Releases: The-Paramedic-Foundation/paramedic-narrative-skill
Release list
Canonical knowledge set with generated renderings
This is a breaking change to the file layout. No clinical content changed. Every threshold, criterion, and standard is the same as 2.1.1. What changed is how the files are organized and how the platform renderings are produced.
Why
The 2.1.1 audit found that safety-critical text had been paraphrased into each platform's instruction file rather than reproduced, and the paraphrases had lost content. The ChatGPT rendering was missing six of eleven forensic triggers, including domestic violence and sexual assault, and its standing disclaimer was truncated. Fourteen of twenty findings traced to one cause: three hand-maintained paraphrases of a single standard drift apart.
Those files are no longer written by hand.
What is different
One always-loaded block, identical on every platform. It holds only the rules that must never depend on retrieval: the editorial-tool boundary, ten NEVER rules, the attribution boundary, PHI, patient and incident separation, during-a-call guardrails, the pre-draft check, style, the standing disclaimer, and a router.
Seventeen topic files in paramedic-narrative/references/, all named WHEN-*.md. Each is named after the situation that requires it, not the subject, and opens by restating its own trigger. Filenames are the retrieval lever, so WHEN-FORENSIC.md is easier for a model to route to than a section buried in a monolith.
system-prompt.md, chatgpt-instructions.md, claude-project-instructions.md, and SKILL.md are now generated from those sources by build.py. Sources live in src/.
A verifier gates the build. It fails if the always-block is not byte-identical in every rendering, if a knowledge file is not routed, if the router points at a missing file, if the ChatGPT instructions exceed 6,500 characters, or if house style is violated. The 2.1.0 defects would not have shipped past it.
documentation-standards-primer.md and narrative-formats.md no longer exist. They duplicated each other and SKILL.md on seven topics, including the entire ACS Field Triage criteria list appearing near-verbatim in both. That content was deduplicated into the seventeen files.
Upgrading from 2.x
Remove the old files first. If SKILL.md, documentation-standards-primer.md, or narrative-formats.md remain alongside the new set, the assistant may consult a retired copy.
Provider profiles and agency configuration files are unaffected and need no changes.
On seventeen uploads
For ChatGPT and Claude Projects this is more work than the previous two files. It is one-time. The reason is that Knowledge retrieval is chunked and not guaranteed per turn, so a monolithic file means a safety rule may not reach the model on a given call. Claude users can avoid it entirely with the .skill package.
Contributing
Pull requests edit src/, never the generated files at the repo root -- those are overwritten by build.py. Run python3 build.py && python3 build.py --verify before opening a PR.
Cross-file parity corrections
Parity corrections across the platform renderings. No change to clinical content,
thresholds, or standards. Agencies that have reviewed 2.1.0 carry no re-verification
burden. Updating replaces the instruction and reference files only; provider profiles
and agency configurations remain valid.
Why this release exists
The repository publishes one standard in five renderings, each compressed for a
different platform's constraints. An audit of all eleven instruction, reference, and
supporting files found that safety-critical text had been paraphrased into those
renderings rather than reproduced, and the paraphrases had lost content. Fourteen of
twenty findings were instances of that single root cause.
The most consequential were in chatgpt-instructions.md, which is the only
always-loaded text on that platform. Knowledge File retrieval is not guaranteed per
turn, so anything abbreviated there may simply never reach the model.
Corrections
chatgpt-instructions.md
- Forensic trigger list restored to all eleven triggers. It carried five, omitting
domestic violence, sexual assault, intoxication-related harm, arson, motor vehicle
collision with impairment or fatality, and threats. Two of those are among the most
common forensic call types in EMS, and their absence could have prevented the
standard from applying at all. - Standing disclaimer restored verbatim, including the AI-tool disclosure and the
no-clinical-advice sentence, both of which had been dropped. - Camera-metadata PHI rule added.
- Controlled substance element list regains container identifier and reconciliation.
- Pre-draft grouped verification restored as an explicit workflow step.
