Skip to main content
This page follows one exam through all five products. It is written as a sequence because that is how a school year works, not because the products push data to each other. Your platform makes every call, and it can stop after any step.

​
Step 1: Grade the exam

1

Create the exam and upload the question paper

Give the exam a name, grade, subject, type, set, and date, then upload the paper as a PDF or as images, and mark it final. CrazyGoldFish reads the sections, questions, and marks out of it.
2

Upload the model answers

Once the questions are extracted, upload the model answer sheet. Where the paper gives marks step by step, those steps are picked up too.
3

Upload each student's answer sheet

Create one answer sheet per student and upload the scans. Each sheet is graded on its own, question by question.
4

Read the marks and check them

You get marks for every answer with the feedback behind them: what was wrong, the marks taken off, and the reasoning. The API accepts a change to a mark, an edit to an answer or a model answer, and an approval flag, from your platform. It does not check who is asking, so your platform decides which of its users may do those things.
5

Publish

Publishing is deliberate and happens in two stages. Until you publish provisional results, they cannot be read from the provisional or final endpoints. Publishing again, as final, opens the final result. Note that a third endpoint returns the same marks and feedback for a marked answer sheet with no publishing check, so publishing is a signal your screens follow rather than a lock. Marks also stay editable after publishing.
6

Handle student queries

A student can question a piece of feedback. The query goes back to the AI, which either keeps the mark or revises it and explains why. The query is then marked resolved.
Everything so far happens inside Exam Evaluation. See the Evaluation Layer for the detail.

​
Step 2: Ask what to do about it

Now your platform makes a second, separate call. Send the answer sheet ID for one student, or the exam ID for the class. Plans are written only from answers that lost marks, and only from exams CrazyGoldFish graded. The AI sees the question number and the written marking feedback, and nothing else: not the question wording, not the model answer, not the marks. See evaluation to action plans.

​
Step 3: Make the material

A plan says what needs work in sentences. Your platform, or the teacher reading the plan, turns that into a topic and asks for material on it.

A practice worksheet

Up to 10 questions with an answer key, each tagged with a learning objective, a difficulty, and a thinking skill. Explanations are optional. The teacher reviews the topics and then the question plan before anything is written.

A reteaching lesson

One lesson of 30 to 90 minutes, in six blocks from introduction to assessment, with what to say, what to use, and what to ask. The teacher reviews the objectives and then the lesson structure.
See action plans to content for what carries over and what does not.

​
Step 4: Review the teaching

After the lesson, your platform can send a recording of it to ClassTrack. The request needs the rubric ID you want the teaching judged against, plus your own teacher ID, the board, grade, and subject IDs, and the date of the lesson. You get a written judgment per rubric point, observations with quotes from the lesson, and suggestions for the teacher. The scores are text, not numbers. See where observation fits.

​
What carries over between steps?

Only three things, and your platform carries all of them.
Nothing in this sequence starts on its own. There is no loop that reruns, no product that watches another, and no result that changes a rubric, a lesson plan, or an exam after the fact. If a step should happen, your platform makes the call.

​
What if you only want one part?

Every step above works alone.
  • Grade exams and stop. Publishing and student queries work without action plans.
  • Create worksheets and lesson plans without ever evaluating anything. The builders need a board, grade, and subject, one source such as a topic, and, for a lesson plan, the lesson length.
  • Run classroom observation on its own. It needs a recording, a rubric, your own teacher ID, the board, grade, and subject IDs, and the date of the lesson. All of those are required.
  • Action plans are the one product with a prerequisite: the exam has to have been graded by CrazyGoldFish first.

​
Frequently asked questions

Each step is a background job, and the time depends on the size of the work: the number of pages in a paper, the number of answer sheets, the length of a recording. Treat every call as a job you collect later rather than wait on.
Yes. You can read the marks as soon as evaluation finishes, without publishing. Publishing opens the provisional and final result endpoints, but it is not a lock: another endpoint returns the same marks with no publishing check, and the API cannot tell a student’s request from a teacher’s. Whether either step happens is up to your platform.
For grading, yes. Assignment Evaluation grades everyday work against a rubric or model answers. But action plans read exam results only, so an assignment cannot start one.
No. A correction is stored against that answer, and a student query triggers an AI re-check of that one answer. Nothing changes your rubrics or model answers for a future exam unless you change them, and nothing the AI does on one exam carries into the next.
No. Every result is returned to your server. CrazyGoldFish sends no email, SMS, or messages to students, teachers, or parents.
The job comes back as failed, and the next step simply does not happen, because you are the one who starts it. Fix the input and start that step again.

​
Next steps

The five products

What each one takes and gives back.

API quickstart

Get a token and make your first call.

Async jobs and webhooks

How background jobs, polling, and webhooks work.

All APIs

Every endpoint, grouped by product.