“I can’t do this” can mean “I can’t hear the difference,” “I don’t understand the instruction” or “I need more practice.” Those deserve different lessons. We’ll use language education to show how Claude Code can preserve that distinction, while Kabo retrieves the public comment evidence.
Kabo, an AI agent for creator research, collects public comments. Claude Code helps two contributors agree on what the labels mean before applying them across the sample.
Prepare the raw sample and teaching codebook
Add a short pilot-labels.md with each contributor’s initial decisions. The learning need and the comment type are different labels; record both instead of forcing one overloaded category.
[ ] video-links.txt — Relevant public lesson URLs. [ ] comment-rubric.md — Request, correction, experience and unclear, with definitions. [ ] lessons.csv — Existing exercises and explanations to avoid duplicating. [ ] teaching-notes.md — What our own audio demonstrations can prove. [ ] research/ — Raw comment sample, classification notes and source references. [ ] pilot-labels.md — Independent labels, reasons and disagreements for the same small comment set.
Connect Kabo, then check the project context
Open the content project, install Kabo for Claude Code and complete sign-in. For a direct server connection, follow the setup guide.
Calibrate two different labels, not one broad theme
For our language lessons, a comment can be a request and concern listening at the same time. One label describes the kind of comment; the other describes the learning need.
Retrieve a small public sample with Kabo, keep the original evidence, and have contributors label the same few rows. The goal is to discover definitions that lead to different teaching decisions, not to make everyone agree immediately.
Read comment-rubric.md and lessons.csv. Use Kabo to retrieve a small relevant public comment sample. For a few rows, preserve the source reference and label comment type separately from learning need. Include the contributor labels I provide and mark unreviewed rows. Show disagreements and their reasons before grouping the full sample.
Preserve the raw sample and its limits
Use ready Kabo public comment tools and save source video, comment reference, collection date, sort order and coverage. Retain missing replies as a limitation. Do not infer a commenter’s nationality or the whole audience’s language background from public text.
Resolve the exercise implied by a borderline case
Consider this authored paraphrase: a learner can hear the sound alone but loses the contrast in a sentence.
- Reviewer A: articulation exercise.
- Reviewer B: listening-in-context exercise.
- Evidence to inspect: does the original comment ask about producing the sound, hearing it, or both?
- Possible decision: request + listening in context, if the text supports that reading.
- Exercise implication: hear a contrast in a sentence before asking the learner to repeat it.
If the source remains ambiguous, keep the row unclear. A majority vote does not supply missing evidence.
Apply the settled distinctions and audit the boundary cases
Once the definitions are agreed, classify the collected sample using the two labels. Ask Claude Code to isolate rows near the boundary between hearing and producing a sound, or between a correction and a request. Review those rows before turning counts into teaching priorities.
The output should preserve the evidence behind a decision, so a second contributor can challenge the proposed exercise.
Use the agreed comment-type and learning-need definitions on the saved public sample. Keep raw text and source references unchanged. List ambiguous rows separately, count only within this sample and check lessons.csv for existing exercises. For each proposed new exercise, explain how the label supports that response and which audio examples we need to record.
Review one exercise brief against its source question
Before writing a batch of lessons, inspect one proposed response together. If the need is hearing a contrast in connected speech, a mouth-position demonstration alone will not answer it. Keep the source question, chosen exercise and required audio in the same brief.
For an existing-library response queue rather than team label calibration, use the Codex comment workflow.
These comments ask how to hear the contrast, not how to position the mouth. Keep them as a listening-exercise brief and remove the articulation demonstration.
Try the exercise against the learner’s question
Use Google Docs for the teaching review, keeping the rubric decision beside the exercise brief. Once the exercise is approved, the script workflow turns it into spoken and audio-demo beats. Update a durable definition only after the contributors agree it should apply again.
Worked examples are illustrative. Product screenshots show existing Kabo web results, not a run of the example in this guide. Use your own materials and the evidence returned in your task.