- The rule against assuming your crew performed an act whose performer is unstated
added. - Body reduced to 7,649 characters against the 8,000 limit to accommodate the above
and leave margin. The WHAT THIS TOOL DOES NOT DO list was removed from this
rendering only; every item in it is stated as an operative rule elsewhere in the
same file, and the full list remains in SKILL.md. - The header now states that the header lines are not part of what you paste.
system-prompt.md
- "threats" restored to the forensic trigger list.
- Neonate temperature threshold restored. This was the only numeric divergence from
SKILL.md across all eight age bands and 43 values. - "medical director-adopted guidelines" corrected to "Chief Paramedic-adopted", the
last instance of pre-2.0.0 terminology.
claude-project-instructions.md
- Incident and patient workspace isolation now carried inline rather than by reference
to a project file. Retrieval is model-initiated, and this rule turns on a passive
judgment call with no keyword likely to trigger a lookup. The file's stated purpose
is to hold safeguards that must never depend on retrieval; preventing cross-patient
contamination qualifies. - Controlled substance element list regains reconciliation.
- Forensic trigger list regains domestic violence, sexual assault, and threats.
- The inversion rule now names the principle it inverts instead of citing a number
that means something different in SKILL.md.
README.md, SKILL.md, and the primer
- README forensic trigger list aligned to the canonical eleven.
- README's summary bullets now name the attribution boundary, patient workspace
isolation, the rule against referencing a structured entry that does not exist, and
the rule against assuming the documenting crew performed an unstated act. - The primer's Neurological scoring list regains NIHSS, which both instruction files
carried and the reference file did not. - Cross-references to "the primer under Scoring Tools" corrected to that section's
actual heading. - "PCR platforms" corrected to "ePCR platforms" in the one boilerplate bullet where
three files had reverted to the bare form.
docs/
- The dictation pocket card text gains the attribution prompts in items 1, 2, and 9
that 2.1.0 described but did not ship. The printable PDF, business card, and mobile
reference image still require regeneration.
Contributor
The neonate temperature omission was independently reported by @kaju-skribh in #2,
with a fix in #3, before this audit was run. Their analysis identified not only the
missing value but why a compression pass loses it: temperature is the only vital
stated for a single age band, so its absence leaves no visible hole in the table.
What is not in this release
The audit also reviewed six production narratives and found defects that require
specification changes rather than text corrections, including an attribution failure
on a multi-agency call and a refusal narrative that does not meet the refusal
standard. Those are tracked for the next release.
Upgrading
Replace the instruction file you use, and replace SKILL.md and the reference files
wherever they are uploaded. Do not leave old and new copies side by side. Full
per-platform steps are in the README under Version Control.
Contributors
Assets 3
Attribution and data-integrity boundary
Structured ePCR entries are attributed entries. They assert that this crew performed this act, and where structured data feeds external reporting they become the official record of what the agency provided. Three categories of care therefore belong outside
structured fields and must be carried in full by the narrative: care performed by another agency's provider, care performed before the documenting crew assumed responsibility, and interventions prepared or considered but not performed.
The skill previously modelled only the forward direction of that boundary: structured fields hold what and when, the narrative holds why, do not restate. It had no concept of content deliberately excluded from structured fields. That produced a live defect.
Core Operating Principle 5 would emit "treatments as charted" for an intervention the agency's own standard forbids charting, leaving that intervention documented nowhere.
What changed
- New section: Attribution and Data-Integrity Boundary (
SKILL.md,system-prompt.md), including the inversion rule that suspends Core Operating Principles 3 and 5 for narrative-only content. - Core Operating Principle 1 extended to other clinicians' clinical reasoning. Reasoning is attributed only to the clinician who stated it.
- Core Operating Principle 5 now requires confirming that a structured entry exists before referencing it.
- Medication Administration Standard: prepared-then-withheld doses, including the full waste trail for a controlled substance drawn and not given; medications administered by another provider or before arrival.
- Workflow Step 3: a single attribution prompt covering all three categories, triggered on multi-agency calls, facility-origin transports, and any call where care was in progress on arrival.
