Design an internal prompt library and explain who maintains it.

Interviewer's question: "Design an internal prompt library and explain who maintains it." Hartwell & Voss LLP is a mid-size law firm handling contract review and litigation support. Priya Varkey is knowledge management counsel there.

How to answer this, stage by stage

Nobody is grading whether you can describe a searchable prompt library. They're grading whether you thought about who keeps it honest six months from now.

Let's learn

The Prompt Shelf is an internal library of vetted prompts for common drafting tasks: first-pass NDA review, deposition summary drafts, standard client-update memos.

Before it, associates at Hartwell & Voss wrote their own prompts from scratch each time. A junior associate, working it out alone, took about three hours to get a usable first draft of an NDA review memo. A senior associate, with prompts they'd refined over years, did the same task in forty-five minutes.

The turn: the time savings were never the actual point. The point was that a junior associate's first draft stopped depending on which senior happened to mentor them, or whether that senior felt like sharing their private prompt tricks.

Same task, same associate. What changes is whether their starting point depends on luck.

At its worst: the firm updates its citation style guide, a widely used deposition-summary prompt never gets touched, and for months, associates pull a template that confidently cites the old format, because "it's in the official library" reads as "someone checked this."

Four branches, and only one of them lets a stale prompt keep looking official.

What I would leave alone: low-stakes, low-traffic prompts, like a template for internal meeting notes, don't need the same ninety-day cadence. A stale version there costs almost nothing if it slips.

The lesson: a prompt library's real product isn't the prompts. It's the guarantee that whatever's showing is still true, and that guarantee needs an owner's name on it, not a publish date frozen in time.

Now here is the same thing as a story

The short version above is what you'd say pitching The Prompt Shelf to Hartwell & Voss's managing partner. Read this one for how the audit actually caught it.

Priya Varkey has run knowledge management at the firm for four years. She built The Prompt Shelf's first version herself, uploading prompts contributed by senior associates who were, for the most part, happy to share what had taken them years to refine.

For its first year, senior associates who used to write their own prompts from scratch started reaching for the shelf instead, the same way anyone reaches for a form that already works. They stopped teaching junior associates the underlying reasoning behind a good prompt, because the shelf already had one that worked.

The near miss came during a routine partner-led audit, the kind that samples associate work product twice a year. An auditing partner flagged a memo citing the firm's old citation format, three months after the firm-wide style guide had changed. Tracing it back, the memo had come from a Prompt Shelf template that nobody had touched since before the update.

Three of these four steps are invisible from outside the library. Only the audit ever looked.

Nobody had deleted the prompt's owner field. It simply had never had one. Priya's first version of The Prompt Shelf let anyone upload a prompt, credited to whoever wrote it, with no expectation that they'd ever look at it again.

Here's the decision I'd take back: launching the library with contribution but no ownership, crediting a prompt to whoever wrote it without ever asking them to keep watching it. That made sense when the shelf held a dozen prompts everyone could mentally track. It stopped making sense once it held two hundred, and nobody could remember which ones anyone was still responsible for.

Replayed with named ownership and a ninety-day re-verification window: the same style guide update happens. The deposition-summary prompt's owner gets a routine reminder at day sixty, checks it against the new format, finds the mismatch, and updates it before a single associate ever pulls the stale version. The next audit finds nothing to flag.

I let the shelf launch without real ownership because asking busy senior associates to commit to "maintaining" something felt like a big ask for a first version. It took a partner's audit and a firm-wide style guide gap to see that the alternative, a library nobody owned, was a much bigger ask on everyone who trusted it.

SPARK, mapped onto one prompt's lifecycleNot a features list. SPARK is what tells you why ownership, not search or formatting, is the actual design.

Each of these is a real feature the firm could build eventually. None of them is this design's job today.

Same stale prompt, two different libraries. Only one of them ever asks anyone to look at it again.

The top-right corner is exactly where a missed re-verification would actually hurt someone.

The recap, one line per letter: situation is inconsistent, hoarded prompt knowledge; payoff is trusting a shared, owned template; anchor is named ownership with a re-verification window; risk is an unowned prompt going quietly stale; and keep out is holding off on automated generation and client-facing use.

And if you want to be sure it really works, try it somewhere elseSame five letters, a hospital's residency program instead of a law firm. Nothing about the two jobs is alike.

Ridgemont Health Network gives residents a library of vetted prompts for drafting routine clinical notes. Dr. Aisha Okonjo is a second-year resident there.

Mapped onto SPARK: situation is a resident writing a discharge-summary note from scratch, leaning on whatever a senior resident happened to teach them on rounds. Payoff is residents trusting a shared, owned template for routine cases while still writing complex or unusual cases themselves. The anchor is structurally identical: every clinical note template has a named attending physician as owner and a last-verified date, auto-hiding if it passes its review window. Risk is the same shape too: if a documentation standard changes and a template's owner has rotated to a different unit, the old template keeps producing notes in a superseded format, and "it's in the hospital's approved library" makes it more trusted, not less. Keep out is no auto-generation from patient chat transcripts, no use for anything beyond routine documentation, no removing an attending's review step.

A different shape of picture than Section 2 used: parts around a center, not a branching tree. The anchor underneath is the same idea.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "every prompt gets a named owner and a re-verification window, and auto-hides if it goes stale, since maintenance is the actual design," and stop.
Cost: if a full re-verification workflow is too much for version one, start with just a visible last-verified date and a manual quarterly review, and say so honestly.
The model gets better, for real: even if the underlying model improves and prompts need less careful wording, an unowned template is still just as likely to cite outdated firm policy, since the model's quality was never the risk here.

Where people run it wrong.
They build the library, launch it, and treat "who contributed this" as the same thing as "who's responsible for it staying accurate."
They let a prompt stay visible and unflagged indefinitely, so a stale one looks exactly as trustworthy as a current one.
They wait for an audit or a complaint to catch drift, instead of building a re-verification cadence into the library itself.

How to use it live. When someone asks you to design an internal prompt library, ask yourself first: six months from now, who finds out first if this specific prompt is wrong, and how? If the honest answer is "nobody, until someone complains," the design isn't done yet.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0