6.3Best practices & pitfalls
The essentials in 30 seconds
- Constrain: palette, medium and framing locked — repetition makes a brand recognizable.
- Log: prompts, versions and parameters kept; generated content disclosed, especially realistic content (5.5).
- Measure: A/B performance (5.4), not volume — select rather than release.
Sonia thread — the criteria stuck on the wall
After one quarter, Sonia stuck four words on the wall in Algiers: constrain, log, verify, disclose. What counted most was not a tool but three habits — selecting instead of accumulating, starting from a validated image for video, checking terms and protected elements — and her biggest narrowly avoided pitfall: a lookalike face that nearly went out without documented consent.
Pick a single habit to anchor in your routine, then move to the next one.
What makes the difference
The difference between amateur use and professional practice is not about tools but reflexes: constrain, log, verify, disclose.
✓ To do
- Select rather than produce: volume serves sorting, not release.
- Constrain deliberately (palette, medium, framing): repetition makes a brand recognizable.
- Start from a validated image for video (image-to-video).
- Log prompts, versions and parameters (reproducibility + proof of process).
- Disclose generated content by default, especially realistic content (5.5).
- Measure performance (A/B), not the number of images.
✗ To avoid
- Letting everyone invent their own prompts (patchwork that dilutes the brand).
- Presenting an AI image as documentation of a reality (0.3).
- Using a real person's face/voice without documented consent (deepfake, 0.4).
- Believing fake detection is reliable and neglecting source verification.
- Pushing --stylize to the maximum and getting a "too AI" look.
- Forgetting to check commercial-use terms and protected elements.
Ready-to-say script — checking your practice before releasing (30 s)
"Selection done? Constraints respected (palette, medium, framing)? Prompts and versions logged? Content disclosed if realistic? Terms and protected elements checked? Documented consent if real face? If any no, I fix before releasing."
Copy-paste template — pro grid before release
Pro grid before release:
[ ] selection done, coherent family viewed TOGETHER
[ ] palette + medium + framing locked
[ ] prompts, version and parameters logged
[ ] generated content disclosed if realistic (5.5)
[ ] commercial-use terms + protected elements OK (0.4)
[ ] no face/voice without documented consentSmall business / tight budget
One habit per quarter is enough: constrain + sort first, the rest later.
Algiers SME
Adapt routines to client peaks: one sorting and traceability slot kept each week.
Remote / hybrid
The #1 discipline at a distance: written selection and disclosure, even for small deliveries.
Cross-cultural
Check local practices rather than assuming: an unusual face or symbol should raise a flag everywhere.
48h challenge — apply tomorrow
Screen your next delivery: zero release without sorting, zero face without consent, zero unverified terms. Fix the first gap found.
- I know the 6 reflexes to apply and the 6 pitfalls to avoid.
- I checked my 3 critical points (sorting, traceability, disclosure).
- I did the 48h challenge or scheduled its date.
Before delivering a realistic visual, which reflex best protects Sonia?
A good AI operator is not the one who generates the most, but the one whose hand you recognize and whom you can trust.