Job Prompts Compliance Pass Fair Housing Phi No Invented Stats
A job prompt that can write a listing, a patient email, or a case study also can discriminate, leak health details, or invent a statistic that sounds true. The fix is not a longer pep talk. It is a second pass with a stop list.
Write the draft with your usual job prompt. Then run a compliance pass that is allowed to delete, flag, and say "I do not have that number." Matching compliance-pass prompts are landing in the PromptDig library (Browse more prompts). Put them next to the job prompt, not inside it.
This is not legal advice and not a HIPAA program. It is a way to stop the model from being casually reckless before a human reviews the work.
The job prompt drafts. The pass audits.
If you cram Fair Housing, PHI, and source rules into the same prompt that is trying to be persuasive, persuasion wins. Models optimize for a complete-looking artifact. A complete-looking rental ad is exactly how you get "great for young professionals" and a made-up crime rate in the same paragraph.
Split the work:
- Job prompt: produce the artifact from facts you supplied
- Compliance pass: read the artifact against a checklist, output a marked copy plus a list of flags
- Human: decide. The pass does not publish.
The pass should not "rewrite for punch." Punch is how the banned phrase comes back. The pass is a red pen.
A skeleton for the pass:
You are a compliance editor, not a marketer.
Do not add new claims. Do not invent numbers, quotes, or reviews.
Read the draft against this checklist: [Fair Housing / PHI / sources].
Output:
1) Clean copy with banned material removed, not rephrased into a wink.
2) FLAG list: line, why, what I must supply if we want the idea back.
3) If a number, diagnosis, or protected-class implication is not in the source notes, delete it.
If you are unsure, FLAG. Do not smooth.
Fair Housing: describe the property, not the people you want
Housing copy goes wrong when it selects tenants by implication. Familial status, disability, race, religion, sex, national origin, and the other protected bases are not "tone." They are the point of the pass.
Ban lists that belong in the prompt (plain language, not a statute dump):
- Who would "fit" or "feel at home"
- Churches, schools, or "safe for" a kind of person as a selling point
- Ability or family composition as a vibe ("walk-up, so not for everyone with kids")
- Source-of-income snobbery if your jurisdiction treats that as protected
- Made-up neighborhood crime or school scores
Allow: square feet, rent, deposit, pet policy as a rule, stairs as a fact, distance to a transit stop as a fact, photos of the actual unit.
The existing Honest Rental Listing Writer on PromptDig is the job prompt half of this: facts, flaws, photo notes. Pair it with the pass. The job prompt should already refuse a target tenant. The pass catches the adjectives that snuck through ("vibrant young energy," "perfect for a couple").
If the pass cannot verify a neighborhood claim from your notes, it should delete the claim. "Quiet street" needs to be your words from a walkthrough, not the model's mood.
PHI: minimum necessary, no extra medical plot
For anything that might touch health (clinic emails, care-nav copy, a case study, a benefits explainer), the failure mode is extra detail. The model will add a diagnosis, a medication, a date of service, or a family history because it makes the story coherent.
Rules to put in the pass:
- Use only identifiers and clinical facts present in the source notes
- Strip extra patient detail that is not needed for this artifact
- No "a 47-year-old woman with..." unless that sentence is already approved
- No portal screenshots, record snippets, or "paste this into the EHR"
- If the job is education, talk about a condition in general, not about a person
Say "this is not a medical record and not medical advice" only if that disclaimer belongs on the artifact. Do not use a disclaimer as permission to get specific.
If your team is a covered entity, this prompt does not replace policy, BAAs, or training. It stops the draft from arriving already wrong.
No invented stats: TODO is a feature
Marketing and research prompts love a number. If you did not supply the number, the pass must remove it. Not hedge it. Remove it.
Examples the pass should kill on sight:
- "9 out of 10 customers"
- A dollar figure, a wait time, a success rate, a ranking
- A quote attributed to a named person or a "recent study"
- A chart in an infographic that was not in the brief
Replace with a TODO the human can fill: [NEED: source for refund timing]. If the piece cannot ship without the number, it should not ship with a generated one.
If a number is not in the brief, omit the callout. The second pass exists so the writer prompt cannot talk its way around that rule.
Store the pass next to the job, and run it every time
Name the files as a pair: job-rental.md + pass-fair-housing.md, job-clinic-email.md + pass-phi.md, job-case-study.md + pass-sources.md. If an artifact might hit all three, use one combined pass with a longer checklist, not a vaguer one.
Do not skip the pass because the draft "looks careful." Careful is the model's costume. Run it. Read the FLAG list. Supply a real fact or drop the claim.
When a pass actually catches something your team would have missed, share that checklist (Share a prompt) with the job name in the title. Browse the library on PromptDig (Browse more prompts) as the paired prompts land.
The point is not to make the model a lawyer. The point is to stop publishing a confident draft that selected a tenant, named a patient, or cited a study nobody ran. Draft, pass, human. In that order.