- Dictation pocket card: items 1, 2, and 9 updated.
- Primer: new Multi-Agency and Prior-to-Arrival Care Attribution section; the regulatory reporting function added under Purpose; additions to Transfer of Care, the IMIST-AMBO T element, and cardiac arrest documentation.
- Narrative formats: universal inversion rule, quality checklist item 8, and IFT template updated for sending-facility care.
- Agency configuration: new Section 4A, Structured-Field Scope and Attribution Boundary, covering external reporting destinations, waivered or specially authorized acts, and the agency's rules for partner-agency care, prior-to-arrival care, and prepared-but-not-performed interventions.
Waiver, variance, and special-authorization language is deliberately jurisdiction-neutral. What is authorized, by which body, and how it must be reported varies by state and between agencies working the same call, so the rule is stated symmetrically and the agency configuration carries local specifics.
New file: claude-project-instructions.md
The documented Claude setup pasted all of system-prompt.md into a Project's Instructions field while also uploading SKILL.md, duplicating the entire standard. Instructions are re-sent with every message, so users paid roughly 17k tokens per turn for content already retrievable from project files, and the two copies drift apart whenever one is updated.
claude-project-instructions.md carries only what must be in context at all times (non-fabrication, the attribution boundary, PHI, controlled substance and forensic standards, workflow, and the standing disclaimer) and points to the project files for
everything else, at roughly 2.6k tokens.
system-prompt.md remains correct for Gemini Gems, direct API use, and any platform without file retrieval.
Upgrading from 2.0.x
This release replaces the instruction and reference files only. Provider profiles and existing agency configuration files remain valid, though agencies should complete the new Section 4A.
Claude Project users: paste claude-project-instructions.md into the Instructions field over whatever is there now. If your Project currently holds the full system-prompt.md in Instructions, replacing it with this file is the upgrade: same standards, a fraction of the per-message cost. Then remove the old SKILL.md and reference files from Files and upload the new ones. Do not leave both.
Claude Skill package users: download paramedic-narrative.skill below, remove the old skill, and upload the new file.
ChatGPT: replace the Instructions field with the new chatgpt-instructions.md, and replace SKILL.md in Knowledge. Do not leave both.
Gemini: replace your Gem's instructions with the new system-prompt.md.
Microsoft Copilot: update Custom Instructions with the new chatgpt-instructions.md and use the new SKILL.md for your per-session paste.
Full details in the README under Version Control.
v2.0.1 -- Incident/patient workspace isolation; mobile reference card fix
Paramedic-Narrative v2.0.1
New: incident and patient isolation
Documentation is now organized in two levels: an incident (dispatch, scene, mechanism, hazards, shared timeline) and one workspace per patient (demographics, history, exam, vitals, treatment, transport, disposition). Facts from one patient are never applied to another unless you say they're shared. Starting a new incident, adding or switching to another patient, continuing the same patient, and referencing a patient's prior encounter are all handled distinctly. You'll only get a check-in question when it's genuinely unclear which of these applies, and switching patients is confirmed with a short neutral line rather than repeating identifying details back to you.
Prior-encounter history is only used when you confirm it's the same patient, and is always labeled as prior history rather than presented as a current finding.
Fixed: mobile reference card image
docs/TPF_ParamedicNarrative_MobileReferenceCard_2026_v1.png had its footer and Paramedic Foundation branding cut off at the bottom edge. It's now sized correctly with the complete footer visible.
Getting the update
- Claude: replace the Instructions field contents with the new
system-prompt.md, and replaceSKILL.md(and any changed reference files) in Files. If your account uses the separate Skill package instead, download the newparamedic-narrative.skillfrom the Releases page and replace it under Settings > Skills (browser required for that step). - ChatGPT: replace the Instructions field with the new
chatgpt-instructions.md, and replace the uploadedSKILL.mdKnowledge File. - Gemini: replace your Gem's instructions with the new
system-prompt.md. - Microsoft Copilot: update your Custom Instructions and use the new
SKILL.mdfor your per-session paste going forward.
