The Prompt Library Nobody Uses

There's a predictable artefact in any team six months into AI adoption: a shared document of prompts, created with real enthusiasm, last edited eleven weeks ago, opened by nobody.

It's such a consistent pattern that it's worth treating as a design problem rather than a discipline problem. Nobody is lazy here. The library failed for structural reasons, and they repeat.

Why they die

Finding one costs more than rewriting. Forty entries with vague names, no categories, and no indication of which still work. Searching that is slower than typing the request again, and people are correct to skip it.

Nobody knows which are good. A library where the excellent and the abandoned sit side by side with equal weight is a library you can't trust, so you don't use it.

They're written for the author. A prompt someone refined over ten iterations carries assumptions from their specific case. Pasted by a colleague, it produces something subtly wrong, and after two of those they stop trying.

They rot invisibly. Models change, processes change, the template's assumptions expire. There's no signal — it just quietly starts producing worse results, and the person who notices assumes they've done something wrong.

Nobody owns it. Shared with everyone means owned by nobody, which means nobody removes the dead entries, and the dead entries are what make it untrustworthy.

What the surviving version looks like

The libraries still in use after a year are strikingly consistent, and mostly by being smaller.

Fewer than a dozen entries. Covering the genuinely repeated tasks, not everything anyone ever found useful. If it's longer than a screen, it's an archive, not a tool.

Named by the job, not the technique. "Weekly client update" gets found. "Structured summarisation prompt v2" does not. People search by what they're trying to do.

Each entry includes what to change. The variable parts marked plainly, and a line about what the template assumes. This is what makes it usable by someone other than the author.

Each says when it was last verified. A date, updated when someone confirms it still works. Old dates are a signal, and a signal is more useful than a maintenance policy nobody follows.

One owner. A person, whose job includes deleting things. Deletion is the maintenance task that matters — a library of eight trusted entries beats one of forty unknown ones by a wide margin.

Store it where the work happens

A library in a separate document loses to the friction of switching. Most teams do better putting templates where the task starts — a saved reply in the mail client, a project template, a pinned message in the relevant channel, a custom instruction inside the AI tool itself.

If your tool supports saved prompts or projects, that's usually the right home: it's zero switching cost and it's where someone will be when they need it.

Make contribution a by-product

"Add your good prompts to the library" reliably produces nothing, because it's a separate task with no immediate return.

What works is attaching it to something already happening. A standing two-minute item in an existing meeting — anything worth adding, anything to delete? — collects more than any amount of encouragement, because it removes the initiative cost and it surfaces the deletions, which nobody ever volunteers.

Prune on a schedule

Once a quarter, someone opens it and asks two questions per entry: has anyone used this, and does it still work?

Delete generously. An entry nobody has used in six months isn't a resource, it's noise reducing the credibility of everything next to it. The library's value is concentration, and pruning is the only thing that maintains it.

Ilse — the library of eight

Ilse leads a customer success team. Their shared library had reached fifty-three prompts and was, by her own description, "a graveyard".

She replaced it with eight, chosen by asking the team what they actually did every week. Each has a plain job name, marked variables, an assumptions line, and a verification date.

Usage went from near-zero to routine. Her explanation is unromantic: forty-five of the entries had been making the other eight impossible to find.

Mateo — the template that only worked for him

Mateo is a data analyst at a healthcare provider. He'd contributed a genuinely excellent analysis prompt, refined over dozens of runs, and colleagues got mediocre results from it.

The reason was that it assumed his data structure — column names, a particular date handling, a filter he applied without thinking. All invisible to him, all fatal to anyone else.

He added four lines: what the template assumes about your data, and what to change if that isn't true. Same prompt, and it started working for other people.

The one thing

Prompt libraries fail because they're too big, unowned, unversioned, and written for their authors. Concentration is the whole value.

Fewer than a dozen entries, named by job, with marked variables and a verified date, owned by one person who deletes things — and stored where the work already happens.