Draft Assignment Feedback from a Rubric with AI at the Cursor

The short answer: speak your marking observations criterion by criterion inside a single spoken AI request, and ask for them to be arranged as structured feedback at the cursor in a compatible Windows text field, such as a comment box or document. Your observations are the only source material. The request tells the AI to flag gaps and not to invent evidence, but the output is a draft that needs your review. You decide what the student sees and handle the grade separately.

This article walks through that workflow for a university lecturer. It covers setup, how to prepare inputs, an illustrative before-and-after example, a reusable template, a review checklist, and a method for measuring time to reviewed output without guessing.

Rubric-based feedback has a predictable shape. For each criterion, a student needs to know what you saw in their work, what would improve it, and what to do next. The thinking happens while you read. The slower part is often turning scattered notes into consistent, readable comments across a stack of submissions.

This workflow produces a first draft of that written feedback from four fields you supply:

• Criterion: the rubric heading, as written in your course materials.
• Observed evidence: what you saw in the submission, such as a paraphrased passage, a paragraph location, or a missing element.
• Improvement: the change that would strengthen that criterion.
• Next action: a concrete step the student can take in the next draft or assignment.

The output is draft text placed at your cursor. It is not a grade, an independent AI assessment of the work, or a finished comment. It carries the evaluative observations you speak, such as what is clear or less clear, and the AI should add none of its own. The AI structures what you say, and you remain the assessor.

SpeakLingo is built around Syping, the combination of speaking and typing: you speak, and text appears where you would otherwise type. The Syping page explains the idea, and the features page gives the wider product context.

Three capabilities matter for this task:

1. Online transcription in 99+ languages. You can dictate your observations in a supported language that suits you and your teaching context.
2. AI contextual formatting. Spoken notes can be shaped into a clearer structure, such as headed criteria with consistent sub-points.
3. Spoken AI requests that produce text at a compatible focused Windows text field. You say the task instruction, and the result is placed at the cursor rather than in a separate AI window.

That last capability addresses a familiar chore. A common pattern is to open a separate AI window, paste notes in, copy the answer out, and paste it into the marking environment. Where the cursor workflow is supported, you can reduce those transfers and some of the manual structuring. Compatibility varies by application, so test your marking environment with a harmless dummy comment before relying on it.

SpeakLingo is accepting waitlist signups, so treat this article as a description of the product's intended workflow, not a report of how it behaves on your machine today. The method of preparing inputs, requesting structure and reviewing output can be adapted to whichever tools your institution permits, but how any particular tool behaves will differ and needs testing.

A few minutes of setup keeps a draft from reaching a student before you have read it.

1. Check the service and your permissions. Online AI requests need an internet connection, an enabled service, your consent and access. Check what your institution permits for student work before you begin.
2. Turn off automatic Enter or send behavior. SpeakLingo has an optional Enter-after-delivery behavior that can submit text in some apps. In a feedback box or messaging field, that could release a comment before you have reviewed it. Turn it off for marking, so delivered text stays in the field.
3. Test with a dummy comment. Place the cursor in the field you will actually use, speak a short test request with invented content, and confirm the text lands where expected.
4. Decide about screen context. Screen context is optional and opt-in. If the marking window shows student names or identifiers, consider leaving it off. If enabled, visible content may be sent remotely along with your audio and text.
5. Keep the rubric and submission open separately. The draft should come from your observations, not from whatever happens to be on screen.
6. Review the draft before sending. Read it in full, edit it, and press save, upload or release yourself.

How the AI request gets its material. Use the supported AI Task workflow for this step, and speak the instruction and the facts it needs together in the same request. That means the rubric criteria, your observations, the improvements and the next actions. Dictation alone does not turn on AI. Earlier dictations, text already in the comment box, and whatever is visible on your screen are not automatically available to the AI. Screen context is a separate opt-in and may be sent remotely, so put everything the draft needs into the request yourself.

The draft can only reflect the evidence you supply. The example prompt for this workflow sets the rule:

Using only these verified inputs, draft assignment feedback from a rubric. Inputs: [Criterion, observed evidence, improvement, next action]. Use a clear professional structure. Flag missing facts; do not invent them.

"Verified" here means you have checked the observation against the submission. A few habits help:

• Read first, then speak. Form your view of the work before you dictate, so each statement comes from something you saw.
• Anchor evidence to the work. "Second paragraph of section two asserts a causal link without a source" gives the draft more to work with than "evidence is weak."
• Say the criterion name as written in your rubric. This keeps headings consistent and makes skipped criteria easier to spot.
• Leave out sensitive details. Do not speak student names, ID numbers, health information, or personal circumstances unless your institution's policy clearly allows it. Use a neutral label for the submission if you need one.
• Keep the grade out of the request. Ask for feedback structure only. Decide marks separately against your rubric and any moderation process you follow.
• Mark one submission per request. A single request containing one submission's inputs keeps evidence from different students apart.

This example is fictional and uses non-sensitive content. The output shown is illustrative, not observed from a real run. The assignment is a short first-year essay on urban green spaces. The rubric has four criteria: Argument, Use of evidence, Structure, and Referencing. The lecturer has read the submission and prepared observations.

Before: the lecturer's spoken inputs

The lecturer speaks these observations as the inputs part of one request:

Argument. Clear position in the intro, that green space should be prioritised in dense areas. By the conclusion the position softens and it's less clear. Should keep the claim consistent. Next time, write the thesis in one sentence and check the conclusion against it. Use of evidence. Two examples in section two, both described but neither linked back to the claim. Should say why each example supports the argument. Next, add one sentence after each example explaining the link. Structure. Sections follow a sensible order, headings help. Transition into section three is abrupt. Add a linking sentence. Referencing. Haven't checked this one yet.

