How do I build a small prompt library so my private AI works the way I do?
Short answer: Write five to eight reusable instructions with a role, a task, a format and a list of what to leave out, then test, version and store them on your own box.
Apply for a place From $10,000 setup, then $48,000 a year. Five places a week.
What is a prompt library and why keep one?
A prompt library is a short set of saved instructions you reuse for the jobs you repeat every week. You write each one once, test it, and then paste or select it instead of retyping it from memory.
The point is consistency. A model answers differently when you phrase a request differently. A saved instruction gives you the same shape of answer on Monday and on Friday, and it lets an assistant or colleague get the same result you do.
Five to eight instructions cover most of a busy person's week. More than that and you will stop using them.
Which five to eight instructions should you write first?
Start with the jobs you already do by hand. These are good first candidates:
- Meeting notes into a list of actions, with an owner and a due date for each one
- An email reply in your own tone, using two or three of your past emails as samples
- A summary of a long file that lists where in the file each claim came from
- A plain language rewrite of dense legal, financial or technical text
- A daily briefing built from your calendar notes and the documents you point it at
- A cleanup pass that fixes spelling and grammar and changes nothing else
How do you write one instruction that works?
Write each instruction with four parts: a role, a task, a format and a list of what to leave out. Short beats clever. If a person could not follow it, the model will not either.
The role says who the assistant is for this job, for example a careful chief of staff. The task says what to do with the text that follows. The format says what the answer looks like, such as a numbered list or three short paragraphs. The leave out list names the habits you dislike, such as adding facts that are not in the text, guessing a date, or using a cheerful tone.
Add one line telling the model to say so when the input does not contain what it needs. That single line cuts down on made up details, though it does not remove them.
How do you handle summaries and facts you must trust?
For any summary or answer you may rely on, ask for a source check inside the instruction. Tell the model to quote the passage each point comes from and to mark any point it cannot tie to the text.
Then you check. Open the file, find the passage and confirm it says what the summary says. The model can be wrong, and a summary can drop the one sentence that mattered. Our page on meeting notes without cloud transcription shows the same habit for recordings.
Where are the prompts stored and who may edit them?
On a private setup, store the library as plain text files or notes on your own box, in a folder only your login can open. That keeps your instructions, which often hold your tone and your client habits, in the cloud account you own.
Decide who may edit. A simple rule works well: you and one named person can change a prompt, and everyone else can only read and use it. Keep a copy in your normal backups. Our security and backups guide covers access limits and backup basics.
Do not put passwords, account numbers or other secrets inside a saved prompt. The prompt should say what to do, not carry the data.
How do you test and version a prompt?
Test each instruction on three to five real examples before you save it, and compare the answers with what you would have written yourself. Mark each answer as usable, needs edits or wrong, and change the instruction until most are usable.
For versions, give each file a name and a number, such as the job name and version two, and keep one line at the top that says what changed and why. Keep the old version until the new one has held up for a few weeks. Retest after you switch to a different model, because the same words can give a different answer. See how to choose an open model for how to compare models on your own work.
What does the Freeborn team set up, and what do you write?
If you hire Freeborn, a person sets up the private assistant in a cloud account you own, places a starter folder for your saved instructions on it, and trains you to use them. You can then keep the setup. Our private AI setup service page lists what is included.
You write the instructions that carry your voice, your priorities and your rules, because only you know them. Freeborn can help you shape the first few during training, but the content and the final review stay with you.
If you would rather build the whole thing yourself, the do it yourself guides show how.
What are the limits?
A prompt library makes answers more consistent. It does not make them correct, and it does not replace your review of anything you send, sign or rely on.
Models change, so an instruction that works today may need a rewrite later. A private setup lowers the number of parties who hold your data, but it does not remove the need for strong passwords, access limits and backups.
This is general information, not legal, medical, financial or tax advice. Ask your own counsel, compliance officer or adviser before you use an assistant on material that carries professional duties. More questions are answered in the private AI answers hub.
What it costs
A place costs a one-time setup from $10,000 (the setup price rises with each place sold), plus $48,000 a year in dues paid up front (or $4,800 a month). Compute is billed to you at cost. Hands-on training and on-call support are extra, by the hour.
Questions
How long should a saved prompt be?
As long as it needs to be and no longer. Most good ones fit in a short paragraph plus a few lines for format and exclusions.Can I share my prompts with an assistant or a colleague?
Yes, if you give them read access to the folder. Decide ahead of time who may change a prompt, so two versions do not drift apart.Should I put examples inside the instruction?
A couple of short examples of your own writing often help the model match your tone. Use samples that hold no private names or numbers.What if a prompt stops working after a model change?
Run your test examples again and adjust the wording. Keep the earlier version so you can compare and go back.Do I need a library if I only use the AI now and then?
Probably not at first. Save an instruction once you catch yourself retyping the same request a third time.Related guides
- Can you get AI meeting notes without sending audio to a cloud transcription service?
- How do you choose an open AI model (Llama, Qwen, Mistral and others) to run privately?
- AI for executives who have no time to set it up
- How can a CEO use AI on board materials, deals and people decisions without leaking them?
- How do you run Llama privately on your own server?
- Private AI answers
- Private uncensored AI: the full guide
- All guides to getting your own private AI agent
Apply for a place
Takes thirty seconds. No payment now.
Adults only (18+). Private uncensored AI · Terms · Privacy · All guides