6.4Toolbox
The essentials in 30 seconds
- Before acting: classified data, RCTF prompt, example, requested criticism, set checking level — otherwise do not launch.
- Markers: an LLM predicts plausible text, one example beats ten lines, whoever signs checks.
- Rule: a prompt library and a charter at hand beat a method known by heart.
Amine thread — the sheets at hand
After one year, both sheets below are posted at Amine's place: the pre-use one next to the screen, the markers one in his drawer with his prompt library. “What saved me, he says in substance, is not knowing the points, it is ticking them each time — especially in rush weeks.”
Adapt the wordings to your context and charter, keep the structure: it is what avoids omissions and incidents.
Ready-to-use markers
A good checklist makes a task fast and omission-free; good markers speed up decisions. To adapt to your context and your organisation's charter.
Before using AI on a task
- Is there company data? (→ approved tool)
- Does my prompt have Role, Context, Task, Format?
- Do I have an example to show (few-shot)?
- Did I ask for criticism (anti-sycophancy)?
- Which checking level: re-read / critical points / source everything?
- Are the facts (figures, quotes, references) checked?
- Is the data classified / anonymised?
- Does the human keep decisions and sending?
Markers & reflexes to deploy
- An LLM predicts plausible text: check the facts
- RCTF: Role + Context + Task + Format (format = most rewarding)
- One example > ten lines of description (few-shot)
- “Criticise like a sceptic” against sycophancy
- Facts/news = assistant with web search + cited sources
- Company data → approved tool; if in doubt, do not paste
- Whoever signs checks; “it's the AI” is no defence
- AI never makes an individual decision (GDPR/AI Act)
Ready-to-say script — launching an AI task properly (30 s)
“Classified data and approved tool if company. RCTF prompt with example. Criticism asked. Checking level set. Facts to check noted. Human on decisions and sending. Launching.”
Copy-paste template — AI task follow-up (1 line per point)
TASK of ...: tool by data (approved if company): OK / to redo
RCTF PROMPT (role, context, task, format): OK / to redo
EXAMPLE provided (few-shot): OK / to redo
CRITICISM asked (anti-sycophancy): OK / to redo
CHECKING LEVEL set: OK / to redo
FACTS checked (figures, quotes, references): OK / to redo
DATA classified / anonymised: OK / to redo
Human DECISIONS and sending: OK / to redoSmall business / tight budget
One-page sheets posted on the wall: data, RCTF, checking and the human deciding suffice.
SME in Algeria
Plan local realities into the sheet: approved tool, team charter, professional or DPO to consult.
Remote / hybrid
Add a written-confirmation line to the follow-up: at a distance, dated checking avoids misunderstandings.
Intercultural
Spell out the headings in writing: what seems obvious (approved tool, checking level) is not obvious to all.
48h challenge — apply tomorrow
Take 1 real task, run the “before using AI” checklist and tick where it stands. File your best prompt in the library.
- I ticked a checklist on a real task.
- I can deploy the pre-use sheet and the markers at the right time.
- I did the 48h challenge or planned its date.
Amine is about to use AI on a file with internal data. What must he tick first?
Tip: keep your standard-prompts library and your team charter at hand; the tool landscape changes fast, the method stays stable.