Writing good prompts
Give context, a clear task, material as a file and the format you want, improve the result in the same chat and save what works.
A prompt is your message to the model. Whatever you leave out, the model fills with generalities or plausible-sounding assumptions.
The five building blocks
| Building block | The question behind it | Example |
|---|---|---|
| Context | Who is writing, what is it about? | "The client is a joinery." |
| Task | What exactly should be produced? | "A reminder about missing documents." |
| Material | What should the model rely on? | Attach a file or activate a brain |
| Format | What should the result look like? | "At most 120 words, as a list." |
| Limits | What must not happen? | "Do not invent amounts. If something is missing, ask." |
Not every message needs all five. But when a result misses the mark, one of them is almost always missing.
In chats and projects the model already knows your name and, in a company workspace, the company name. Letters and mails are therefore signed correctly.
Example: a client email
Before
Write an email to Brunner about the missing documents.
What comes back is a generic email without a concrete list and with a deadline the model chose itself.
After
Write an email to Mr Marco Brunner, managing director of Holzwerk Brunner GmbH.
For the 2025 annual accounts we are still missing: the December bank statements, the inventory as at 31 December and the leasing contracts for the two vans. We need them by 15 January so the accounts are finished at the end of January.
Tone: friendly and committed. At most 120 words, the missing documents as a list.
Recipient, list, deadline with its reason and format are fixed. The model has nothing to guess.
Example: preparing a decision
Bäckerei Mehlwerk GmbH needs a delivery van for about CHF 58,000: a purchase with a bank loan or a lease over 48 months. I am preparing a meeting with the management.
First ask me up to three questions about information that is missing for a fair comparison. Then a table with the pros and cons of both options and a list of the assumptions I need to clarify. Only calculate with figures I give you.
The questions stop the model from inventing interest rates or residual values.
Keep working in the same chat
The first result is a draft. Short follow-ups are enough, the model knows the conversation: "Shorter, at most 80 words.", "We already have the inventory.", "Turn this into a Word file."
- If your first message already went wrong, correct it with Edit and resend instead of adding corrections.
- For a new topic, start a New chat. A long chat costs more credits with every message, and older parts are summarised at some point.
Material and format
- Files instead of pasted text. Attach documents via Add. The file stays available for the whole chat, tables and sections are kept, and Custodos edits Word, Excel and PowerPoint files directly. See Attaching files and images.
- A brain for recurring material. If you need the same documents again and again, they belong in a brain. See Building a good brain.
- Order the format precisely. "As a table with the columns client, deadline, status", "as an Excel file with formulas", "three subject lines of at most eight words". Words or sentences instead of "short".
- Make gaps visible. "Mark every statement that does not come from the attached documents."
Reuse good prompts
- Save as prompt. Under your message, choose Save as prompt and replace what changes each time with placeholders such as
{{Client}}. With Whole workspace the whole team can use it. See Saving and sharing prompts. - Response style. What you would write into almost every prompt, such as "short and factual", you enter once under Response style, your role under About me. See Personal settings.
- Agent. A task the team does regularly becomes an agent. See Building good agents.
Frequently asked questions
Custodos sets spelling and number format by your workspace's country: the billing address or, while there is none, the country chosen when the workspace was created. For Switzerland that means ss instead of Eszett, and 1'000. If no country is known, the model follows the way you write. If it still writes differently, add to your Response style, for example "ss instead of Eszett, keep the umlauts".
That is pseudonymisation: detected values go to the model as placeholders and are put back into the answer. Which values these are is set by your workspace's protection level. See Protecting personal data in chat.