From prompts to Practice Loops

A good prompt gets you a good draft once. A Practice Loop turns that into a repeatable workflow with a job, a standard, a stopping rule and a record of who signed it off. This page explains the idea using Practice Loops for Nursing, an open project from the same Clinical Quality AI (CQAI) community as this toolkit.

Where this fits

The Clinical Prompt Library helps you write a strong single prompt. A Practice Loop wraps a prompt in governance: it is closer to a care pathway than to a question. To try one, see the Practice Loops tool page.

Why a prompt isn't enough

A one-off prompt has no built-in answer to the questions a governance lead would ask:

• What standard was the output measured against?
• Who checked the reasoning, and when?
• What happens if the input contains identifiable data?
• Who accepted the result, and where is that recorded?

People who are handed a finished AI output tend to accept it. Over-reliance on automated advice (automation bias) is well documented in clinical decision support, and checking the reasoning, not just the result, is one of the ways to reduce it (Goddard, Roudsari and Wyatt, 2012).

The six pillars

Every Practice Loop runs through the same cycle:

Pillar	What happens	Nursing parallel
Trigger	The nurse starts the loop with de-identified input.	Assessment
Task	The job is bounded, including what the AI must not decide (for example progression, fitness to practise or disciplinary outcomes).	Scope of practice
Standard	The loop names the standard it will be judged against (NMC proficiencies, the Code, local policy) before it drafts anything.	Evidence-based practice
Verification	The draft is scored against ten explicit checks.	Audit against a standard
Iteration	Anything scoring below 8 out of 10 is revised and re-scored, up to three rounds.	Evaluate and adjust
Human sign-off	The output stays a DRAFT until a named professional accepts, corrects or refuses it.	Accountability

Two human gates

The loops map onto the nursing process with two points where the loop stops and waits for the nurse:

1. Gate 1: check the reasoning. Before drafting, the AI shows how it has framed the problem: what it thinks the issue is, which standard applies, and what is fact versus inference. The nurse confirms or corrects this. For example, repeated lateness alongside a disclosed caring responsibility is a wellbeing matter, not a conduct concern. Correcting the framing here is far easier than editing a finished document built on the wrong premise.
2. Gate 2: accept accountability. The nurse reviews the final draft and signs it off, refuses it or escalates. Nothing is final until a named person says so.

This mirrors the NMC Code: you can delegate a task, but not your accountability for it (see The NMC Code review and AI).

Stop rules and records

A Practice Loop also defines when to stop and escalate instead of producing an answer: safeguarding concerns, patient safety risks, suspected bias or not enough information. Every run writes a timestamped audit record (what ran, what was flagged, what scored low, who signed), so a team can later answer governance questions with evidence rather than memory.

The PRACTICE capabilities

The project also sets out eight capabilities a professional needs to direct AI safely, spelling PRACTICE:

Capability	The question to ask
Purpose	What is this explicitly not allowed to conclude?
Restrict	What exactly did I put in, and could I defend that to an information governance lead?
Anchor	Which standard applies, and why that one?
Challenge (Gate 1)	What did I correct before it drafted anything?
Test	On what basis did I decide this was good enough?
Iterate	What changed as a result of what I found?
Confirm (Gate 2)	Who is accountable for this, and did I actively accept that?
Evidence	Could I show someone else that all of this happened?

These work as a teaching framework even if you never install the tool: they are questions any nurse can ask of any AI output.

Teaching activity: build a loop on paper

Time: 40 minutes · Group size: 3–4

1. Choose a task (5 min): for example, turning placement meeting notes into an action plan.
2. Bound it (10 min): write the one-sentence job and three things the AI must never decide.
3. Anchor it (10 min): pick the NMC standards the output should be checked against, and write five of the ten verification checks.
4. Set the gates (10 min): what must the AI show at Gate 1? Who signs at Gate 2, and what would make them refuse?
5. Compare (5 min): try the live simulator and compare its loop with yours.

---

Source: Practice Loops for Nursing (PolyForm Noncommercial 1.0.0), described and summarised here with attribution.

Related: Clinical AI SDLC · GOV.UK AI guardrails for nursing · AI Nursing Constitution