Your provider profile and agency configuration files are unaffected.
v2.0.0 -- Consolidated release: privacy, trust boundary, clinical accuracy, Chief Paramedic terminology
Paramedic-Narrative v2.0.0
New since your last update
- 14 narrative formats, not just SOAP. SOAPE, SOAPIER, CHART, CHARTE,
DCHART-E, CHRONOLOGICAL, HEAD-TO-TOE, DRAATT, AT CHART, FACT, a dedicated
REFUSAL/NON-TRANSPORT template, an IFT (interfacility transfer) template, and
a CUSTOM format your agency can define. Say "use CHART for this one" to
override per call. - Photos plus dictation, together. Add a monitor photo, an ePCR screen, a
medication label, or facility paperwork alongside your spoken or typed
description. Everything gets transcribed exactly and confirmed back to you
before it's used — nothing is ever guessed from a blurry photo. - Document a call in pieces, whenever you have a minute. Start on scene,
add more during transport, finish after you clear. The skill tracks what
it still needs and never asks you to repeat something you already gave it. - Documenting hours later? Say so. Instead of a blank interview, you get
targeted questions built from what's already known — the kind that jog
memory instead of demanding free recall. - On-request IMIST-AMBO handoff prep, assembled from what you've already
told the assistant during a call — say "handoff prep" or "IMIST-AMBO now." - A training example after every draft, modeling a clean structured
handoff built from your own call data. Clearly labeled, never part of the
PCR, never a critique of the handoff you actually gave. - On-request prearrival notification note — a copy-paste block matched to
your notification platform's fields, built only from what you've stated.
Corrections and clarifications in this release
Clinical reference accuracy
- The adult respiratory-rate reference range now matches NASEMSO's National
Model EMS Clinical Guidelines (Universal Care, Table 1), which the vital
threshold table cites directly. - HEART Score and CURB-65 guidance notes that troponin and urea/BUN are lab
values: document a full score only if that value was obtained, and use
CRB-65 (no urea component) otherwise. - PERC rule guidance notes it applies only when a patient has already been
assessed as low pretest probability for PE. - Cincinnati Prehospital Stroke Scale and "CPSS" are the same instrument (a
general stroke screen), distinct from VAN (a large-vessel occlusion
screen). - ACS Field Triage guidance reflects the American College of Surgeons' 2021
revision: penetrating injury is an Injury Patterns criterion, the
physiologic criterion is motor GCS below 6, and systolic blood pressure
thresholds are age-banded.
Privacy guidance
- One consistent rule for photographing anything that might contain patient
information, stated the same way everywhere it appears: redact identifiers
before you take the photo, not after. If you can't fully redact it, dictate
the values instead. Applies the same way to a monitor screen, an ePCR tab,
or paperwork.
A brief check-in before a full write-up
- Before a complete after-the-fact narrative, the assistant gives you one
short list of anything still open, so you can confirm, correct, or say you
don't recall it in one pass. Live handoff prep and prearrival notes stay
immediate. - Anything the assistant calculates on your behalf (for example, an estimated
gestational age worked out from a due date) is always flagged for your
confirmation, the same as anything else unverified.
Setup instructions
- ChatGPT and Gemini mobile setup: building or editing a Custom GPT or a Gem
is a web-only feature on both platforms; the mobile apps are for using one
you've already set up. - Microsoft Copilot: a declarative agent (Copilot Studio / Agent Builder) can
hold a persistent knowledge source, as an alternative to the per-session
paste workflow already documented.
New dictation references
- The printable pocket card is sized as an actual card.
- Two new formats of the same 12-point dictation skeleton: a wallet-size
business card, and a tall reference image sized to view on a phone screen
while you're dictating.
Agency configuration template
- Section 2 is now Chief Paramedic Endorsement. Existing configuration files
need no changes; use the current field names the next time you revise one.
Before you use v2.0.0
A small number of additional clinical-reference items are open for Chief
Paramedic confirmation and unchanged in this release. This tool documents
your clinical reasoning; it does not supply it.
