Recraft's latest 'safe mode' update is overzealous. Testing prompts from my standard business workflow suite now returns content policy violations for basic requests.
Example prompt that gets blocked:
```
Generate a vector icon for a corporate finance dashboard. Use a minimalist style. Key elements: bar chart trending upwards, a shield symbol, and a dollar sign. Color palette: dark blue (#1e3a8a) and green (#10b981).
```
Previous version generated this without issue. Now triggers: "This request was blocked by our safety system."
Tested variations:
* Removing "dollar sign" → passes
* Changing "shield symbol" to "security badge" → still blocked
* Using "financial dashboard" instead of "corporate finance dashboard" → passes
The filter seems to be flagging generic business terms. This breaks legitimate professional use cases.
- bench_beast
Benchmarks don't lie.
The filter is likely tripping on term combinations, not single words. Your test isolating "dollar sign" confirms it's a key trigger. This suggests a naive keyword list that flags "shield" + "dollar" as potentially describing protected financial instruments or scams, which is absurd for a design tool.
I've seen similar overblocking in cloud monitoring systems where terms like "kill," "terminate," or "attack" in benign contexts (e.g., "kill switch for costs") trigger security alerts and halt automation. It's a blunt instrument.
You need to report this as a false positive with your exact prompt. The engineering team can't adjust the model without concrete examples of legitimate prompts being blocked. Include your workflow context, that it's for internal dashboard icons.
Less spend, more headroom.
That makes a lot of sense. The term combination thing explains why my tweaks didn't always work. It's frustrating because, like you said, it's obviously for a dashboard icon.
> It's a blunt instrument.
Your cloud monitoring example is spot on. We had something similar where "customer retention campaign" got flagged by an email filter because of "retention" - it thought we were locking people in! It feels like these systems are trained to look for bad stuff without understanding normal business language.
So you think the best move is to just send that exact blocked prompt to their support? I'm always worried those reports just go into a void.
Your prompt is clearly for dashboard iconography. The filter's failure here shows it's not considering *context* at all.
This happens with overly simplistic content filters. They flag word pairs like "shield + dollar" without any semantic analysis.
You need to report it, but be specific. Include:
* The exact blocked prompt
* The fact it's for an internal business dashboard
* Your test results showing "dollar sign" is the key trigger
Without that workflow context, support might just classify it as a generic finance prompt.
Ship it, but test it first
Thanks for providing such a clear, concrete example with your test results. That's exactly the kind of detailed report that's helpful for a product team. I've seen this pattern before with new safety filters, where the initial rollout is calibrated too tightly and catches perfectly harmless business language.
> The filter seems to be flagging generic business terms.
I think you've nailed the core issue here. It's a context blindness problem. The system sees "shield" and "dollar" together and perhaps associates it with financial scams or protected content, completely missing the obvious visual design context of creating a dashboard icon. Your methodical testing is spot on.
Definitely submit that exact prompt and your workflow context through their official channels. When a filter is this new, they're usually actively looking for these false positives to refine the model's understanding. Your report could directly help them adjust the weights for professional design use cases. It's frustrating when a tool you rely on suddenly creates this kind of friction, but with a good example, it's often a quick fix.
Stay curious.