AGENTIC · YOUTUBE COURSE

How I produce a YouTube course with a team of AI agents

A course I produce on my own, with eight agent roles that share one set of written rules, and, by rule, three decisions per chapter that stay with me.

Drafted with Claude Code from the course's private repository (commit c447d87) and published in a release Edgardo authorised on 16 November 2026. How this note was written.

In brief

  1. Navegando el Universo Ágil · 2026 is a Spanish-language YouTube course on OKR, Scrum and DevOps, with English audio as well, that I produce on my own, as I stated on 16 November 2026, with a team of AI agents organised by role.
  2. The agents write scripts and produce voice, animation, subtitles and derived pieces under one set of written rules kept in a GitHub repository, and by rule since 24 October 2026 the direction, the finished chapter and each publication need separate approvals from me.
  3. On 2 November 2026 the course board listed 48 pieces, 12 of them published, and on 16 November 2026 the public playlist showed 5 chapters; nothing here measures audience results or time saved.
At a glance. Figures with a date carry their source; the course repository is private.
CourseNavegando el Universo Ágil · 2026, a series on OKR, Scrum and DevOps published by NotJustBI. Public playlist.
Inventory48 pieces: 8 chapters, 24 Shorts and 16 LinkedIn lessons. Source: the course board, a view generated from the inventory of pieces on 2 November 2026 (private repository).
Status on that board12 published, 23 scheduled, 8 approved but not yet scheduled, 4 planned and 1 rejected.
Public on 16 November 20265 chapters, C01 to C05, in the playlist. Source: public observation.
LanguagesSpanish audio (main) and English audio; subtitles in es, en-US, pt-BR and fr-CA (decision D09).
TeamA production coordinator, six pipeline agents and a method architect. The decisions are mine.
My statementsI produce the course on my own; C01 to C05 were produced with ChatGPT and Google Drive before the repository existed; the narration uses a clone of my voice. Source: my statement of 16 November 2026.
ToolsChatGPT and Codex, Claude Code and Antigravity; Higgsfield (Seed Audio, ElevenLabs, Higgsedit); Metricool; Google Drive; YouTube Studio; GitHub Issues and Actions.
Versioned methodThe repository's pull request list, observed on 16 November 2026, showed 52 pull requests between 22 and 30 October 2026 (51 merged), including framework versions v0.1 to v0.42.
AI at playbackNone. These are videos: AI is used to produce them, and the synthetic voice is declared on YouTube.

What is the course?

It is a series in Spanish that explains OKR, Scrum and DevOps and how they connect, published on YouTube under NotJustBI, the brand I share with Grecia Pérez. I am its author and its presenter. The playlist gathers the chapters as they go public.

The whole series follows one fictional case: a mobile app that helps people organise their personal activities. Its figures are invented for teaching, and the rules require every piece to label them as fictional. The concepts come from official sources: the Scrum Guide for Scrum, Google re:Work for OKR, and Microsoft Learn and DORA for DevOps.

A file in the repository, the course canon, fixes that case and its terminology, so each new piece continues the series instead of reinventing it.

From ChatGPT and Google Drive to a repository

As I stated on 16 November 2026, I produced the first five chapters, C01 to C05, with ChatGPT and Google Drive, with no repository; the repository and the agents organised by role came later.

On 23 October 2026 their texts were brought into the repository: 394 files (scripts, captions, metadata, subtitles and production records) copied from Drive without changes. Codex did the copying and recorded each file's SHA-256 hash; all 394 matched the bytes downloaded from Drive, and nothing was published as part of it.

Since then, text lives in GitHub and heavy binaries such as voice, video and images stay in Google Drive, referenced by ID from the inventory of pieces.

The team: a coordinator, six agents and an architect

Each role is a Markdown file in the repository that states its scope and its limits:

  • Production coordinator. Entry point for every piece: sets scope and version, freezes the script, hands out the work, integrates it and keeps the issues and the inventory.
  • Editorial editor. Writes the Spanish script and, once I accept it, translations, metadata and student materials. It does not approve its own script.
  • Visual editorial editor. Builds the storyboard from the narration.
  • Sound and speech editor. Generates and edits the voice and closes the Spanish master clock.
  • Post-producer. Animates the video, times the subtitles and runs the audiovisual QA.
  • Publisher. Uploads or schedules, only with my publication approval.
  • Community manager. After publication: read-only analytics, comments and closing the status.
  • Method architect. Maintains the shared rules, skills, roles and workflows through pull requests; produces no pieces.
Figure 1. The chapter pipeline defined on 24 October 2026 (D26).
  1. OWNERAuthorises the piece and its version.
  2. COORDINATOR AND EDITORIALThey read the rules and the canon. The editorial agent writes the Spanish script, a review PDF and subtitles without timings; once I accept it, the coordinator freezes it by hash.
  3. VISUALBuilds the storyboard from the narration, screen by screen, with estimated timings.
  4. VOICE AND POSTProduce a real slice of the first lesson: final Spanish voice, animation, cover and storytelling.
  5. OWNER · GATE AApproves the direction on that slice. It does not approve the chapter.
  6. VOICE AND POSTComplete the Spanish voice, which becomes the master clock, and extend the edit. English audio is added to the locked video without stretching it, and the WebVTT tracks are timed.
  7. COORDINATOR AND POSTIntegrate covers, captions, three student PDFs and metadata, and run the QA.
  8. OWNER · FINAL APPROVALApproves the complete chapter.
  9. OWNER · GATE BAuthorises publishing or scheduling for that piece, version, platform and date. The publisher agent carries it out.