Getting the update
- Claude: replace the Instructions field contents with the new
system-prompt.md, and replaceSKILL.md(and any changed reference files)
in Files. If your account uses the separate Skill package instead, download
the newparamedic-narrative.skillfrom the Releases page and replace it
under Settings > Skills (browser required for that step). - ChatGPT: replace the Instructions field with the new
chatgpt-instructions.md, and replace the uploadedSKILL.mdKnowledge
File — don't leave the old one in place alongside it. - Gemini: replace your Gem's instructions with the new
system-prompt.md. - Microsoft Copilot: update your Custom Instructions and use the new
SKILL.mdfor your per-session paste going forward.
Your provider profile and agency configuration files are unaffected — only the
skill itself needs to be replaced.
v1.5.0 -- Prearrival notification note
On request during transport (say notification prep, prearrival note, or the platform name, e.g., Pulsara note), the skill now produces a compact copy-paste block matched to prearrival notification platform fields: patient type mirroring the provider stated working impression (never assigned when the provider has not stated one), a one-line chief complaint, a brief Narrative/Notes block with age and sex, presentation, key findings, latest vitals as provided (for platforms that auto-extract vitals and demographics from the note), treatments and response, and ETA, plus destination as stated.
No patient name or date of birth passes through the AI session; identifiers are entered directly into the notification platform by the provider. The v1.4.0 concurrent-use guardrails apply in full: assembly only from facts already collected, patient care precedes documentation, and the provider verifies every element before sending. The skill never selects activation type, acuity category, or destination.
The agency configuration gains a Prearrival Notification Platform field (Section 8) and an ON/OFF prompt setting (Section 7). The attached .skill package is rebuilt from v1.5.0 sources.
v1.4.0 -- IMIST-AMBO handoff prep and retrospective training example
Two additions centered on transfer of care.
Concurrent intake and handoff prep. Fragments can now be fed during transport, not only after the call. On request (say handoff prep or IMIST-AMBO now), the skill assembles a spoken-style IMIST-AMBO report from the facts collected so far, with uncollected elements listed in a single line and never filled in. Guardrails are explicit: patient care precedes documentation, the skill never solicits input mid-call, it assembles only facts the provider already collected (it never suggests assessment, treatment, or destination), and the provider verifies every element before speaking it to a receiving clinician.
Retrospective handoff example (training stimulus). Because the chart is written after transfer of care, every completed draft now closes with a labeled TRAINING USE ONLY model of what a structured IMIST-AMBO handoff for that call would sound like, built solely from provider-supplied data with [VERIFY] carried through. It models structure and never critiques the handoff actually given. Agencies can turn either feature off in the configuration (Section 7).
All platform files updated; chatgpt-instructions.md remains within the 8,000 character limit. The attached .skill package is rebuilt from v1.4.0 sources.
v1.3.0 -- Multi-format support, photo plus dictation intake, delayed recall
Learnings from field usage, now native in the skill:
Narrative formats. Fourteen formats supported natively: SOAP (default), SOAPE, SOAPIER, CHART, CHARTE, DCHART-E, CHRONOLOGICAL, HEAD-TO-TOE, DRAATT, AT CHART, FACT, dedicated REFUSAL/NON-TRANSPORT and IFT templates, and CUSTOM agency-defined section orders. Format is declared in the agency configuration with per-call override. Section definitions and quality checklists in the new references/narrative-formats.md.
Photo plus dictation intake. Monitor screens, ePCR screens, medication vials, facility paperwork, and handwritten notes are transcribed exactly and verified with the provider -- never inferred from blur ([ILLEGIBLE] rule), with photo-vs-dictation conflicts flagged as discrepancies.
Suggested verbal report format. A 12-point dictation skeleton; printable pocket card in docs/ (text and one-page PDF).
Asynchronous and delayed recall support. Fragment accumulation across sessions, resume-anywhere status lines, memory-jogging recall questions anchored to what is known, gap surfacing by call type, honest gaps ([VERIFY], never filled), and late-entry notation per agency configuration.
