Skip to content

GPT-6 Astra + Higgsfield MCP Made This ENTIRE Video in One Chat

2026-09-05 · long tutorial / working session · Adil · recording, 11:06 · watch on YouTube

Topic: Prompting & AI assistants (also: Post-production & finishing, Pre-production & story, Video generation & shot direction, Audio, voice & music) · mode: Tutorial

In this tutorial of 5 September 2026 Adil shows how GPT-6 Astra, connected to Higgsfield MCP, produced the video itself, including his talking head generated from a short clip of him. GPT-6 Astra coordinates, Higgsfield generates the footage and editing apps finish it: a production brief, a script split into one-thought takes, a tested presenter, real screen captures, After Effects graphics and a DaVinci Resolve edit. He stresses that the creator judges each result, and that a realistic generated presenter should be disclosed to the audience.

What is said

The parts of the video and what is said in each (★ — how much it matters to the video).

00:00:00 — This entire video was made in one chat: the four input files and the goal

  • ★★★ Adil says this video was made with GPT-6 Astra and Higgsfield MCP, including his talking head, which was generated from a short clip of him. — 00:00:00 · GPT-6 Astra, Higgsfield MCP, AI avatar
  • ★★ Adil started from four files: a short video of himself, an After Effects project, a render of its motion graphics, and a finished YouTube video as a style reference. — 00:00:00 · After Effects, motion graphics, YouTube
  • ★★ Adil says a folder of generated clips is only the beginning: a finished video needs a clear story, the right pacing, useful demonstrations, consistent sound and an actual export. — 00:00:28
  • ★★★ In Adil's workflow GPT-6 Astra coordinates the production, Higgsfield generates the footage and the editing apps turn the pieces into the finished video. — 00:00:28 · GPT-6 Astra, Higgsfield MCP

00:00:58 — Connect GPT-6 Astra and Higgsfield MCP, then a small connection test

  • ★★ Adil explains that MCP (Model Context Protocol) gives an assistant a standard way to work with external tools, here Higgsfield generation and media management. — 00:00:59 · Higgsfield MCP
  • ★★ To connect Higgsfield to an assistant: in ChatGPT add the Higgsfield plugin and sign in; other compatible clients connect through the Higgsfield MCP server, following the client's official setup instructions. — 00:00:59 · Higgsfield MCP, ChatGPT
  • ★★ Before asking the assistant for a whole video, Adil advises a small connection test: have it inspect the generation tools and the shot's requirements, then generate one short sample. — 00:01:26
  • ★★★ Adil's production brief for GPT-6 Astra gives the topic, audience, a 10 to 12 minute running time and a horizontal YouTube format, and names the presenter reference, the existing graphics and the previous video. — 00:01:54 · GPT-6 Astra, YouTube

00:02:03 — Write the production brief, use the style reference, check the technical claims

  • ★★ Adil's reusable prompt structure for the brief: describe the finished result, name each input file and its role, specify the style and where files are saved, and ask for editable projects. — 00:02:17
  • ★★ Adil uses the earlier YouTube video only as a style reference: the new script still needs its own argument, carrying over the visual language while teaching this workflow. — 00:02:17 · YouTube
  • ★★ Adil would have GPT-6 Astra check the technical claims before the narration is generated, since model names, access, inputs and duration limits change; an uncertain feature is verified in official documentation or left out. — 00:02:17 · GPT-6 Astra

00:03:21 — Build the script and plan your takes

  • ★★ Adil structures a tutorial script as: the finished result, the ingredients, the production, and the checks that make it publishable, with every section answering the viewer's next question. — 00:03:30
  • ★★★ Adil splits the presenter script into short takes, each one complete thought ending at a natural sentence boundary, which makes generation easier to direct and gives clean editing points. — 00:03:30
  • ★★ With narration split into takes, a sentence that needs a correction is regenerated alone instead of paying to rebuild several minutes of narration. — 00:03:55
  • ★★ Adil warns against calculating running time from word count alone: word count gives a first estimate, but the generated performance sets the real duration, checked in the timeline. — 00:03:55
  • ★★★ For a generated presenter of himself Adil used a short excerpt of his own video, with his face, voice, gestures, framing and studio, and advises a clear reference with steady light, visible face and audible speech. — 00:04:16 · AI avatar

