Six months into a company-wide subscription, two groups appear. A handful of people for whom AI has become a daily tool, and everyone else, who open it occasionally and leave disappointed. The difference is not technical. It is five habits, and all five fit in one sentence each.
They start from the company's documents
Not from a blank page
They ask for one clear task
With a format and a length
They check the sources
Especially figures and dates
They give feedback in the same chat
They do not open a new one
They save what worked
As an instruction or a template
What was saved becomes the starting point next time
Why the fourth one matters
The usual reflex, when the first answer is poor, is to close the chat and start again with different wording. That is precisely the wrong move: you throw away everything the model had worked out and begin from nothing. Correcting inside the same conversation — “too long”, “drop the pricing part”, “write it for a client, not for a colleague” — gets there in three steps, where rephrasing never gets there at all.
The fifth is the one that compounds
Once you finally land the right shape for a quote, a complaint reply or a meeting summary, that shape has to leave the conversation. Saved as a standing instruction or a template inside the project, it becomes the starting point for everything that follows. Unsaved, it is gone the same day.
Prompts do not accumulate. What you saved out of them does.
Terms used here
- Correction loop
- The request — answer — note — new answer sequence, carried out inside one conversation.
- Template
- A validated shape, kept in the project, that later requests of the same kind start from.
Where to start
- 1Pick one task you do weekly and do it only with AI for a month.
- 2Always correct inside the same conversation. Do not open a new one.
- 3At the end, save the good version into the project and start from there next month.