All platform files updated: SKILL.md, system-prompt.md, chatgpt-instructions.md (7.0k chars, within the 8k limit), agency-config-template.md, README.md. The .skill package attached here is rebuilt from the v1.3.0 sources.
v1.2.0: Add Copilot support, mobile setup instructions
Paramedic-Narrative v1.2.0
What's new
Microsoft Copilot support
Paramedic-Narrative now works on Microsoft 365 Copilot. The setup is a hybrid: chatgpt-instructions.md saves once into Copilot's Custom Instructions field (persistent), and paramedic-narrative/SKILL.md is pasted at the start of each session. The behavioral rules persist. The clinical reference standard loads per session. This is a limitation of M365 Copilot's architecture, not of the tool -- there is no persistent Knowledge File feature equivalent to ChatGPT Custom GPTs. Full setup instructions are in the README and on the website.
Mobile setup instructions (iOS and Android) The README now includes platform-specific setup instructions for getting files from GitHub on a phone, installing on Claude, ChatGPT, Gemini, and Copilot via mobile app, and using the tool on shift with voice input, camera input, and ePCR transfer. Key finding: the GitHub mobile app does not support raw file downloads -- use Safari or Chrome on your phone instead and tap Raw on any file to access the plain text.
Distribution table updated
The multi-platform file table now correctly maps chatgpt-instructions.md to ChatGPT and Copilot, system-prompt.md to Gemini and direct API, and SKILL.md to ChatGPT Knowledge File and Copilot per-session paste.
Files changed
| File | Change |
|---|---|
| README.md | New Copilot installation section, new Mobile installation section, updated quick-start list, updated distribution table |
| chatgpt-instructions.md | Version bump only |
| system-prompt.md | Version bump only |
| paramedic-narrative/SKILL.md | Version bump only |
| ETHICS.md | Version bump only |
| LICENSE.md | Version bump only |
| agency-config-template.md | Version bump only |
| provider-profile-template.md | Version bump only |
| paramedic-narrative.skill | Repackaged at v1.2.0 |
Upgrading from v1.1.0
- Claude users: download the new
.skillfile and replace the existing one in your Project - ChatGPT users: no change needed -- your Custom GPT setup from v1.1.0 continues to work
- Gemini users: no change needed
- Copilot users: new platform -- see README Installation section for setup steps
- Agency administrators: version bump only, no configuration changes required
Known limitations
Microsoft Copilot: M365 Copilot Custom Instructions has a character limit that system-prompt.md exceeds. Use chatgpt-instructions.md in the Custom Instructions field. SKILL.md must be pasted at the start of each session. Some organizations disable Custom Instructions through IT policy -- if this applies to you, paste both files per session.
GitHub mobile app: Raw file downloads are not supported in the GitHub mobile app. Use your phone browser (Safari on iOS, Chrome on Android) to access raw file content.
v1.1.0 -- ChatGPT two-file setup
What's new in v1.1.0
ChatGPT two-file setup
ChatGPT Custom GPTs have a character limit on the Instructions field that the full system prompt exceeds. This release adds a condensed chatgpt-instructions.md (5,743 characters) designed for the Instructions field, with paramedic-narrative/SKILL.md uploaded as a Knowledge File. The full clinical reference standard is preserved -- it moves from the instructions field to the knowledge file where ChatGPT retrieves it during conversation.
Files changed
chatgpt-instructions.md-- new, condensed instructions for ChatGPT Custom GPTssystem-prompt.md-- header updated to clarify this is for Gemini, API, and platforms without character limitsREADME.md-- ChatGPT installation section rewritten with two-file setup instructions- All files -- version bumped from 1.0.0 to 1.1.0
Upgrading from v1.0.0
- Claude users: download the new
.skillfile and replace the existing one in your Project - ChatGPT users: replace your Instructions field content with
chatgpt-instructions.mdand uploadSKILL.mdas a Knowledge File - Gemini users: no change needed,
system-prompt.mdcontinues to work as before