Guide for creating or revising a prompt that an agent uses to generate qualitative feedback on a single student submission. Use this skill when building feedback prompts for exercises, assignments, worksheets, or similar student artifacts where each submission is reviewed in an isolated agent session together with the exercise and didactical concept.
Write a prompt for an agent that reads one student's submission and produces one qualitative feedback report.
You are writing the prompt, not the feedback report itself.
The generated prompt must make isolated agent runs produce feedback that is:
The prompt must not assume shared memory across students. It must define the review lens, structure, tone, and prohibitions clearly enough that different agent sessions produce comparable reports.
Each student submission is reviewed in a separate agent session.
The reviewing agent receives:
That means the feedback prompt does not need to restate the full exercise. It should instead tell the reviewing agent how to use the provided materials consistently.
The prompt should make explicit:
A weak prompt leaves the reviewing lens implicit. A strong prompt makes it explicit.
The prompt must instruct the agent to act as a feedback writer, not a grader.
The agent's job is to describe:
The agent must not assign grades, points, scores, percentages, pass/fail labels, rubric levels, rankings, or other quantitative judgments.
A human teacher reads the report and determines the grade independently.
The generated prompt should explicitly instruct the reviewing agent to use the provided materials in this way:
The agent should use the exercise and didactical concept to interpret expectations, but should ground all claims about student performance in the student's submission.
Do not let the agent drift into generic pedagogy or teacher-facing didactical commentary inside the report.
Every prompt must define one primary validation lens for the exercise.
This quality bar is the main question through which the agent interprets the submission.
Good qualities of a quality bar:
Example:
Could a competent frontend developer build a working interactive prototype from this document alone, without needing extra clarification?
Use exactly one primary quality bar unless the exercise genuinely requires two tightly connected lenses. Do not create a vague bundle of half-related criteria.
All major sections of the report should align with this quality bar.
You can assume that the reviewing agent has access to the exercise and the didactical concept in the same session.
The feedback prompt therefore does not need to repeat the full exercise text.
It must still define all of the following clearly:
Do not leave key review criteria implicit just because the exercise and didactical concept are available.
The prompt must force the agent to ground every important claim in the submission itself.
Instruct the agent to:
Do not let the agent invent strengths to sound balanced.
Do not let the agent infer hidden understanding from polished wording alone.
Do not let the agent assume intent when the document is ambiguous. Describe the ambiguity instead.
The prompt must define how to handle incomplete, uneven, or misaligned submissions.
Instruct the agent to:
A report should still be useful when the submission is incomplete.
Define a fixed report structure in the prompt.
Every report must include these universal sections:
Document Overview
2 to 3 sentences. State what the student produced, how complete it appears, what major parts are present, and whether advanced aspects were attempted.
Strengths
Up to 5 bullet points. Each point must refer to something concrete in the submission. Do not invent strengths for balance.
Gaps and Weaknesses
Up to 5 bullet points. For each point, state what is missing, vague, inconsistent, or underdeveloped, why that matters in light of the quality bar, and what a stronger version would need to do. Point toward improvement. Do not rewrite the missing part for the student.
Consistency Check
A short paragraph. Check internal consistency, terminology, states, rules, transitions, and cross-references. If the document is largely consistent, say so briefly and specifically.
Actionable Next Steps
Up to 3 bullet points. Prioritized, concrete, and realistic. Answer the question: if the student had one more hour, what should they improve first?
Summary
A short closing paragraph that characterizes the current state of the submission through the quality bar.
Between Consistency Check and Actionable Next Steps, the prompt may add up to 3 domain-specific sections.
Use them only when the exercise has distinctive quality risks or artifacts that are not already covered well by the universal sections.
Good reasons to add a domain-specific section:
Bad reasons:
Prefer 1 or 2 strong domain-specific sections over 3 weak ones.
For each domain-specific section, define:
Embed these rules directly in the generated prompt so the reviewing agent applies them consistently.
Prefer:
The document defines the main endpoints clearly, but it never states what happens when the booking already exists.The terminology is mostly consistent, except for the switch between "member" and "customer" in the error cases.Avoid:
This is excellent work.The document is poor.Overall, it seems like the student perhaps understood the task in general.Do not use em dashes or en dashes to join clauses or add asides. Rewrite using a new sentence, comma, colon, or parentheses.
The generated prompt must explicitly prohibit the reviewing agent from doing the following:
When writing or revising the feedback prompt, adapt it to the exercise at hand.
That means:
Do not produce a generic feedback prompt that could apply equally well to almost any exercise.
good, weak, complete, or clear without telling the agent how to interpret them for this exercise.Before finalizing the prompt, check the following: