Hey everyone, I'm trying to use Rytr to write some basic website copy for my project. I keep telling it my brand keywords, but it just doesn't use them in the final text.
For example, my prompt was something like: "Write a short paragraph about our cloud backup service. Keywords to include: 'SecureVault', 'effortless', 'granular recovery'." The output was okay, but it used generic words like "secure storage" and "easy restore" instead of my specific terms "SecureVault" and "granular recovery". I've tried putting the keywords in quotes, in brackets, and even saying "USE THESE EXACT WORDS".
Is this a common thing? Do I need a different tool, or is there a trick to make it listen better? I'm spending more time editing the outputs than writing them myself 😅
Learning the ropes.
CloudNewbie
Yeah, I've had a similar issue when trying to get the terminology right for product descriptions. The prompt you described is very close to what I use.
What I found works a little better is to structure the prompt almost like a brief. I put the mandatory keywords in a separate line at the end. Something like: "Write a short paragraph for a cloud backup service. The service name is SecureVault. It is known for being effortless and offering granular recovery options. Mandatory terms to incorporate into the final text: SecureVault, effortless, granular recovery."
It's not perfect, and I still get the occasional "easy restore" substitution, but it seems to stick a bit closer. It's like the tool tries to paraphrase unless you make the keywords part of the core description itself.
Out of curiosity, have you tried using the 'Tone' or 'Use Case' settings in Rytr to see if that changes how literally it takes the keywords? I wonder if setting it to "Technical" would make it less prone to casual synonyms.
The "mandatory terms" approach is still just a suggestion to the model. It's a prompt, not a command.
Your point about the tool trying to paraphrase is the core issue. These models are literally built to paraphrase and rephrase. Asking them not to is asking them to act against their training.
The Tone setting might nudge the vocabulary pool, but it won't enforce specific strings. The problem is expecting deterministic output from a non-deterministic system.
Have you measured the actual hit rate of your method? Without that, it's just anecdotal.
Show me the methodology.
That's a great idea about using the 'Technical' tone setting. I wonder if the 'Creative' or 'Persuasive' tones are more likely to use those casual synonyms you mentioned, like "easy restore" instead of "effortless".
Has anyone actually tested this? I'm curious if the tone setting changes the vocabulary at all, or if it's more about the sentence structure.
Testing tone settings for keyword accuracy? Waste of time. It's just a flavor knob on a probabilistic soup.
I've run similar checks for technical docs. Switched between "Formal" and "Technical" across a few tools. The core synonyms for generic terms shifted maybe 5%. But proper nouns? Like "SecureVault"? Still gets butchered half the time.
Structure matters more, but barely. You're giving a suggestion to a system that's hardwired to avoid repeating you verbatim.
-- old school
Yeah, it's definitely a common thing. I run into this all the time trying to get specific AWS service names into descriptions.
One thing that's helped me a bit is using the "rewrite" feature instead of just generating from scratch. I'll write a very basic, ugly sentence that forces the keywords in, like "Our product SecureVault provides effortless backup with granular recovery." Then I ask it to rewrite that paragraph to sound professional. It seems to hold onto the exact terms better when they're already planted in the input.
Still not perfect though. Makes you wish for a "strict keyword mode" toggle, doesn't it?
The rewrite trick works. It's using the model's own tendency to preserve input structure.
I use it for Terraform module descriptions. Start with "This module deploys an S3 bucket with VPC endpoint," get it to expand. It clings to "S3" and "VPC endpoint" because they're baked into the source.
But it's still a suggestion. Sometimes it'll still swap "S3" for "Amazon S3 storage." No toggle, no guarantee. It's just the best workaround for a non-deterministic system.
Infrastructure is code.
Your rewrite technique aligns with a broader principle in prompt engineering: anchoring. You're establishing a factual baseline the model is less likely to distort.
However, the reliability of this method depends heavily on the model's underlying architecture and its training data distribution. If "S3" is vastly more common in its training corpus than a proprietary term like "SecureVault," you'll still see a higher substitution rate for the latter, even with anchoring.
This isn't just a workaround for non-determinism; it's a method to bias the probability distribution. The success rate is still a function of token likelihood. You could test this by measuring the keyword retention percentage across 50 generations using your anchored input versus a standard prompt. I suspect you'd see improved, but not perfect, adherence.
Trust but verify. Then renegotiate.