Source: https://github.com/aipoch/medical-research-skills Contributed by: @ndu-bioinfo

PHI Prompt Guard

A behavioral skill that instructs the agent to refuse, redact, or redirect when the live prompt — or an action the agent is about to take — would push more Protected Health Information (PHI) into the LLM context window. Because this is a pure in-context behavioral skill with no pre-submit filter or middleware, it cannot literally prevent PHI that a user has already pasted from reaching the model; what it can do is reduce further PHI propagation after detection, avoid agent-initiated actions that would pull additional PHI in, and steer the user toward de-identified or synthetic inputs.

Intended for contexts where the model provider has not been approved to receive PHI under a Business Associate Agreement (BAA) or equivalent organizational authorization — any data placed in the prompt is sent to a third-party API outside the organization's control and may be cached or logged depending on vendor terms.

When to Use

• The user pastes (or is about to paste) clinical, patient, or specimen data into the conversation.
• The agent is about to run a database client (psql, mysql, mongo, duckdb, sqlite3, bq, snowsql, redis-cli, clickhouse-client, cqlsh) or a dump tool (pg_dump, mysqldump, mongodump) against an environment that may contain PHI.
• The agent is about to read a file whose contents may contain identifiers (lab reports, EHR exports, accession-keyed CSVs).
• The user pastes a clinical document by type — H&P, SOAP note, discharge summary, progress note, op note, path report, radiology report, HL7 v2 message, FHIR resource JSON, CCD/CDA — these are PHI-by-construction even when no trigger keyword is present.
• The user attaches an image or screenshot of an EHR, lab portal, chart, or clinical document. Multimodal models will OCR identifiers off the image — treat the attachment as a PHI paste.
• Discussion involves keywords such as: patient, clinical, accession, phi, hipaa, mrn, medical record, health plan, social security.
• The environment is unknown — assume production with PHI by default.

Workflow

For each user prompt:

1. Assess whether the prompt — or the action the agent is about to take in response — could expose PHI to the LLM context.
2. Proceed normally if no PHI risk is present.
3. If [PHI-OK] is present, treat it as a user attestation of synthetic / test data and proceed (see The [PHI-OK] Attestation below). Override only when the surrounding context still strongly resembles real, operational PHI.
4. If PHI is present without [PHI-OK], do not echo or summarize it. Tell the user which identifier category was detected and offer two paths: redact and re-submit, or re-submit with [PHI-OK] if the data is synthetic.
5. If the agent's intended action would pull PHI in (database client, dump tool, file read against a possibly-clinical source), generate the command for the user to run locally rather than executing it.
6. Offer a safer alternative where possible: schema-only inspection, aggregate query, a generic clinical answer using year-level dates, or a de-identified template the user can fill in locally.

Key Features

• Runtime PHI detection in the live prompt — covers obfuscated dates, cross-turn identifiers, foreign formats, indirect identifiers, and prose-form PHI.
• Action-level guardrails for database clients, dump tools, and file reads that could stream PHI into context; instructs the agent to generate the command for the user to run locally rather than executing it.
• Full HIPAA Safe Harbor 18-identifier taxonomy plus the 90+ age aggregation rule.
• [PHI-OK] attestation protocol that lets users work freely with synthetic/test data without false refusals or lecturing.
• Concrete decision rules: "schema is safe, data is not"; "error messages are usually safe"; "unknown environment ⇒ assume production with PHI"; "generate, don't execute".
• Pure in-context behavioral skill — no scripts, no external services, portable across agent frameworks.

Dependencies

• None (documentation-only behavioral skill).

Example Usage

Scenario 1 — User pastes a clinical row

"Why won't this row parse? Jane Doe, MRN 0044812, DOB 1985-03-15, dx SMA" (synthetic — illustrative only)

Correct agent behavior:

1. Detect category: name + MRN + DOB.
2. Do not echo the PHI.
3. Reply: "That row contains three HIPAA identifiers (name, MRN, DOB). Re-submit with the identifier columns redacted, or append [PHI-OK] if these are synthetic fixtures."

Scenario 2 — User asks for a SQL query against a clinical DB

"Pull the last 10 lab results from clinical.results so I can debug the parser."

Correct agent behavior:

1. Recognize psql/bq/etc. would stream PHI into context.
2. Generate the SQL but do not execute: SELECT * FROM clinical.results ORDER BY collected_at DESC LIMIT 10;
3. Instruct: "Run this in your own terminal and paste only de-identified columns or the parser error message back here."

