Dev.to AI ๐Ÿค– Ai ๐Ÿ‘ 0 ๐Ÿ“– 6 min read

I built a prompt library for nurses, and the guardrails turned out to be the hard part

I built a prompt library for nurses, and the guardrails turned out to be the hard part A short field note on shipping 116 AI prompt templates to a profession that cannot afford a confident wrong answer. This is the ada

I built a prompt library for nurses, and the guardrails turned out to be the hard part

A short field note on shipping 116 AI prompt templates to a profession that cannot afford a confident wrong answer. This is the adapted version of an article that first appeared on the AI Stack Lab blog. Same product, same limits, different edit.

Nursing is one of the fastest-adopting professions for consumer AI, and for an unglamorous reason: documentation burden. Charting, handoffs, discharge teaching, preceptor feedback, care plans โ€” the writing never stops, and most of it follows a template the nurse already knows by heart. That is exactly the shape of work an LLM is good at.

So the demand is real. The supply is mostly junk. Search "AI prompts" on any marketplace and you get tens of thousands of generic bundles where six of the two hundred prompts are healthcare-adjacent and none of them know what SBAR means.

I spent a while in that gap last month and shipped the result: NurseAI Kit, 116 nurse-specific prompt templates in one offline HTML file, one dollar on Gumroad. This post is about what I learned building it, because the interesting part was not the prompt writing. It was the constraint design.

A prompt is a schema, and schemas need invariants

The first version of the SBAR prompt produced beautiful SBAR notes that occasionally contained vitals I never supplied. The model was not hallucinating wildly โ€” it was pattern-completing. An SBAR block with an empty "Assessment" line invites an assessment, and the model will supply a plausible one.

That is a bug class, not a bug. In nursing writing, a plausible-but-invented number is worse than a blank. So the fix could not be "better prompt wording" alone. It had to be an explicit contract in the prompt text:

Keep everything strictly to what I gave you; add no clinical details, no vitals, and no numbers I did not provide. If a section is missing something important, insert a bracketed [CHECK] marker instead of inventing content.

The [CHECK] convention is the whole trick. It gives the model a legitimate way to be incomplete. Without one, "be accurate" is just a wish; with one, the failure mode becomes visible to the reviewer instead of hidden in fluent prose.

Every one of the 116 prompts carries a version of that contract, plus two other invariants I would not ship without:

  • No medication doses, no diagnosis, no treatment content. Not a disclaimer at the bottom of a page โ€” a scope sentence inside the prompt text itself, so it applies at generation time.
  • A visible draft label. Outputs come back marked as a draft for a licensed professional to review and own. In testing this caught a real problem: a rewritten discharge sentence came back more specific than the source handout, and the review step is what flagged it.

If you take one thing from this post: when you hand a prompt to a regulated profession, the interesting engineering is the guardrail, not the personality.

The structure the model cannot guess

Generic prompts fail in a second, quieter way: they produce a hybrid artifact. Ask for "a handoff" and you get half a note, half a spoken report, written for nobody.

The fix is to encode the artifact's real shape. Concretely, from the kit:

Provider call script with an escalation ladder. Situation / Background / Assessment / Recommendation, each block under three lines so it can be scanned while dialing โ€” and then the part every generic prompt pack omits: how long to wait before the second call, when to escalate to the covering provider or house supervisor, and what to document after each attempt.

Plain-language handout rewrite. Paste the current handout, specify a reading level, and the prompt keeps every clinical fact, warning sign and follow-up instruction intact, cuts sentences to under about fifteen words, and ends by listing the phrases it was unsure how to simplify so you verify those against your education standards.

Post-night-shift recovery routine. A timed sequence from clocking out through waking, a postponement list for household demands, and a worst-day fallback โ€” with an explicit instruction not to suggest any substance to aid sleep and to redirect persistent sleep problems to a primary-care provider.

Each of those is only useful because the prompt asks for an artifact with a known shape. The model supplies prose; the schema does the work.

Packaging mattered more than I expected

I assumed the deliverable would be a document. It isn't. The deliverable is: how many steps sit between "I have 90 seconds and an ugly handoff to write" and "I have a usable draft."

A PDF or a hosted doc fails that test โ€” it is a file to scroll, on a phone, between med passes. So the kit is a single self-contained HTML file. Double-click it, no install, no account, no network. Live search, category sidebar, click to copy. It opens beside whatever ChatGPT, Claude or Gemini tab you already have.

Zero external references: no CDN, no fonts, no telemetry. Everything, including all 116 prompts, is embedded in that one 213 KB file. I verified it by grep, not by faith โ€” zero external http references in the markup.

The nine categories, if you are wondering whether your work is covered:

Documentation        11   SBAR drafts, shift summaries, charting phrase banks
Communication        16   provider call scripts, de-escalation, family updates
Care Planning        16   SMART goals, fall risk, delegation worksheets
Patient Education    11   handout rewrites, teach-back sets, discharge checklists
Career               17   STAR stories, resume bullets, cert study plans
Education/Precepting 14   preceptor feedback, orientation plans, simulation outlines
Admin & Leadership   14   agendas, budget justifications, project proposals
Wellness              9   shift recovery, burnout self-checks, debriefs
Research & Evidence   8   PICO builders, paper summaries, evidence briefs

Every prompt file has the same anatomy: a use-case line, the {variables} it needs, the full prompt, a practical tip, and a safety/scope line. I checked that programmatically across the bundle โ€” 116 files, all 116 with at least one placeholder, a tip, and a safety line. No stubs.

What I deliberately did not build

This is where a product post usually goes quiet, so let me be loud instead.

  • No clinical decision support. It does not interpret labs and never will.
  • No EHR integration, no API, no CRM hookup. It is text you paste into a chat window. I claimed nothing else in the listing, and I am not claiming it here.
  • No time-savings number. The honest answer depends entirely on how fast you review, and reviewing is the slow part. A "saves 3 hours a week" claim would be fiction.
  • No ratings, reviews, or customer counts. There are none to quote, so none appear.

Who should skip it

If you are a nurse who already writes your own prompts and knows exactly what you want from each artifact โ€” skip it, you have solved the problem this kit solves.

Skip it if you want clinical decision support, expect doses or diagnostic content (the prompts refuse that category by design), do not use an LLM at all, or will not review drafts. Skip it if you need a specific EHR integration.

Buy it if your AI outputs at work still feel generic and you have been meaning to build this prompt library yourself for months.

NurseAI Kit โ€” 116 AI Prompts and Workflow Templates for Nurses. One offline HTML file, 116 markdown prompt files, README. $1.00 USD launch price on Gumroad: nandminc.gumroad.com/l/nurseai-kit

Disclosure: this is our own product โ€” we publish it, so a purchase through that link supports this writing directly and costs you nothing extra. I wrote this so you can decide whether it is worth a dollar before spending it.

Honesty block. Example figures are illustrative planning inputs โ€” not market rates and not earnings. Nothing here is legal, tax, insurance, licensing, veterinary, automotive-repair, or financial advice, and nothing here is clinical advice: the kit is workflow tooling for licensed professionals, and AI output is a draft you review and sign. No claims of CRM, map, API, or invoicing integrations the bundle does not ship โ€” it ships one HTML file and 116 markdown prompts.

Questions before you buy, or a workflow the kit should cover next: [email protected]. I answer my own mail.

๐Ÿ“ฐ Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes โ€” full credit and traffic to the original publisher.