Projects, standing instructions and progress notes

If you write letters for the same programme every week, explaining the style, terms and rules again in every chat gets tiring fast. Many AI apps offer projects or workspaces: a place where you store standing instructions and reference files once, and every chat inside it carries that context automatically.

Standing instructions are like system instructions: rules the model reads before your first message. Good instructions have one important property: they can be checked. Compare these two:

• "Write professionally and engagingly." Nobody can check whether a letter is "engaging".
• "Write in standard Indonesian, address the recipient as 'Bapak/Ibu', at most 200 words, always put the programme number in the first line, and do not promise dates that are not in the material." You can check every item in one read.

Checkable instructions bring two benefits. The model is more consistent, because the rules are clear. And you can spot when the model breaks them, because you know exactly what to look for. Keep standing instructions short: ten rules that are followed beat fifty that collide.

Reference files are documents the model reads in that project, such as a style guide, a glossary or approved sample letters. Choose carefully. Stale files make stale answers. Files with personal data make that data readable in every chat. And outside files, such as documents sent by other parties, can contain sentences that pose as commands to the model. This attack is called prompt injection, and it can also arrive through a connector that links the AI to other apps; how to handle it is in the lesson on connector risk and prompt injection.

Keep one project per piece of work or per client. One project for every client means one client's documents can colour the answers for another.

Long work, such as compiling an annual report or tidying a hundred documents, rarely finishes in one chat. A conversation can grow so long that the model starts to lose its beginning, a session can stop at a usage limit, or you have to switch devices. App memory is not the answer; it is brief and not fully under your control (see the lesson on data settings and memory).

The answer is a progress notes file you own: one short document you update every time a step is done. It has four parts:

1. Goal: one sentence about the end result.
2. Done: the finished steps, with where their results are.
3. Decisions: what has been decided and why, so it is not argued again.
4. Next: the exact step to do next.

When you start a new chat, paste the notes and ask the model to continue from "Next". A new chat with tidy notes often works better than an old chat that has grown too long.

Standing instructions need upkeep too. Whenever you see the model break the same rule twice, for example it keeps saying "student" when your programme uses "participant", add one firm rule. Whenever an old rule no longer applies, for example the programme number changes, delete or update it rather than leaving it piled next to the new one. Write the date of the last change at the bottom of the instructions, so colleagues using the project know the version. If the project is shared, agree who may change the instructions; instructions quietly edited by many people soon become a pile of rules that contradict each other.