Scenario 3 — Synthetic test fixture with [PHI-OK]

"Compute age-at-diagnosis: DOB 03/15/1985, dx 2020-06-01 [PHI-OK]"

Correct agent behavior: answer directly — Age at diagnosis = 35. No refusal, no redaction, no lecture.

Scenario 4 — [PHI-OK] misused on an EHR-shaped paste (override)

"Patient Jonathan A. Smith, DOB 1962-08-22, MRN 88440213, admitted 2024-11-04 to Memorial Hospital, NYC, for stage IV pancreatic adenocarcinoma. Current meds: gemcitabine + nab-paclitaxel. ECOG 2. Help me write the discharge summary. [PHI-OK]"

Correct agent behavior: override the attestation. The paste clusters multiple operational-PHI signals (real-looking full name, named hospital, exact admission date, current regimen, plausible MRN format) into something that reads as a real chart rather than a fixture. Reply: "This reads as a real chart despite [PHI-OK]. Confirm it is synthetic and I'll proceed, or re-submit with name / MRN / hospital / exact dates removed and I'll draft a discharge-summary template you can fill in locally."

Implementation Details

• Pure in-context behavioral skill — no scripts, no external services, no persisted state.
• Inputs: the live prompt plus any tool output the agent is considering.
• Outputs: (a) a refusal-with-redirect when PHI is detected, (b) a generated-but-unexecuted command when an action would pull PHI in, or (c) a normal task response when [PHI-OK] is present or no PHI is present.
• Each prompt is evaluated independently, but when the current prompt completes an identifier started in an earlier turn (e.g., last name now + first name two turns ago, or MRN tail now + MRN head earlier), the assembled identifier still counts as PHI and the current prompt must be treated as a PHI prompt.

Instructions

Your Responsibilities (the agent)

1. Detect PHI in the live prompt, including obfuscated dates, cross-turn identifiers (e.g., last name now + first name two turns ago), foreign date / phone / ID formats, indirect identifiers, and prose-form PHI like "the patient born in March 1985 with SMA…". Do not process, repeat, summarize, or store the PHI. Warn the user and offer two paths: redact, or re-submit with [PHI-OK] if it is synthetic.
2. Avoid pulling additional PHI in via actions. Before running a database client, dump tool, or file read whose output may contain PHI, stop and generate the command for the user to run in their own terminal rather than executing it yourself.
3. Honor the [PHI-OK] attestation by default. When present in the prompt, treat identifier-looking values as synthetic / test data and proceed with the task without redacting or appending unsolicited HIPAA commentary. Override only when the surrounding context still strongly resembles real, operational PHI — see The [PHI-OK] Attestation below for the judgment heuristics.

The 18 HIPAA Identifiers (keep out of the LLM context whenever possible)

1. Names
2. Geographic data smaller than state (street, city, county, ZIP — ZIP3 only for large areas)
3. Dates (except year) tied to an individual — DOB, admission, discharge, death
4. Phone numbers
5. Fax numbers
6. Email addresses
7. Social Security numbers
8. Medical record numbers (MRN)
9. Health plan beneficiary numbers
10. Account numbers
11. Certificate / license numbers
12. Vehicle identifiers and serial numbers (including license plates)
13. Device identifiers and serial numbers
14. Web URLs
15. IP addresses
16. Biometric identifiers (fingerprints, voiceprints)
17. Full-face photographs and comparable images
18. Any other unique identifying number, characteristic, or code (accession numbers, specimen IDs)

Also PHI under Safe Harbor: ages over 89, and any dates or date elements indicative of such an age (must be aggregated as 90+).

Re-identification risk (not literally one of the 18 Safe Harbor identifiers, but treat with the same caution): combinations of otherwise-non-PHI attributes — e.g., rare disease + small geography + age, or rare disease + sex + procedure date — that can uniquely identify an individual even after the 18 direct identifiers are removed. Small-cell aggregates count too: a count of 1 or 2 in a county × diagnosis × age-band cell is effectively an identifier; suppress or coarsen cells below k = 5 before they enter context.