This is useful raw material but awkward to send as it stands. The criteria run together, the tone is clipped, and the last criterion is incomplete.

The spoken AI request

With the cursor in the feedback box, the lecturer adds the instruction around those inputs:

Using only these verified inputs, draft assignment feedback from a rubric. Inputs: [the observations above]. Use a clear professional structure with one section per criterion: observed evidence, improvement, next action. Flag missing facts; do not invent them.

After: structured draft at the cursor (illustrative)

A draft along these lines might appear. Actual output will vary and always needs checking:

Argument Observed: Your introduction states a clear position that green space should be prioritised in dense areas. By the conclusion, that position is less clear. Improvement: Keep the claim consistent from introduction to conclusion. Next action: Write your thesis in a single sentence, then check your conclusion against it before submitting.

Use of evidence Observed: Section two includes two examples. Both are described, but neither is linked back to your claim. Improvement: Explain why each example supports your argument. Next action: Add one sentence after each example that states the link to your claim.

Structure Observed: Sections follow a sensible order, and the headings help the reader navigate. The transition into section three is abrupt. Improvement: Signal the shift in focus between sections. Next action: Add a linking sentence before section three begins.

Referencing Missing information: No observed evidence, improvement, or next action was provided for this criterion. Add these before using this feedback.

What the lecturer does next

The request asked for gaps to be flagged, and Referencing is flagged as incomplete. The lecturer checks the reference list against the submission, adds the observation in their own words, and edits any phrasing that does not sound like their voice. They confirm each detail matches the work and decide the mark separately. Only after that do they save or release the comment.

The draft contains no grade, no new AI-added verdict on the quality of the argument, and no invented comment about referencing. The evaluative remarks it does contain, such as the clear position, the less clear conclusion and the sensible section order, come from the lecturer's own spoken observations. If a grade, an AI-introduced judgement or an invented referencing comment had appeared, the lecturer would remove it. The reshaping of the lecturer's own words is the useful part, and every claim about the student's work still comes from the lecturer.

Keep this near your marking setup and adapt the labels to your rubric.

Spoken input pattern (repeat for each criterion):

Criterion: [rubric heading]. Evidence: [what I saw, with location]. Improvement: [what would strengthen it]. Next action: [one concrete step].

Request wrapper to speak around the inputs:

Using only these verified inputs, draft assignment feedback from a rubric. Inputs: [the criterion patterns above]. One section per criterion with observed evidence, improvement, and next action. Keep a supportive, professional tone. Flag any criterion with missing facts. Do not invent evidence, quotes, or marks. Do not suggest a grade. Do not add evaluative judgements beyond my observations.

Output skeleton to expect back:

[Criterion] Observed: … Improvement: … Next action: …

[Criterion with a gap] Missing information: …

Optional additions you can speak as needed include a word limit for each section, a preferred sentence length, or an instruction to address the student as "you." Keep additions short so they do not compete with your evidence. For other professional scenarios built on the same pattern, see the use cases page.

The table also works as a pre-dictation check. If you cannot supply a specific, verified observation for a criterion, that criterion is not ready to dictate, and the flag in the draft is the right outcome.

Use this after the draft appears and before you save or release anything.

• [ ] Every comment traces to something I observed in this submission.
• [ ] No quote, page reference, or example appears that I did not supply.
• [ ] All rubric criteria are present, and every flagged gap is resolved.
• [ ] Improvement and next-action statements are specific enough for the student to act on.
• [ ] Tone is respectful and consistent with how I give feedback in this course.
• [ ] No grade or mark band appears, and no evaluative judgement appears that I did not supply myself.
• [ ] The grade was decided separately against the rubric and any moderation process I follow.
• [ ] No student identifiers or sensitive personal details appear in the text.
• [ ] Wording is clear for a reader whose first language may differ from mine.
• [ ] Automatic Enter or send behavior was off, and I have read the whole draft.
• [ ] I have checked my institution's guidance on using AI tools with student work.
• [ ] I pressed save, upload, or release myself.

These are illustrative writing practices. Your department's assessment policies take priority.

Do not assume the workflow saves time; measure it on your own marking. The useful measure is time to reviewed output: the elapsed time from finishing your reading of a submission to holding feedback you have reviewed and are willing to release. Drafting is only one part of that.

A simple method:

1. Choose a comparable batch. Pick submissions of similar length and difficulty, or alternate between methods on the same assignment.
2. Define start and end. Start when you finish reading and have formed your observations. Stop when the feedback is reviewed, edited, and ready for you to release.
3. Include the whole path. Count speaking the request, any re-dictation, gap-filling, editing, and fix-ups for compatibility problems in the target field.
4. Record short notes. Write down when a draft needed heavy rewriting, when a field did not accept text, or when a criterion was flagged as missing.
5. Compare your own figures. Set your usual method against this one over several submissions, and judge whether the difference matters for your workload.

Quality belongs in the comparison too. If edits grow, or comments feel less specific than your usual standard, record that alongside the clock.

• Compatibility varies. The workflow depends on a compatible focused Windows text field. It will not suit every application or learning platform, so test first.
• AI requests need connectivity and access. They require an internet connection, an enabled service, your consent, and access. Transcription has its own mode, setup and access requirements, so an AI request being unavailable does not show that transcription is available. Offline English transcription is only a fallback if it has been configured in advance, and AI requests stay unavailable until you reconnect.
• Data may leave your device. Audio and text may be sent remotely, and so may optional screen context if you opt in. Consider what is acceptable for student work.
• Language scope. Transcription supports 99+ languages. That is not translation, and it does not guarantee AI output in every language, so read the result closely if you work in a language other than English.
• Judgement stays with you. The draft can misstate, over-smooth, or omit something. Assessing the work, deciding marks and releasing feedback are your responsibilities.