The output is empty, or not what I asked for
Nine times in ten it is a placeholder that did not resolve, an image the step never received, or a prompt the model read more literally than you meant.
In this article (6)
Start in the same place every time: the Data tab on the step, which shows the exact resolved input per item. It costs nothing and it answers most of these in seconds.
Literal braces in the output
If the resolved input still contains {product_name}, the placeholder did not match a column. An unknown placeholder is deliberately left as written rather than blanked — so:
- Check the spelling and case against the batch card's column names.
- Check you are using the right form:
{item.<column>}(or the short{<column>}),{previous.text},{step.<label>.text}. - If you renamed a step, update any
{step.<old label>.text}pointing at it.
See referring to data.
Empty fields on a triggered run
The classic: a payload mapping pointing at a column your flow does not have. It produces blanks with no error, because that is the same "leave unknown things alone" rule. Fire one event, then check Recent deliveries and the run's Data tab.
Empty cells from a spreadsheet
- A header renamed by hand no longer matches the flow's column — download a fresh template and compare.
- An image URL that is not publicly reachable, needs a login, or is over 12 MB is refused; that item has no image.
The model ignored the image
- Only some models can see. A vision step must run on a model that accepts images — Claude (Anthropic) is the reliable path for scoring or describing an image, which is why the Ad + QA loop insists on a Claude pool for its reviewer.
- Check whether the step should use the item's original images or the previous step's output — that is a per-card option, and getting it backwards is common in multi-step image flows.
The text is fine but wrapped in code fences
Some models wrap output in ~~~ fences however firmly you ask them not to. There is an opt-in "strip code fences" setting on a step for exactly this; the Catalog enricher template has it on.
It runs, it just is not good
- Try one item and iterate on the wording there. Fifty items of a bad prompt is fifty times the price of learning the same thing.
- Put the non-negotiables in a Brand guide / notes card (or the workspace brand guide) so every step obeys the same rules instead of each prompt re-stating them.
- For images, be concrete about surface, light and props — the template prompt suggestions are written to show the level of detail that works.
- Remember free test mode proves wiring, not quality: placeholder output tells you nothing about the prompt.