Building a system
From one-off AI wins to something you can rely on
Most people who use AI seriously have a folder of moments where it worked brilliantly, and no way to get that result again on demand. The gap between those two things is the whole difference between a novelty and a tool.
The symptom is recognisable. You got something excellent out of it in March. You cannot remember what you typed, you have rewritten it badly twice since, and each attempt drifts further from the version that worked.
The problem is storage, not skill
You are not getting worse at this. You are retyping from memory every time, and memory is a poor place to keep a working process.
Every good result you do not save is thrown away. Every time you start from scratch, you repeat the same discovery work. That is the entire reason the wins never compound.
Build a library, not a pile
A folder of saved chats is better than nothing and still nearly unusable, because you cannot find anything in it. Give each saved process a small amount of structure.
- A name that says when to use it. "Reply to a delivery complaint", not "customer email v3 final".
- The instruction itself, complete, ready to paste.
- What to fill in — mark the gaps clearly, so future-you does not have to reread it to see what changes.
- One example of good output, so you can tell whether today's attempt is up to standard.
- The known failure — "it tends to invent a delivery date; check that."
Five lines of structure turns a chat you cannot find into something you or anyone else can run in thirty seconds.
Where to keep it
Wherever your team already looks. A shared document, your notes app, the wiki you already have. The best system is the boring one you will actually open.
Resist buying a dedicated prompt manager in month one. The constraint is the habit of saving things, and no software supplies that. Add tooling once you have thirty entries and the document is genuinely unwieldy.
Review it, or it rots
Prompts decay. Models change, your prices change, the policy you referenced was updated in June. A library nobody has checked in a year will confidently produce outdated answers, which is worse than having none — because now you trust it.
Once a quarter, run each saved process once and check the output is still right. Twenty minutes, and it catches the ones that quietly stopped working.
Date every entry when you write it and when you last checked it. An undated prompt gives you no way to judge how much to trust it.
The test that says it is a system
Hand one to someone else and leave the room.
If they get a usable result, you have built something real — it exists outside your head, survives your holiday, and can be handed to a new starter. If it only works when you are there to steer, you have a habit rather than a system, and the knowledge is still trapped.
Start with three
Do not attempt to document everything. Pick the three jobs you do most often and write those up properly.
Three reliable processes you can run in seconds are worth more than forty half-remembered chats, and getting to three takes about an hour. Everything after that is easier, because you will have stopped thinking of each good result as a lucky accident and started treating it as something worth keeping.