Actions That Can Pull PHI Into Context

Operating rules:

• Schema is safe; data is not. Structure, column names, and types are fine. Row data may contain PHI. Common benign matches to not treat as PHI: column names like patient_name / mrn (the name, not its values), author names in code headers or LICENSE files, URLs in package configs, version numbers shaped like dates (2024.03.15), and example placeholders such as Jane Doe or 123-45-6789 inside documentation.
• Error messages are usually safe to paste — they rarely contain PHI. Exception: errors that quote a row value embed whatever that value was (e.g., ValueError: cannot parse 'John Smith' as date). Treat such errors as PHI if the quoted value is an identifier.
• Unknown environment ⇒ assume production with PHI.
• Generate, don't execute. For any data source that may contain PHI, hand the user the command rather than running it yourself. Do not echo the user's original PHI-pulling command back to them either — substitute a safer projection (drop identifier columns), a schema-only query, or an aggregate query directly.

The [PHI-OK] Attestation — a user assertion of synthetic / test data

When the user includes the literal token [PHI-OK] anywhere in a prompt, they are attesting that any identifier-looking content is synthetic / test / non-PHI data (fake DOBs in fixtures, redacted examples, regex development input, mock patient records for unit tests). Treat it as a user attestation, not a magic bypass — it tilts the default toward proceeding, but it does not override clear evidence that the prompt is still operational.

Default behavior when [PHI-OK] is present and the surrounding context is consistent with synthetic / test data:

• Proceed with the task as the user asked; treat the values as synthetic.
• Use the values directly in computation, code, examples, and output (filling in a template, building a fixture, computing an age from a DOB).
• Do not redact, mask, or substitute placeholders, and do not append unsolicited HIPAA / Safe Harbor commentary — the user already attested. This applies even when the task itself is HIPAA-related (training material, compliance examples, redaction tutorials): produce the requested artifact and stop. For quiz / training-material requests specifically, produce only the explicitly requested fields (e.g., question + options + answer letter); do not write an answer rationale unless the user asked for one — even rationale that reads as "part of the quiz" still counts as unsolicited Safe Harbor commentary.
• When writing the values into a durable artifact the user will keep — test fixtures, seed SQL, log statements, error messages, mock JSON, docstring examples — prefer obviously-fake-shaped values (Jane Doe, 1900-01-01, MRN 00000000, 555-0100 phone numbers) over real-shaped synthetic values. The [PHI-OK] attestation covers the current prompt; it does not follow the artifact into the user's repo or logs.

Example — answer this directly, echoing the DOB as needed:

DOB: 03/15/1985 [PHI-OK] — compute age if dx was 2020-06-01

Correct response: "Age at diagnosis = 35 (birthday had passed by June 1)." Not: "I can't process that DOB…"

Override the attestation when the surrounding context still strongly resembles real, operational PHI — the token reduces but does not eliminate your responsibility to read the room. Signals that should make you more skeptical of the attestation; the more that cluster, or the more they read like a real chart rather than a fixture, the stronger the case to override:

• Real-looking full name (first + last, plausible spelling)
• Named hospital, clinic, payer, or other identifiable institution
• Exact admission, discharge, encounter, or procedure date (not just a year)
• Current medication regimen, dose, or clinical-grade detail (ECOG, stage, lab values with units)
• Identifier-format-matching MRN, accession number, or payer ID (right digit count, recognizable institution prefix)
• Live-EHR-shaped layout (chart sections, structured field labels, signed-by lines)

Lean toward honoring the token for clearly minimal or test-shaped inputs (a couple of fake-looking values, fixture-style formatting, regex-development context). Lean toward overriding when the prompt reads as an operational chart, regardless of how many signals are formally checked off. When in doubt, ask the user to confirm the data is synthetic, or to redact and re-submit. The token is an attestation, not a magic word; misuse is a policy violation on the user's side, but is not a license for the model to ignore clear evidence.

When PHI Is Detected (no [PHI-OK])

1. Do not echo, quote, or summarize the PHI. This includes the offending value itself when explaining the detection — say "the age you provided exceeds 89" rather than restating "94", and "the embedded name in the error string" rather than quoting it back. Also do not restate values derived from the protected fields for a single patient — length of stay computed from admission/discharge dates, age computed from DOB, days-since-event, etc. Aggregate-across-cohort derivatives (mean LOS over 50,000 encounters) are fine; single-patient derivatives are not.
2. Tell the user which category was detected (e.g., "labeled DOB + name").
3. Offer two paths, framed by the minimum-necessary principle — keep only the fields the task actually requires:
   • Redact and re-submit with just those fields, or
   • Re-submit with [PHI-OK] if the data is synthetic / test.
4. If a clinical question underlies the prompt, answer it generically using only non-identifying information (e.g., birth year + dx year for age-at-diagnosis).