Evaluate a produced document¶
Once a document has been generated, how do you know whether it is any good? You check it against an evaluation rubric — a list of criteria written in advance, kept in the catalog and attached to the document.
This is the catalog's third kind of entry, alongside instructional routines and formatters:
| Entry kind | What it describes | Applies to |
|---|---|---|
| Routine | The teaching structure of a session | A lesson |
| Formatter | How the document looks | A document |
| Evaluation rubric | The criteria the result is judged by | A document |
Two shapes of rubric¶
A rubric carries a scale, and the scale decides what an evaluation produces. Two shapes recur:
| Shape | Scale | What it decides |
|---|---|---|
| Scored rubric | A numeric scale (for example 0 to 4) | A score, the weighted average of its sections |
| Approval checklist | Yes / No | A go / no-go. A single "No" blocks — there is no average |
The two do not replace one another: one measures how good the document is, the other says whether it may go to print. The same document can score well and still be blocked by a single "No" — which is exactly why both are often attached to it.
The scale, the sections, their weights and the criteria all come from the rubric as written in the catalog. The tool knows no rubric in advance: your programme files its own there (often taken from an official annexe), and that text is what counts.
A rubric is better off partial
A rubric need not reproduce in full the source document it is drawn from. Whatever a formatter already guarantees at generation time — margins, font size, pagination, illustration style — need not be re-checked afterwards: you do not want to produce a document in 9 pt and then tick "No". Keep the criteria that require reading the produced document.
Attach a rubric to a document¶
A rubric in the catalog does not apply itself — you attach it to the document it judges.
"Apply the approval rubric to this document."
As with a routine or a formatter, attaching creates an independent copy under the document. Later edits to the catalog's rubric do not reach documents already served.
One document may carry several rubrics — and normally does: one for the score, one for approval. The evaluation reports every one of them.
A rubric attached twice counts twice
Nothing stops you attaching the same rubric to a document twice — the evaluation would then report it in duplicate. If in doubt, ask "which rubrics are attached to this document?" before attaching.
Run the evaluation¶
"Evaluate this document."
The tool gathers the rubrics attached to the document plus the produced file, and then Claude reads the document and assigns the scores — the tool never judges by itself. You get, criterion by criterion: the score (or the Yes / No), and the reasoning behind it.
Ask for locations
A score without evidence is worthless. A well-written accuracy criterion asks for each error to be cited with where it appears — and a score or a "No" that arrives with no passage cited is worth asking about again.
What an evaluation cannot do¶
Some criteria assume a field test: how many pupils understand the instructions, how long an exercise takes to solve. These cannot be inferred from reading a document.
The rule there is to say so rather than invent a score. A fabricated score on an unobservable criterion is worse than a blank: it creates false assurance.
Likewise, a criterion may only make sense across the whole document: anything about balance or spread — between exercise types, between contexts, between characters — means nothing page by page. Those are judged after a complete read.