Prompt Chain: End-to-End Shift Documentation Workflow

What This Builds

You'll create a complete shift documentation system that goes from voice note → structured clinical note → EHR-ready text, using a multi-step prompt chain. At the end of your shift, instead of spending 45–60 minutes charting from memory, you'll have 3–5 minutes of cleanup work: paste your transcripts into the chain, review the organized output, and copy into Epic or Cerner. This system can save 30–45 minutes per 12-hour shift.

Prerequisites

• Otter.ai set up on your phone (see Level 3 voice documentation guide)
• ChatGPT Plus subscription ({{tool:ChatGPT.price}}/month) for Custom GPT access — or a free account if you use the prompt chain manually
• One full shift of voice notes collected via Otter.ai (start with Level 3 first)

The Concept

A prompt chain is a multi-step AI workflow where the output of one step becomes the input of the next. Think of it like a car wash — your raw voice note enters at one end, passes through several cleaning and formatting stages, and comes out as a polished, structured clinical note ready to paste into the EHR. Each step builds on the last without you having to re-explain the context.

Step 1: Voice notes (captured in Otter.ai during the shift) → Step 2: Organize by patient and clean up transcription errors → Step 3: Structure each entry into clinical note format → Step 4: Format final output matching your EHR's note structure

Build It Step by Step

Part 1: Establish Your Voice Note Habit (If Not Already Done)

This workflow only works if you've been collecting voice notes during the shift. If you haven't set up Otter.ai yet, complete the Level 3 voice documentation guide first.

The goal: at the end of your shift, you have 15–30 voice notes in Otter.ai, organized roughly by time. They don't need to be perfect — just complete enough that the information is there.

Part 2: Create the Prompt Chain in ChatGPT

Open ChatGPT (free or Plus) and start a new conversation. The chain runs in three sequential prompts:

Prompt 1 — Organize raw transcripts:

I'm a respiratory therapist. Below are my raw Otter.ai voice transcripts from a 12-hour shift. Each transcript is a brief voice note about a patient assessment or treatment. Please:
1. Group the notes by room/patient identifier
2. Put them in chronological order within each patient
3. Clean up obvious transcription errors (e.g., "PH" → "pH", "speedy 02" → "SpO2")
4. Do not add information or change clinical content — only organize and clean

Raw transcripts:
[PASTE ALL OTTER.AI TRANSCRIPTS HERE]

Prompt 2 — Structure into clinical note format:

After ChatGPT organizes and cleans the transcripts, send this follow-up in the same conversation:

Now, for each patient, convert the organized notes into a structured clinical documentation entry suitable for a respiratory therapy EHR note. Format each patient entry as:

Patient: [Room/ID]
Date/Time: [from transcript timestamps]
Assessment: [patient status, breath sounds, SpO2, work of breathing]
Treatment: [what was done — device, medication, duration, settings]
Response: [patient response to treatment — objective and subjective]
Plan: [next steps, orders pending, recommendations]
Clinician: [leave blank — RT will fill in]

Maintain clinical language. Don't add information that wasn't in the original voice notes.

Prompt 3 — Final QA check:

Review all the structured notes above and flag any that:
1. Are missing a treatment response (I may have forgotten to mention it)
2. Have inconsistent clinical data (e.g., SpO2 before treatment listed higher than post-treatment)
3. Are missing key required fields

Give me a brief list of items to manually verify or add before charting.

Part 3: Copy Output into EHR

1. Review the structured notes from Prompt 2
2. Address any flags from Prompt 3's QA check
3. For each patient, copy the structured entry → paste into your Epic or Cerner respiratory therapy note field
4. Add your name, credential, and timestamp
5. Sign and lock the note per your department's policy

Part 4: Optional — Build This as a Custom GPT

If you use this workflow regularly, build it as a Custom GPT (see the Level 4 Custom GPT guide) with these instructions pre-loaded so you don't have to type the three prompts every time:

System instruction: "You are an RT documentation assistant. When I paste voice transcripts, run these steps automatically: 1) Organize by patient and clean errors, 2) Structure into clinical note format with Assessment/Treatment/Response/Plan sections, 3) Flag missing information. Format output for direct EHR entry."

With this GPT saved, your workflow shrinks to: paste transcripts → review output → copy to EHR.

Real Example: Medical/Surgical Floor Shift

Setup: You covered a 32-bed medical floor, administered 18 treatments across 12 patients over 12 hours. You recorded a 20-second voice note after each treatment.

Input (paste into Prompt 1):

Sample of raw transcripts:

• "8:15 room 214 albuterol treatment done uh patient SpO2 93 before 97 after no side effects continue Q4"
• "8:45 room 219 patient refused treatment said feeling better educated on importance call the charge"
• "10:20 room 214 second treatment SpO2 stayed at 96 good improvement keep going"
• "11:00 room 231 incentive spirometry teaching new post-op patient hip replacement good effort volumes around 1200 needs encouragement"

Output (what the chain produces after all 3 prompts):

Patient: Room 214 Assessment: SpO2 93% pre-treatment on 2L NC; improving to 96-97% post-treatment; tolerating well Treatment: Albuterol 2.5mg SVN 0800 and 1000 per Q4h order Response: SpO2 improved from 93% to 97% post-0800 treatment; maintained at 96% post-1000 treatment; no adverse effects reported Plan: Continue Q4h albuterol per order; assess for step-down to MDI at next evaluation

Patient: Room 219 Assessment: Patient declined ordered treatment Treatment: Albuterol SVN refused at 0845 Response: Patient verbalized feeling improved; declined treatment citing subjective improvement Plan: Educated on importance of scheduled treatment for symptom control and prevention; charge RT notified; physician notification pending Clinician: ___________________

Time saved: 18 treatments × 3 minutes per note = 54 minutes batch charting from memory → paste and review in 8 minutes

What to Do When It Breaks

• ChatGPT misinterprets clinical abbreviations → Add a glossary to your first prompt: "Note: SVN = small volume nebulizer, MDI = metered dose inhaler, Q4 = every 4 hours"
• Notes are grouped under wrong patient → Your voice note didn't clearly state the room/patient — add room identification as the first thing you say in each voice note
• Output is too verbose for EHR field → Add to Prompt 2: "Keep each patient entry under 150 words — EHR space is limited"
• Prompt 3 flags too many items → Your voice notes are missing response documentation — practice ending each voice note with "Patient tolerated [well/with difficulty]"

Variations

• Simpler version: Skip the prompt chain and just use Prompt 2 alone — paste organized Otter transcripts (manually grouped by patient) and get direct clinical note output
• Extended version: Add a 4th prompt that generates the shift handoff report from the same transcripts — one set of voice notes, two outputs: individual patient notes and shift handoff summary

What to Do Next

• This week: Collect voice notes for one shift, run the 3-step chain, compare charting time to a normal shift
• This month: Refine your voice note format based on what Prompt 3 consistently flags as missing
• Advanced: Build the chain into a Custom GPT so the 3 prompts run automatically — reduces workflow to paste + review

Advanced guide for respiratory therapist professionals. Voice-to-documentation workflow uses Otter.ai (free/paid) and ChatGPT (free/{{tool:ChatGPT.price}} for Plus).