fair_housing_marketing_review v1.0.0

---

name: "fair_housing_marketing_review" description: "Review housing marketing for discrimination risk before it ships: scan listing copy, photos, scripts and ad targeting against the Fair Housing Act advertising rule with the bundled scanner, work passage-linked findings by protected class, check Meta and Google housing targeting limits, handle the older persons exemption and local source of income law, and rewrite flagged copy into neutral property facts that keep the selling points. Use when reviewing a rental or sales listing, a housing ad campaign, a flyer, a showing script, listing photos, or platform targeting settings."

---

Fair Housing Marketing Review

Purpose

Housing ads carry a legal rule most marketing does not: under 42 USC 3604(c) it is unlawful to publish a notice, statement or advertisement for the sale or rental of a dwelling that indicates any preference, limitation or discrimination based on race, color, religion, sex, handicap, familial status, or national origin. HUD's regulation at 24 CFR 100.75 applies the same rule to photographs, illustrations and symbols, to things said to agents and prospects, and to choosing media or locations that keep part of the market from learning a home exists. Courts judge the wording by the ordinary reader test: what a normal reader would take from it, not what the writer meant.

This skill runs that review as a workflow with tooling. scripts/scan_copy.py scans copy for flagged phrases and checks a JSON record of campaign targeting settings against the Meta and Google housing rules. The scanner starts the review. The human read in the workflow finishes it, because context is where most of these calls are decided.

Scope: this skill covers US federal law and the two big ad platforms, plus the local law checks that change the answer. It is a review workflow, not legal advice and not a clearance. Anything genuinely contested goes to counsel with the review record attached.

Workflow

1. Assemble everything that will be published, not just the listing text

Collect the full set before reviewing any of it:

• Headline, description, captions, flyer and sign text, and any email or social post built from them
• Photo and video set, including thumbnails, alt text and on-screen text in video
• Showing script, phone script, chatbot replies and any recorded narration
• Campaign settings: platform, category declaration, age, gender, geography, audiences, exclusions, and the note describing who the audience is meant to be
• The syndication list: every portal, feed and channel the material will run on

A phrase fixed in the listing and left live in the feed, the flyer or the chatbot is still published. Review the set as one thing.

2. Load the local protected class list before judging anything

The federal list is race, color, religion, sex, handicap (disability), familial status and national origin. HUD's April 2024 digital advertising guidance reads sex to include sexual orientation and gender identity, and reads familial status to cover children under 18 in the home, pregnancy, and a person securing legal custody of a child.

State and local law adds classes that change real answers: source of income, age outside the older persons exemption, marital status, sexual orientation and veteran or military status are common additions, and the list differs by state and city. Pull the current list for the property's jurisdiction from the state or local fair housing agency and write it at the top of the review record. Two scanner flags depend on it:

• Run with --soi-protected where source of income is protected. It turns no Section 8 and no vouchers from review items into blocks.
• Treat every federal review item that names a locally protected class as a block in that jurisdiction, and say so in the record.

3. Run the scanner on the copy

python3 scripts/scan_copy.py --copy listing.txt

Add --hopa-verified only when step 6 has verified an older persons exemption for this property. Add --json when the findings need to go into a file or a ticket.

The scanner reports two severities:

• Block: wording that states a preference, limitation or exclusion tied to a protected class on its face. Examples in the rule set: no children, no wheelchairs, Christian home, white family home, citizens only, men only, no assistance animals.
• Review: wording whose risk depends on context. Proxies (young professionals, quiet professional building), coded phrases (safe neighborhood, exclusive community), steering language (top schools, neighborhood is changing), age language without a verified exemption (senior community, mature tenants), and screening exclusions (no Section 8, must be employed).

Exit code 1 means at least one block. Exit code 0 with review items still means work to do. Exit code 2 means bad input.

The scanner deliberately does not flag wording HUD treats as acceptable: master bedroom, desirable neighborhood, walk-in closets, great view, walk to bus stop, kosher meals available, kids welcome, family room. A review that blocks those is overstating the law and teaching writers to ignore the review. references/phrase-table.md holds the full table, with fine items listed on purpose.

4. Do the human read the scanner cannot do

Read the whole set once as an ordinary reader who belongs to each protected class in turn, and check four things the phrase rules cannot see:

1. People described anywhere. Search the copy for every sentence about a person rather than the property: who lives there now, who the area suits, who should apply, who the seller hopes for. Each one is a finding, scanner hit or not.
2. Photos and video as a set. Who is shown enjoying this home, and who is shown in the photos for the other listings in the same campaign? Is an accessibility feature photographed or quietly left out? Selective human imagery is a statement under 24 CFR 100.75(c)(1), which names photographs and illustrations.
3. Differential information. Would every caller get the same facts, the same amenity list, the same schools answer, the same availability answer? Marketing that gives one audience the full picture and another a shorter one is steering with a distribution list.
4. Neighborhood talk. Any sentence about what an area is like, who is moving in, or how it is changing. Compare it against the blockbusting rule in 42 USC 3604(e): inducing a sale with representations about the entry of protected class persons into a neighborhood. Property facts with a source are fine. Demographic weather reports are not.