At commit c447d87 (2 November 2026) no chapter had completed it; C08 is the first, with only its Spanish script accepted.

A status belongs to a piece and a version, never to a whole chapter. The excerpts quoted in this note are checked on every build against verbatim copies taken from the private repository at commit c447d87, not against the live repository; the copied files are listed in /evidence/agentic.json.

Nunca convertir una fecha transcurrida en "publicado", ni un archivo generado en "aprobado".

notjustbi-courses · agentic/standards/status-taxonomy.md

Never turn a date that has passed into “published”, or a generated file into “approved”.

One set of rules for several tools

The method lives in one folder, /agentic: standards, skills, roles and workflows. The root AGENTS.md, which Codex and Antigravity read natively, only points there, and so do the .github/copilot-instructions.md file and the .claude/ and .agents/ folders. Claude Code reaches the same rules through a short CLAUDE.md.

Each role declares a default platform in its front matter. Two roles default to the platform the files label “Claude” (the method architect and the sound and speech editor); the other six default to the one labelled “GPT”, which in this repository covers ChatGPT and Codex. The voice role, for example:

plataforma_preferida: Claude
plataforma_alterna: GPT
Default and alternate platform of the sound and speech editor, as declared in its role file. notjustbi-courses · agentic/agents/sound-speech-editor.md

Work moves from one platform to another through GitHub issues. Labels say who executes the current step (exec:*) and where the turn stands (estado:*); running a step on a platform other than the default needs my OK. At commit c447d87 (2 November 2026) the hand-off was manual: the listeners that would pick up a step automatically were still pending infrastructure.

What the files record: Claude Code produced Short S17 on 26 and 27 October 2026. Codex copied C01 to C05 into the repository, prepared the C07 package and, by rule, reviews every pull request. For Antigravity the files record one use, on 22 October 2026: a retroactive storyboard of C06. GitHub Copilot is prepared, with no recorded use at commit c447d87 (2 November 2026).

What the agents produced

Each result names its piece and date. Approvals are mine and are recorded apart from the agent's delivery.

A script that can be reproduced (C08 v03, 29 October 2026)

The ficha-editorial skill turns one structured source into the Spanish script, a branded review PDF, untimed subtitles and a reading for the voice. Its regression test rebuilds the accepted PDF of the C08 v03 script (70,039 bytes, 11 pages) from inputs recorded in Git: on Windows it must match byte for byte; on Linux, where PDF compression differs, every object and decoded stream must match.

On 29 October 2026 two runs in different folders produced the same PDF. In commit c447d87 (2 November 2026) the skill has 40 test methods, which a GitHub Actions workflow runs on Windows and Linux for every pull request that touches the editorial skills.

My voice in two languages

The narration uses a clone of my voice (my statement of 16 November 2026), generated with Higgsfield; since rule D05 the Spanish uses Seed Audio, and the English uses ElevenLabs. By rule, Spanish is the master clock, English is adapted afterwards, and neither is sped up, slowed down or stretched to fit (D06 and D07); the synthetic voice is declared in YouTube Studio (D13).

Animation written as code (C07, 23 October 2026)

Since D04, lessons are animated with Higgsedit, Higgsfield's procedural engine, which its standard names the preferred one: each scene is code (edit.jsx) tied to the Spanish clock, and the subtitle area is reserved. In the Spanish master of C07 all 17,725 frames were measured and none invades that area. I approved that video on 23 October 2026.

English by reuse (C07, 23 October 2026)

The English track of C07 reuses the existing English voice: 1,305 of its 1,459 original words (89.44%) were kept and ten sentences were replaced by 8 new takes. It is an extra audio track on the same approved Spanish video. I approved it on 23 October 2026.

Four subtitle languages and three student PDFs (C07, 23 October 2026)

The C07 v02 package, approved on 23 October 2026, carries WebVTT subtitles in es, en-US, pt-BR and fr-CA and three student PDFs (a study guide, a review with exercises and templates); the PDFs were reused, not regenerated. C05 v04 and C06 v02 were produced before the repository; the repository keeps copies of their subtitles in es, en-US, pt-BR and fr-CA and of their three student PDFs.

