Quest Ladder
TeachFORGE.ai
Quest Ladder / The AI brief
The description, the AI brief and content-type preferences
The difference between a generated type you rewrite and one you publish.
A vague brief is answered vaguely, at speed.
The description does the work
On the AI lanes the Description field on the Details tab is the brief. Everything else — format, subject, age group, marks — is structure; this is the only place you say what the question is actually for.
A useful description says three things: what understanding is being tested, what a learner who has not got it typically does instead, and anything that must or must not appear.
One line produces one line's worth of quality. Two or three sentences of genuine intent is the single cheapest improvement available on these lanes, and it costs less time than re-running a bad generation once.
The advanced brief
Description brief → Advanced… opens a brief with four tabs — Sections, Dynamic parameters, Student paths and Active learning blocks — so you can steer each part separately rather than hoping one paragraph covers all of them.
Its section list covers the ten parts of a type: title and metadata, stem or question text, options or choices, correct answer, hints, solution steps, post-question learning, constraints and tone, examples, and internal notes.
Use it when one part of a type keeps coming back wrong. Writing hints must not state the formula in the hints section is far more effective than saying it in the main description and hoping.
Content-type preferences
What content types should each part use? tells the generator what to build things out of, in three groups: content block formats, hint styles and answer value types.
It matters most where the natural choice is wrong. A physics stem left to itself may arrive as plain text where it wanted LaTeX; asking for LaTeX in the preferences gets it first time rather than after an edit.
Reading the compiled prompt
Before any run you can see the whole compiled prompt, with its size as ≈ tokens. It is worth reading once, early on: it shows exactly how your description, your settings and your exemplar are being presented.
The panel warns when a prompt is large, because some chat tiers truncate long pastes — it suggests downloading the text and attaching it instead. If you are pasting into an external model and getting oddly partial results, that is usually why.
Where a run fails to parse, Paste JSON manually is the fallback, and the exact reply from the model is available to look at.