@mattpocockuk: Going live, doing a full feature build using: - /grill-with-docs - /handoff - /prototype

Going live, doing a full feature build using:

• /grill-with-docs
• /handoff
• /prototype

https://t.co/NK1Bz4gCq6

TL;DR: Matt Pocock attempts to add a “delivery calendar” feature to his course video manager during this live stream. The main effort goes into aligning terminology with the AI agent, clarifying the data model (relationships between courses, creative briefs, and videos), and exploring the impact of coding standards on Agent Experience (AX).

Stream Background: Warming Up in Chat, OBS Glitches

After going live, Matt interacts with the audience, mentioning he was feeling under the weather recently and is now recovered. He uses OBS for streaming and runs into a window-always-on-top issue, which he fixes. About 250 viewers are in chat, including people from Indonesia, whom Matt thanks. He also mentions using WSL (Windows Subsystem for Linux) to bill the GPU cost to his company, and notes he primarily uses Claw Code for development but plans to try Codec.

Build Goal: Adding a Delivery Calendar to the Course Video Manager

Matt explains the core task of this stream—adding a delivery calendar to his “Course Video Manager.” This is an internal tool he uses to run his entire business, already containing courses, creative briefs, TL Draw diagrams, etc., but lacking a calendar-based scheduling concept. The delivery calendar will give all deliverables (courses, videos, etc.) a clear due date, helping him plan and review progress.

He plans to use the “Grill with Docs” tool for end-to-end development. The codebase is a React Router app, fully TypeScript, heavily dependent on external tools like ffmpeg.

Delivery Calendar Data Model: A Tug-of-War Over Terminology

This is the core technical part of the stream. Matt talks with the AI agent via Grill with Docs, trying to define the data model for “deliverables.”

Terminology Conflicts and Branch Decisions

After reading the project’s context.mmd file (226 lines), the AI agent points out that Matt’s existing vocabulary doesn’t perfectly align with the “delivery calendar” concept:

• “Creative brief” is already a wrapper for video ideas.
• “Course” has no scheduling concept.
• “Video” is a production artifact.
• “Task” doesn’t exist in the domain.

The AI proposes at least three possible models, emphasizing: “Once the language is set, half the work is done.” Matt agrees, but then enters a back‑and‑forth discussion about the specifics of entities.

Core Debate: What Exactly Is a Deliverable?

• AI’s initial proposal: A deliverable is a parent type for Course, Creative Brief, or Task. The calendar shows instances of any of these, replacing Google Docs.
• Matt’s feedback: Courses are obviously deliverables, but what about videos? There are two kinds: standalone videos (already wrapped by creative briefs) and videos within courses. He thinks the real deliverable is “a video created from a creative brief”—the creative brief is lightweight (takes 5 minutes to create), while the video is time‑consuming and needs scheduling.
• AI’s second proposal: Introduce “Ghost Video” to allow planning ahead within a course. Matt rejects this idea because ghost videos only make sense during the creation of a planned course, not for a calendar scenario.
• Final approximate model: Matt believes what goes on the calendar is “a commitment to publish a video.” A more accurate data model: Deliverable = Course or Creative Brief. When a creative brief is created and eventually sent as a video, that counts as a delivery. The video can be treated as a secondary artifact.

Matt laments: “Sometimes alignment is easy, but alignment like this is actually quite difficult.” The conversation takes up a lot of time, but Matt considers this “nitpicking” necessary because language determines all subsequent implementation.

Excluding “Tasks”

Matt explicitly decides that the V1 version will not include “tasks.” He sees tasks as essentially pointers to courses, which can later be created as entities in the database, but are not core to the delivery calendar.

Coding Standards and Agent Experience (AX)

During the conversation, Matt responds to a viewer’s question about whether code quality matters for AI agents. He emphasizes the importance of Agent Experience (AX):

“If you have junk code in your codebase, the agent will only produce more junk code.”

He believes coding standards are now more important for projects than before because agents are more sensitive to code quality than humans. Good code quality improves agent productivity; bad code leads to more bad code.

Quick Q&A on TDD and the Goal Feature

Near the end of the stream, a viewer asks: “Do you think it’s reasonable to use TDD with the Goal feature to complete tasks, instead of bringing Ralph into Sand Castle and running it? What are your thoughts on the Goal feature?”

Matt admits he hasn’t tried Goal yet, but essentially it’s the Ralph loop. He is cautious about how Goal is specifically implemented, mainly whether it uses “smart zones” and how it populates context. He doesn’t go into depth, just calling it a big question.

Stream Mishap: Lost Recording

During the stream, Matt accidentally loses a video recording (code snippet), leading him to decide not to record this stream and only keep the live version. He reconnects to avoid further losses.

Summary

Although this stream had less actual coding, it deeply demonstrated the real process of “using AI to assist in designing software architecture”—the main challenge lies in understanding business intent and precisely defining domain language. Matt concludes: “When you sit down to plan, ‘I’ll release this idea in three weeks.’ Does the video entity exist? No, it doesn’t.” This directly highlights the difficulty of aligning the data model with real‑world timelines.

Source: YouTube Live – @mattpocockuk: Going live, doing a full feature build using: /grill-with-docs /handoff /prototype (https://youtube.com/live/qzX05m11o20?feature=share)