00:04:24 — Generate an AI talking head from your own video

  • ★★★ Adil writes each talking-head request in two sections: how he should look and sound (clothes, desk, background, camera angle), then only the words he will say. — 00:04:45 · AI avatar
  • ★★ Adil tells the model to start each take promptly, finish the complete sentence and leave a small clean hold at the end. — 00:04:45
  • ★★ Adil has the presenter footage generated without music or built-in graphics, so he controls those layers later in the edit. — 00:05:14
  • ★★★ Adil generates the opening take first and inspects it: whether the face stays consistent, the voice resembles the reference, the eyes look natural, the lips follow the words and product names are pronounced right. — 00:05:14 · lip sync

00:05:29 — Check facial consistency, voice and lip sync, and review every take

  • ★★ Once the sample take works, Adil reuses the same direction and the same source reference for the remaining takes, changing only the dialogue and timing. — 00:05:43
  • ★★ Adil keeps a simple take record: take number, prompt, result and selected file. — 00:05:43
  • ★★ A generated talking-head take can look good while dropping words, adding a phrase or ending too early, so Adil compares each take's speech with the script at normal speed. — 00:06:11
  • ★★★ Adil says that in an agent-run production the assistant organizes the work and the creator judges the result, fixing a failed section and trying again. — 00:06:11

00:06:32 — Add visual demonstrations and screen recordings

  • ★★ Adil matches each visual to the narration: the input file when discussing references, the prompt when explaining its structure, the result when discussing generation. — 00:06:39
  • ★★★ For a technical tutorial Adil uses real interface captures, because generated screenshots can put controls in the wrong place or invent buttons; he crops them and makes the text readable. — 00:06:39
  • ★★ Adil says a tutorial needs the right visual when the explanation changes, not a different visual every second. — 00:07:15
  • ★★ For motion graphics Adil reuses existing assets: the After Effects project gives editable compositions and its rendered video gives finished animation. — 00:07:15 · After Effects, motion graphics

00:07:32 — Animate motion graphics in After Effects

  • ★★ Adil uses a workflow diagram to show how the tools connect, and keeps on-screen text readable and long enough on screen. — 00:07:55 · motion graphics
  • ★★ Adil's first pass in DaVinci Resolve lays the talking-head takes in script order, listens to the whole story and trims the dead space at the joints. — 00:07:55 · DaVinci Resolve

00:08:11 — Assemble the video in DaVinci Resolve

  • ★★ Adil makes the first edit pass about the spoken explanation, because graphics cannot rescue a section that is confusing or repetitive. — 00:08:24
  • ★★ On the second edit pass Adil adds demonstrations and motion graphics above the presenter, uses a closer crop for emphasis and returns to the wider shot when the explanation opens up. — 00:08:24 · motion graphics

00:08:52 — Audio levels, music and sound design; keep the project portable

  • ★★ Separate generations differ in volume and background noise, so Adil matches the dialogue levels and listens to every joint with headphones. — 00:08:53
  • ★★ Adil keeps the project portable: script, generated takes, graphics, captures and exports stored together with clear names, plus an editable Resolve project and the After Effects working project. — 00:08:53 · DaVinci Resolve, After Effects

00:09:36 — Final quality checks, export and disclosure

  • ★★ Before the final export Adil watches the whole timeline for missing frames, accidental gaps, unreadable text, clipped speech and abrupt changes in the voice. — 00:09:38
  • ★★ Adil checks the exported file itself, picture, sound and transitions, because a good-looking timeline does not guarantee a correct render. — 00:09:38
  • ★★★ Adil says a realistic generated presenter should be disclosed to the audience; in this video the disclosure is part of the opening and the demonstration. — 00:10:05 · AI avatar
  • ★★★ Adil's complete workflow: define the result, organize the inputs, verify the script, test the presenter, generate the takes, build the visuals and finish the edit. — 00:10:05

00:10:18 — The complete AI video workflow, and starting with a one-minute video

  • ★★ For a first try Adil advises one clear topic, one good reference clip and a short script: make a complete one-minute video, inspect it, then expand the process. — 00:10:33

Threads in this video

Mentioned

Central: AI avatar · GPT-6 Astra · Higgsfield MCP

Also: After Effects · ChatGPT · DaVinci Resolve · lip sync · motion graphics · YouTube

← Previous: Higgsfield 3D Jutsu: Build and Animate 3D Scenes from a Prompt · Next → Changelog: GPT-6 Astra. Now in Higgsfield MCP and Supercomputer