophancy for approval.
- Publishing without checking figures, quotes and references.
- Imitating a living artist's style in commercial use.
- Letting AI decide (hiring, sanction, appraisal).
- Multiplying catch-all threads (one topic = one thread).
- Waiting for “the perfect model” to lower your guard on checking.
Ready-to-say script — checking your use before sending (30 s)
“Approved tool for this data? RCTF prompt with example? Criticism asked against sycophancy? Facts checked to the level of the stakes? Human on decisions and sending? If any no, I fix before sending.”
Copy-paste template — monthly self-check (8 questions)
THIS MONTH (yes / to improve):
1. Tool picked by data (approved if company)?
2. RCTF prompts and systematic iteration?
3. Example provided (few-shot) for formats and style?
4. Criticism required against sycophancy?
5. Proportionate checking before signing?
6. Classified / anonymised data, doubt = no pasting?
7. Human on decisions and sending?
8. Prompt library and charter fed?
MY PRIORITY NEXT MONTH (only one): ...Small business / tight budget
One habit per quarter is enough: if in doubt do not paste first, the rest after.
SME in Algeria
Adapt rituals to peaks and travel: one standard-prompts and charter slot kept each week.
Remote / hybrid
The no. 1 discipline at a distance: one topic = one thread, written checking before any sending.
Intercultural
Check local practices rather than assume: AI mentions and usage rights vary.
48h challenge — apply tomorrow
Fill in the monthly self-check above and pick a single priority to anchor (e.g. require criticism on each file).
- I know the do reflexes and the pitfalls to avoid.
- I filled in the self-check and picked a priority.
- I did the 48h challenge or planned its date.
Before sending, Amine checks his use. Which reflex best protects his work?
Golden rule: AI multiplies those who can steer and check it — framing the prompt, checking the facts, protecting the data, keeping the human on decisions beat mere speed.