The 3 places a filter touches an answer
A filter can act on the input, on the generation, or on the finished reply. Input filtering produces a fast canned response. Generation filtering produces a hedged answer. Output filtering produces a reply that stops mid-thought or gets replaced. The mark each of the 3 leaves is different enough to read.
People picture a filter as a gate that either opens or closes. In practice there are 3 gates at different points, and the gate that acted determines what you should do next.
On this page
Before: the question is classified
Your text is examined before the model works on it. When this stage decides, the reply arrives quickly and reads as prepared, because it was. The distinguishing sign is speed: an answer that appears far faster than usual was not generated, it was retrieved.
During: the wording is steered
The system prompt and the tuning shape the reply while it is being written. Nothing is blocked and nothing is replaced; the sentences simply arrive softened. This is the stage most people mean by a filtered answer, and it is the hardest to see, because there is nothing missing to notice.
After: the reply is checked
A finished reply can be scanned before it reaches you. When this stage acts, you sometimes see the first part of an answer and then a stop, or a complete reply replaced by a short notice. Text that appears and then vanishes belongs to this stage and to no other.
Why the stage matters to you
Rephrasing helps against the first stage, because it works on surface features. Asking for structure helps against the second stage, because ordering is a formatting instruction. Neither helps against the third stage, and recognising the third stage saves you from trying 5 rewordings that could not have worked.
Reasonable questions
Do all services use all 3 stages?
Most use at least 2, and the mix differs. A service can drop the first stage entirely and keep a strict third one, which reads as openness until an answer disappears.
Is the system prompt a filter?
It behaves as one at the second stage. It sets tone and boundaries in ordinary language rather than as a rule list, which is why its effect is felt as style rather than as blocking.
Can I see which stage acted?
Not directly, and the marks are readable: unusual speed points at the first stage, softened wording at the second, and an interrupted or replaced reply at the third.
Watch how a reply here arrives, and at what speed.
Open unfiltered chat