A competitor study is several questions wearing one title: what gets published, what the videos teach and what viewers ask. They may need different tools. We’ll compare beginner guitar lessons to show how to compose the evidence, review the coding and choose an original demonstration.
Kabo, an AI agent for creator research, returns different kinds of public evidence. This recipe keeps channel, video and optional comment records linked to the right comparison claim.
Choose the client and confirm the connection
Connect https://kabo.sh/mcp-pub in a compatible client and complete Kabo sign-in. The MCP setup guide covers the connection; installation options are available for Codex and Claude Code.
The worked client is Claude Code, with an authorized content folder for the source index and comparison matrix. Add the direct remote server only if it is not already configured: claude mcp add --transport http kabo https://kabo.sh/mcp-pub. Use /mcp for authentication and connection status.
Prepare the account set and evidence links
[ ] channels.md — Real public account URLs and the common comparison scope. [ ] scope.md — Shared dates, included formats and collection limits. [ ] source-index — Actual channel/video identifiers or URLs and retrieval dates. [ ] comment-sample — Optional rows with the actual parent video reference. [ ] finding-map — Each proposed finding, its evidence type and linked source rows.
Keep three evidence sets separate before joining them
Our guitar-lesson comparison can use channel inventory, video evidence and optional comments. Each answers a different question. Ask the client to map the ready tools to those questions and record the parent source for every retrieved item.
The main job here is preserving those relationships. For the choice of an original series, the Codex comparison guide handles the editorial decision.
Read channels.md, scope.md and rubric.md. Inspect ready Kabo tools. Map channel inventory, video explanation and optional comment questions to the appropriate tools. Define the identifiers and source URLs our normalized tables will preserve. Keep coverage and retrieval date with each evidence set. Show the plan before collection.
Collect a common comparison frame
Retrieve the available uploads in the shared window, recording actual coverage and pagination limits. Keep Shorts separate from long-form lessons. Store source dates and format judgments. For returned metrics, retain observation age and any relevant account context instead of creating an all-time leaderboard.
Join by the actual source, not a similar topic
Create our own normalized index using real returned identifiers or source URLs. A comment about a first barre chord supports a question under its actual source video. It cannot explain the public performance of another lesson simply because both teach the same chord.
If a parent relationship is unavailable, leave that evidence unjoined. Do not match rows by a convenient title fragment.
Collect the agreed public evidence with Kabo. Preserve channel-to-video and video-to-comment relationships using returned identifiers or actual source URLs. In our normalized matrix, attach every content observation to its evidence record and every coverage claim to the collection scope. Keep unmatched evidence separate. Show a few linked rows for review before writing findings.
An illustrative join check
These are labels for an authored example, not tool-returned IDs:
- Channel A → Video A1: a lesson in the accepted publication window.
- Video A1 → Comment A1-C: a learner question retrieved under that video.
- Channel B → Video B1: another lesson with its own public metric snapshot.
- Valid note: the A1 comment illustrates a question in the A1 sample.
- Invalid note: that comment explains why B1 received more views.
The final table replaces these labels with actual source references and keeps sample limits visible.
Audit the evidence used by each finding
Ask the assistant to separate inventory findings, content-description findings and comment-question findings. Missing transcript or visual evidence limits what we can say about a demonstration. A small comment sample limits what we can say about learner needs. Neither should be silently promoted into a causal performance explanation.
For each proposed comparison finding, list its evidence type and linked source rows. Remove claims depending on an unmatched comment, uninspected visual sequence or incomplete inventory presented as complete. Keep the supported parts and state the missing relationship or evidence needed.
Deliver the comparison with its source map
Return the matrix, source index and a short note identifying unresolved joins and coverage gaps. Save files only through the client’s authorized folder access; a compatible chat client can return copyable tables. The production team can then choose a lesson without losing the evidence behind the comparison.
The comment came from a different lesson than the comparison row. Keep it as a separate audience question and remove the claim that it explains that video’s performance.
Review the matrix with its source map
Review the matrix in Google Sheets. Use the transcript recipe for spoken-method evidence and the Claude Code comparison guide when the team needs to calibrate its coding definitions.
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.