Write each finding with the exact passage quoted, the protected class, the source it rests on (statute, regulation, HUD guidance or platform policy), and the decision: cut, rewrite, or keep with the reason recorded.

5. Check the targeting settings

Write the campaign settings into the JSON schema in the scanner header and run:

python3 scripts/scan_copy.py --targeting targeting.json

Then work references/targeting-checklist.md by hand. The load-bearing platform rules:

• Meta: housing ads must be declared in the Housing special ad category at campaign creation. Audiences must include ages 18 to 65 plus and all genders. ZIP code targeting is unavailable, and a selected city, address or pin carries a minimum 15 mile radius in the US. Excluding detailed targeting selections is unavailable, and lookalike audiences are not offered for the category.
• Google Ads: in the US and Canada, housing ads cannot be targeted by gender, age, parental status, marital status or ZIP code, and demographic bid adjustments must be removed. City, region and radius targeting are allowed, radius at least 1 km.
• Any platform: an audience built from a variable highly correlated with a protected class is a proxy decision. HUD's 2024 guidance names ZIP code as the example and states that a decision made in whole or in part by an automated system can still violate the Act. Who never sees the ad is the question, and it has to be answered in plain words in the review record.

Check delivery after launch as well as settings before it. A skew visible in the delivery breakdown that nobody acts on becomes the advertiser's decision. Note what was checked, when, and what changed.

6. Work the exemptions and edge cases explicitly

Do not let an exemption ride on a label. Each of these gets a yes or no with evidence in the record:

• Older persons housing. Lawful age language needs a verified Housing for Older Persons Act position: for 55 or older, at least 80 percent of occupied units with a resident 55 or older, published policies showing intent to operate as older persons housing, and HUD compliant age verification on file. For 62 or older, occupancy solely by persons 62 or older. No paperwork, no age language, and adults only comes out regardless.
• Shared living. A roommate ad for a genuinely shared unit sits in different case law from an ad for a whole dwelling. Flag it, stop the gender or faith preference wording, and route it to counsel instead of editing around it.
• Religious names and symbols. A building or business name containing a religious reference, or a religious symbol standing alone, may indicate a preference under HUD guidance unless the ad carries a nondiscrimination statement. Add the statement.
• Services described factually. Kosher meals available, an on-site chapel, holiday greetings such as Merry Christmas: HUD guidance treats these as descriptions of services or secularized terms, not preferences. They stay, and they do not license a preference line next to them.
• Accessibility facts. Stairs, floor level, step-free entry, grab bars: state them truly. Facts about the property are always allowed. Judgments about who can manage them are not.
• The seller asked for it. A client instruction to aim copy at a kind of person, or to leave a group out, is itself a finding. 24 CFR 100.75(c)(2) covers expressing a preference to agents and brokers. Record the instruction, decline the wording, and offer the neutral rewrite.

7. Rewrite by swapping the person for the property fact

Every flagged phrase was doing a selling job. The rewrite keeps the job and drops the person:

| Flagged | What it was selling | Rewrite pattern | | --- | --- | --- | | perfect for young professionals | short commute, work space | name the commute time with its source, the desk space, the broadband fact | | ideal for a growing family | room to spread out | bedroom count, second living room, yard size, storage | | quiet professional building | low noise | insulation and construction facts, quiet hours that apply to everyone | | safe neighborhood | peace of mind | low-traffic street, lighting, the property's own locks and alarm | | walking distance to the church | walkable area | measured distances to named amenities, given neutrally | | great for empty nesters | single level ease | main floor bedroom and bathroom, step-free entry, low maintenance yard | | exclusive community | quality and scarcity | the real features: lot size, amenities, build standard, price facts | | top schools | school quality | assigned school and district by name, with the lookup source and date |

Rules for rewrites: use only facts you can source, keep the length and energy of the original line, and never soften an exclusion into a hint. A hint that survives the scanner and still tells a class of readers to stay away has failed the ordinary reader test, which is the test that counts.

8. Worked reviews

Worked review 1: a rental listing. The draft in examples/listing-bad.txt is three lines long and fails hard:

The scanner reports 4 block and 10 review. Blocks: adults only, no children, no wheelchairs, Christian home. Review items include perfect for young professionals, empty nesters, quiet professional building, safe neighborhood, walking distance to St. Mary's Church, mature tenants preferred, family-friendly, top schools and no Section 8. Run in a jurisdiction that protects source of income, the same file reports 5 block and 9 review, because no Section 8 becomes a block. The fixed version in examples/listing-fixed.txt reports 0 block and 1 review: the surviving item is the no pets pairing check, cleared in the rewrite by stating that assistance animals are accommodated. The rewrite kept every selling point that was real: sun, size, the 2023 kitchen, parking, the cul-de-sac, the station, the park, and schools given as assigned names with a dated district lookup instead of a ranking.

