Prompt library
Ready-made briefs for decision memos, difficult email and meeting notes.
Read the guide →By task
Work and productivity
These are the tasks where AI saves the most time and where a quiet error is least likely to be caught, because nobody re-reads the meeting notes.
The rule that prevents most bad outcomes
For workplace tasks the dangerous failure is not a wrong answer. It is an invented one that nobody checks: a decision recorded that was never made, an owner assigned who never agreed, a deadline that came from nowhere.
The instruction that prevents it is short. Tell the model what to do when information is missing, because the default is to produce something plausible.
Turn these notes into a decision record.
[paste notes]
1. Decisions made — only what was actually decided, and by whom
2. Open questions — discussed without resolution
3. Actions — task, owner, deadline
If an owner or deadline was never stated, write "not assigned". Do not infer one.
Do not summarise discussion that led nowhere.
Quote the line that supports each decision.“Not assigned” is the whole trick. Without an explicit fallback, a model asked for owners and deadlines will supply them, because that is what was requested. Give it a permitted way to return nothing.
Before you paste anything
This is the most common real-world incident in professional AI use, and it has nothing to do with prompting. Customer names, employee matters, health details, contract terms and unreleased figures all carry obligations the moment they enter a third-party tool.
Use a business account
Consumer tiers may use your input to improve models. Business and enterprise agreements normally do not, but check the specific terms rather than the marketing page.
Know that training and retention are different questions
“Not used for training” says nothing about how long data is kept, where, or who can see it during a support case.
Strip identifiers when you can
Most analysis works just as well on “Customer A” as on a real name. Anonymise first and the question mostly disappears.
Know what your organisation has approved
Personal accounts used for work, and integrations connected without oversight, are where this goes wrong quietly.
Planning, and the sycophancy problem
Asking a model to review your plan tends to produce encouragement. These systems are tuned toward agreeable, helpful-sounding output, which is exactly wrong when you want the flaw found. The fix is to make disagreement the assigned task rather than an option.
Here is my plan. [plan]
Do not tell me what is good about it.
1. What is the strongest argument that this fails?
2. Which assumption, if wrong, breaks the most?
3. What would a sceptical colleague ask that I have not answered?
4. What is the cheapest test that would tell me early whether this is wrong?
Be specific and concrete. Vague concerns are not useful.Note that this is a task definition, not a personality trick. Community folklore is full of clever phrasings claimed to remove agreeableness, and we have not found rigorous evidence for any of them. Assigning the critical job directly is what reliably works.
Personal data in AI tools
The practical version, not the policy document.
Read the guide →AI literacy for staff
What the EU expects organisations to teach people who use AI.
Read the guide →