Transcripts are great for finding how someone explains an idea. They’re less useful when a summary strips away who said it, when, and with what uncertainty. We’ll use public history lectures as an example, with Codex handling the local notebook and Kabo providing available transcript evidence.
Kabo, an AI agent for creator research, retrieves available transcript text. The notebook preserves attribution and uncertainty before any historical claim enters narration.
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.
This example uses Codex with authorized local file access. Configure [mcp_servers.kabo] with url = "https://kabo.sh/mcp-pub" if no Kabo connection exists, then complete MCP sign-in. In the CLI, codex mcp login kabo starts that flow. The transcript notebook is a client-written file, not a file Kabo publishes.
Prepare the source log and verification questions
[ ] video-links.txt — Public sources and the question each might help answer. [ ] research-question.md — Explanation choices, historical claims or terminology to compare. [ ] source-log.md — Language, returned timing, retrieval date and status. [ ] verification-questions.md — Claims needing independent primary evidence. [ ] notebook/ — Attributed paraphrases, source notes and editorial decisions.
Inspect transcript readiness and the current input schema
A history-video producer uses Codex as a file-enabled MCP client to compare how public lectures explain the introduction of a printing technique. They need attributed notes, not a reconstructed transcript or an automatic fact check.
Ask the client which transcript tools are ready and what inputs they require. Kabo exposes public_video_fetch_transcript where available. Do not infer that every public video has usable text or that every source language will be returned as expected. Keep missing text as a research gap.
Inspect transcript readiness and the current schema for public_video_fetch_transcript. For [public URLs], retrieve available text and retain source language and returned time ranges. If a request returns a job_id, preserve it. Don’t summarize a queued acknowledgment as if it were the transcript.
Complete the data retrieval before summarizing
If the tool returns text immediately, retain it with its source metadata. If it returns a job_id, use get_research_data_job and inspect the terminal state, then read the listed artifacts with read_research_data_artifacts. A queued acknowledgment is not a transcript. This path is distinct from get_analysis_run for saved video analyses.
Use “the speaker says” until we’ve checked the historical claim. That small attribution protects the research notebook from turning a plausible lecture into our own unsupported narration.
The YouTube research recipe adds candidate discovery when we still need to choose the relevant source videos.
Complete any data job through get_research_data_job, then read only its listed object_refs with read_research_data_artifacts. Write short attributed notes with source URL, language and supported timing. Separate speaker claims, independently verified facts and our interpretation. Compare explanations without merging disagreements. Mark uncertain names for audio review and create verification-questions.md.
Write notes that preserve attribution
Have Codex save a notebook with short paraphrases, source URLs, language and supported time ranges. Separate the speaker’s claim, a fact you have independently verified and your editorial interpretation. Preserve uncertain names or technical terms for checking against the audio rather than silently “correcting” them.
A notebook entry you can inspect later
For a printing-history video, a useful notebook entry keeps several different jobs apart:
- Source: Attach the actual video URL.
- Speaker’s claim: Add a concise paraphrase after retrieval.
- Time range: Use only returned or manually checked timing.
- Verification question: Does an independent primary source support the date?
- Visual question: What does the demonstration show that the transcript omits?
- Editorial action: Resolve those questions before scripting.
Compare explanations without flattening disagreement
Ask what the lectures agree on and which differences may reflect scope, terminology or an actual disputed claim. Use short excerpts only where exact wording matters. Do not assemble a replacement copy of a lecture or translate a paraphrase into an apparent direct quotation.
Leave the notebook with a verification plan
Create questions-to-check.md beside the research notes if local file access is authorized. Include visual demonstration questions that spoken text cannot answer. Historical claims need appropriate independent evidence before they become narration; transcript retrieval does not perform that verification for you.
The transcript’s technical name looks unreliable. Preserve the uncertain wording, link the source and add an audio/primary-source check instead of silently choosing a plausible term.
Write narration from the verified notebook
Use Google Docs to develop the script from the verified notes. Keep source references with the claim checklist and inspect the original video wherever the historical demonstration matters visually.
Use the source dossier for a missing explanation source. Use video analysis when the unresolved question concerns the demonstration rather than its spoken description.
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.