I Deleted 40 Saved Prompts and Kept 4

My prompt library had grown to forty-one entries. Folders, tags, a naming convention. I opened it last week and realised I had used six of them in three months, and four of those six did all the work.

So I deleted the other thirty-seven. Here is what survived, and why the ones that died had to die.

1. Interview me first

Before you do this, ask me the three questions you most need answered to do it well. Wait for my answers. Do not start.

Almost every bad output I get comes from a prompt that was missing something I knew and never said. This one moves the burden: instead of me guessing what context matters, the model tells me what it is missing. The questions are usually obvious in hindsight and never obvious in advance.

2. Make it falsifiable

What would have to be true for this answer to be wrong? List the three assumptions you made that I did not give you.

A model will happily produce a confident answer built on three invented premises. This makes it show them. Half the time one of the assumptions is wrong and the whole answer collapses — which is worth knowing before I act on it, not after.

3. Force the specifics

Rewrite that with a number, a name or a date in every claim. If you do not have one, write MISSING instead of describing it in general terms.

Vagueness is the default failure mode, and it hides well: text with no specifics reads smooth and says nothing. MISSING turns invisible gaps into a checklist.

4. Cut what I already know

Cut this in half. Remove anything a reader in my field already knows. Keep the parts that would change what they do on Monday.

The model writes for a reader who knows nothing, because it does not know who is reading. This tells it. What survives the cut is usually the only part that mattered.

Why the other thirty-seven died

They were phrasings, not moves. Twelve variations of "write a blog post about X" that differed only in adjectives. Eight "act as a senior expert" openers that changed nothing measurable. A dozen formatting instructions I now keep in a rules file instead of retyping.

A prompt is worth saving when it changes what the model does, not how it sounds. Interviewing me, exposing assumptions, demanding specifics, cutting the known — those are four different moves. The thirty-seven I deleted were one move in thirty-seven costumes.

The test

Open your own library. For each entry, ask: if I removed this, would the output change in a way I could measure? Not "would it feel different" — measure. Most will fail, and that is the point. Four prompts you actually use beat forty you maintain.

Want the calm version of AI news like this, once a week? Subscribe to the Sharp AI Hub newsletter →