Ask any model for a renewal email and it will hand you something perfectly fine. Grammatically clean, on-topic, faintly damp. The problem is that "fine" is the ceiling of a first draft, and most people stop there because the output *looks* finished.
The second pass is where the quality lives. Not you rewriting it. The model critiquing itself, against a standard you set, before you ever read it.
Here's the mechanic: models are much better at evaluating text than generating it. Generating is a guess about what comes next. Evaluating is a comparison against criteria, and comparison is the thing these systems are genuinely good at. So you split the job in two. Draft, then judge, then revise. Same conversation, three moves.
The move that does most of the work
After it drafts, don't say "make it better." That's a coin flip. Say this instead:
> Before I read that, score it 1-10 on each of these: specificity to this property, whether a resident would actually open it, whether the ask is clear in the first two lines, and whether any sentence could appear in a competitor's email unchanged. For any score under 8, quote the exact sentence that lost the points and say why. Then rewrite.
Two things make that work. You named the criteria, so it isn't inventing a standard that flatters the draft. And you made it quote the offending sentence, which forces it to point at something real instead of writing a vague self-review that reads like a performance appraisal.
The "could appear in a competitor's email" test is the one that earns its keep. Run it on almost any AI-written property copy and watch it flag its own opener. "Nestled in the heart of" is not describing your community. It's describing 40,000 communities.
Make it argue against itself
Scoring catches flabby writing. It doesn't catch bad thinking. For that, you want an adversary.
Say you've had it build a concession recommendation for a lease-up sitting at 61% with 14 units to go before the summer slowdown. It gives you a confident plan: one month free, applied at move-in, promoted through the ILS listings and paid social. Reasonable. Now:
> Argue the opposite case. What would a skeptical asset manager say about this recommendation, specifically about the effect on renewal rates in 12 months, on existing residents who paid full rate, and on how the ILS listing prices out against the four comps within a mile? Then tell me which of those objections you can't answer.
That last clause is the whole trick. "Which of these can't you answer" is where the model stops being agreeable and admits the recommendation rests on a rent-growth assumption nobody has checked. You walk into the call having already survived the hard question. That's a different meeting than the one where the regional asks it first.
The rubric you reuse
The self-critique gets sharper when the criteria are yours and they don't change. Write a rubric once, per output type, and paste it every time.
For an ILS listing description, mine would be: 1. Every claim is verifiable from the property fact sheet. 2. Describes the unit and the community, never the person who should live there. 3. Amenities named specifically. "In-unit washer/dryer, full size" not "modern conveniences." 4. Fewer than 5 adjectives total. 5. No sentence that would work for any other property in the submarket.
Number two is not a style preference. Models default to writing about the resident, because that's what marketing copy on the internet does. "Perfect for young professionals," "ideal for a growing family," "great for students." Every one of those describes a protected class or proxies for one, and it's a fair housing problem the moment it publishes to your ILS feed. The fix is one instruction, held every time: describe the unit, the community and the lease terms. Never the person. Put it in the rubric so the critique pass catches it even when your prompt forgot to say it.
Where it breaks
Self-critique has a real failure mode: the model will find problems whether or not they exist, because you asked it to. Ask for three weaknesses and you'll get three, including one it invented to fill the slot. So ask for "any weaknesses, and say 'none' if there are none," and check that it's ever said none. If it never does, your criteria are too vague.
It also can't catch factual errors about your property. If it invented a dog park, the critique pass will happily assess how well-written the dog park sentence is. Facts come from you. Judgment is what you're renting.
Run it today
Take the last thing you shipped. A renewal campaign, a tour follow-up sequence, a Google Ads description set. Paste it in cold, with your rubric, and ask for scores with quoted evidence.
You're going to disagree with about a third of what comes back. Good. The third you agree with is free, and you'd have shipped it anyway.