Worked review 2: the same copy for a verified 55 or older community. With exemption paperwork verified and the scanner run with --hopa-verified, the adults only line downgrades from block to a consistency check and the file reports 3 block and 11 review. The other blocks stand: an older persons exemption covers familial status only, so no wheelchairs and Christian home are still blocks, and no children still has to be reworded to the exemption's actual basis, housing for older persons, rather than a bare exclusion. The lesson: an exemption is class specific and paperwork specific, never a general pass.

Worked review 3: a Meta campaign. The settings in examples/targeting-bad.json describe a campaign with no category declaration, ages 28 to 55, women only, two ZIP codes, a 5 mile radius, a lookalike audience, exclusions for public housing interests, and an audience note aimed at homeowners and parents in affluent areas. The scanner reports 8 block and 1 review. Every block is a settings change, not a copy change: declare the category, open the age band to 18 to 65 plus, include all genders, drop the ZIPs, widen the radius to 15 miles, drop the lookalike and the exclusions, and rewrite the audience note around geography and listing views. The corrected settings in examples/targeting-good.json scan clean.

9. Sign off and keep the record

A review is done when the record holds: the material reviewed, the local class list used, the scanner output for copy and targeting, each human finding with its passage and decision, the exemption evidence where one was claimed, the final rewritten copy, and the name of the person who approved shipping it. Keep the record with the listing file. If a complaint or a tester ever asks how this ad was reviewed, the record is the answer.

Anything the reviewer cannot settle (a claimed exemption with partial paperwork, a shared living preference, a client insisting on flagged wording, a delivery skew with no clear cause) goes to counsel as an open item. The ad does not ship with the open item inside it.

Output Contract

Return all of the following:

1. The review record header: property, channels and platforms, jurisdictions, and the local protected class list used, with its source
2. Scanner output for the copy and for the targeting settings, with the flags used (--hopa-verified, --soi-protected) and why each was justified
3. Findings table: exact quoted passage or setting, protected class, source (statute, regulation, HUD guidance or platform policy), severity, and decision
4. Photo and media findings: imagery set, placement channels and language editions checked
5. The rewritten copy and the corrected targeting settings, each flagged item traced to its replacement
6. Exemption and edge case outcomes, with the evidence relied on or the gap that remains
7. Open items routed to counsel, stated plainly. An ad with an open item in it is reported as not cleared to ship

If the material is clean, say so in the same format with the checks named. Clean means the scanner ran, the human read happened, and the targeting was checked, not that nothing was looked at.

Operating Rules

1. Describe the property, never the person. Any sentence about who should live somewhere, who lives there now, or who an area suits is a finding until rewritten into property facts.
2. Intent does not clear wording. The ordinary reader test decides. Kindly meant preferences are still preferences.
3. No exemption by label. Senior, retirement, adult and 55 plus language ships only on verified paperwork for that specific community.
4. Targeting is part of the ad. Settings, exclusions, audiences and delivery get reviewed and recorded with the same care as the copy, on every platform, including portals, feeds, print and signs.
5. Proxies count. A variable that tracks a protected class closely is treated as the class until someone shows, in writing, why it does not in this use.
6. Local law can only add protection, never remove the federal floor. When federal and local answers differ, the stricter one governs the copy.
7. The scanner is a flagging tool, not a ruling. A clean scan plus no human read is an unfinished review. A scanner flag plus a context that genuinely clears it gets recorded with the reason, not silently dropped.
8. Never soften an exclusion into a hint. If the rewrite still tells a protected class of readers the home is not for them, it fails.
9. Do not invent law. Penalty amounts, case outcomes and platform rules enter the record only with a source checked during this review. Where a point is unsettled, the record says unsettled and the item routes to counsel.
10. Worked examples in this skill are fictional. Their addresses, ZIP codes, distances and dates are illustrations, never facts about a real property.

References

• Phrase table: wording by protected class with block, review and fine calls, the basis for each, and the HOPA conditions in full.
• Targeting and delivery checklist: Meta, Google, portal and print checks, the proxy questions for any audience, and delivery monitoring.
• Worked examples in examples/: listing-bad.txt, listing-fixed.txt, targeting-bad.json and targeting-good.json, used to test the scanner.
• Primary sources: 42 USC 3604 and 3607 at uscode.house.gov; 24 CFR 100.75 at law.cornell.edu; HUD, Guidance on Application of the Fair Housing Act to the Advertising of Housing, Credit, and Other Real Estate-Related Transactions through Digital Platforms, April 29, 2024; Google Ads personalized advertising and housing policy pages at support.google.com; Meta special ad category help, checked against the settings the scanner enforces.