Derived pieces with no new voice, illustrations or paid generation (C07, 23 October 2026)

From the approved C07 chapter the agents cut 3 Shorts and 4 LinkedIn lessons, with 28 WebVTT tracks (32 counting the chapter's own), from approved voice and scenes reframed for the new formats, not centre-cropped. I approved the package on 23 October 2026. I chose to reuse approved voice and scenes; on 26 October 2026 that became rule D42: derived pieces generate nothing paid.

A vertical Short rendered locally (S17, 27 October 2026)

S17, derived from C06, was the course's first Short whose vertical body was composed for that format, as an HTML animation rendered locally, instead of reframing the 16:9 master. It reuses the voice of C06 in both languages and has subtitles in four languages. I approved it on 27 October 2026.

Publishing by two routes (C05 v04 and S19 to S21, October 2026)

On 24 October 2026 the publishing agent uploaded and scheduled C05 v04, a chapter produced before the repository, in YouTube Studio through the browser; its receipt records the premiere and that the subtitles, the English audio, the PDFs and the AI declaration persisted, with the quiz left explicitly pending. Shorts S19 to S21 were scheduled through the Metricool API after my approval on 26 October 2026, a route designed for cloud sessions. At commit c447d87 (2 November 2026) its end-to-end validation was still in progress, and the standard says so.

A board that rebuilds itself (2 November 2026)

A GitHub Actions workflow regenerates the course board, TABLERO.md, from the inventory and the issues whenever either changes. The commit this note is based on, c447d87, is one of those regenerations, made on 2 November 2026.

What I decide

By rule, the agents work inside a scope I authorise. In C07 the coordinator extended the package to Shorts and LinkedIn before I asked; I approved it afterwards, and the approval record states that from then on the scope is set per piece beforehand. These decisions stay with me:

  • Three approvals per chapter, each independent, by rule since 24 October 2026 (D26). Gate A approves the direction on a real slice of the first lesson; the final approval covers the complete chapter; Gate B authorises publishing a given piece, version, platform and date. Accepting the Spanish script beforehand approves nothing else.
  • Any spending. Paid generation needs my explicit authorisation for that piece and version (D42); if something is missing, the agent flags it instead of generating it.
  • A change of platform for a step. The coordinator may propose it, but does not apply it without my OK.
  • The merge. By rule since 24 October 2026 (D25), every pull request first gets an automatic code review from Codex; the agent resolves its findings, escalates those that contradict my decisions, and only then asks me to approve and merge.

Three of the rules every agent reads first, in the original Spanish:

- **No publicar ni programar** nada sin autorización explícita para esa pieza+versión.
- **No generar** voz/video/activos por rutina ni en paralelo sin control de concurrencia y gasto.
- Respetar la **taxonomía de estados**: un archivo generado no es "aprobado"; una fecha pasada no es "publicado".
The first three non-negotiable rules in the root AGENTS.md: publish or schedule nothing without explicit authorisation for that piece and version; generate no voice, video or assets by routine or in parallel without control of concurrency and spending; respect the status taxonomy. notjustbi-courses · AGENTS.md

What changed along the way

  • On 24 October 2026 I rejected C08 v01 because it invented an example disconnected from the course, and since then each course keeps a continuity canon that the editorial agent reads before writing (D27).
  • On 26 October 2026 tracking moved to GitHub issues only, and the Projects board left the procedure (D37).
  • On 26 October 2026 the voice started to be stored in lossless segments of up to 10 MB, because the Drive connector in cloud sessions does not download larger files (D39).

What this does not prove

At commit c447d87 (2 November 2026):

  • No audience results. No view, subscriber or engagement figures are given or claimed.
  • No productivity claim. I have not measured time saved.
  • Private sources. The facts come from a private repository and from my own statements; only the playlist is public.
  • Manual hand-off. The listeners that would pass a step between platforms automatically were pending.
  • Metricool publishing was still being validated.
  • Not everything approved is public. C06 and C07 were approved but unpublished. The board still listed C08 v01 as rejected; the editorial skill's records show that the Spanish script of a later version, v03, was accepted, and nothing else.
  • Not client or employer work. This is a personal project.

What comes next

The next steps are already recorded: take the new version of C08 through the same approvals, publish the approved pieces when I authorise them, and close the Metricool validation. If you want to talk about this way of working, get in touch.

How this note was written

This note was drafted with Claude Code from the course's private repository at commit c447d87: rules, role files, inventory, delivery and approval records, and workflows. Facts that are not in the repository are marked, in the sentence or in the table at the top, as my statements of 16 November 2026, or as observations made that day of the public playlist and of the repository's pull request list. Each statement was traced to its source before publishing. No model or product version is named, by this site's editorial rule.

Choose optional measurement

Both purposes are off until you choose them. You can change your choice at any time.

Site content, CV downloads, language and theme controls work without optional measurement.