Skip to main content
When something goes wrong

When something goes wrong

Most of what looks like a failure is one of a small number of things, and nearly all of them are recoverable without starting over.

Common symptoms

What you seeWhat to do
Nothing happens for a whileThe model is thinking before it writes. On the deeper engines this is normal and can last a while. The clock is the tell — it means work is happening.
The answer stops mid-sentenceThe engine stopped early. What it wrote is kept — ask it to continue, or ask again. You do not lose the thread.
The work is marked as truncatedThe build ran past what one turn allows and was cut. Vresk keeps going in the background by itself and says so under the reply — or ask for something smaller.
A widget shows nothingSome widgets need to be run before they draw. Look for the Run control in the panel.
A deck looks wrong after exportCheck the slide in the panel first. If it is right there and wrong in the file, ask for that slide to be simplified — complex custom layouts are where fidelity is hardest.
An edit rewrote more than you askedStep back a version and restore. Then ask again, more narrowly — one change per message.
Your limits says it couldn’t read your allowanceThat means exactly what it says: reading your allowance failed on our side. It says nothing about your account or how much you have used. Reload in a moment; the figure comes back when the read does.
Search didn't runThe small card above the answer says exactly why: web search is unavailable, today's searches are used up, or it isn't set up here. In every one of those cases the answer was not checked against the web, and the card says that too.

When something ambitious runs long

A single turn has a time limit, and it is a real platform limit rather than a policy we invented. The interesting asks — a full deck, a substantial build — can exceed it.

When that happens Vresk continues the work in the background on its own and tells you so, right under the reply — a card with a running clock and a Stop button. You can close the tab; the finished result lands back in the conversation, and you are told the moment it is ready. This needs a signed-in account: a background run is real work on our side, so guests get the honest cut instead. The full story of background runs — allowances included — lives with limits.

Or make the ask smaller

“Draft the outline first” then “now build it” is often faster end to end than one enormous request, and you get to correct the structure before anything is designed around it.

When the output is disappointing

Before rewriting your prompt, try a different engine. These models have genuinely different characters, and a result that feels thin on one is often solid on another — particularly for design-heavy work like decks.

If you are choosing blind, run the ask across a few engines side by side once and use what you learn from then on. More on choosing an engine.

Tell it what missed

“This is too generic — it never mentions our actual constraint” gets a better second attempt than “try again”. The model cannot see your reaction, only your words.

When the answer is wrong

Different problem from a disappointing answer: a confident answer with a wrong fact in it. Models state wrong things in exactly the same tone as right ones — that is the nature of the technology, and the product's job is to give you the tools to catch it. These are those tools.

  • Check the citation chips. When an answer drew on the web, its sources appear as chips, and only the chips are real links — Vresk builds them from the actual search results, and a web address the model merely types in its prose deliberately does not link. A claim with no chip behind it is the model speaking from memory.
  • Ask it to check. “Verify that against the web” turns the claim into a search with sources you can inspect yourself. How web search works.
  • Get a second opinion. The same question to a different engine is the cheapest independent check there is — run them side by side.

When a message is declined

A safety check of Vresk's own wraps every model turn — it reads what you asked and what the model wrote back, whichever engine did the writing. When it declines a turn you get a plain refusal that says so, never a fake answer or a silent failure; in a side-by-side compare, the message says explicitly that the decision was Vresk's and not the engine's.

If the safety check itself cannot run, Vresk declines the turn rather than skipping the check. The message says exactly that: nothing about what you wrote was judged, a retry has already been made for you, and these outages usually clear within a couple of minutes.

No in-app appeal today

There is no way to contest a declined message inside the product yet. If you believe a decline was wrong, we do want to hear it — tell us here.