Three games, two AI tools, one kit

The module spec is still the spine: 12 weeks, 8 competency units, Godot 4, persistent project memory, competency-based assessment. The v1 plan stays as the record of the questions. These are the answers.

Neo's question: integrate Codex Astra into the kit, or run Astra as its own platform for designing and building the game, with Claude Code as the back end? Answer: one kit that both tools read, with the jobs split the way RUSH proved they should be. That gives us the "separate platform" feel for the student without building a platform.

What RUSH proved

police sentri : RUSH is a 3D chase game in a Malaysian pasar, built in Godot for phones. Between 15 and 16 September 2026 it went from a native prototype to a live game on Google Play, and was submitted to the App Store the same afternoon, all under the Todak Studios accounts and driven entirely from the fleet. Two AI tools did it, in the same repository, each on its own jobs:

The lesson is not "Astra is creative, Claude is technical". It is that the split worked by job, inside one repository, with one tool briefing the other in writing and one tool holding the keys. That is exactly what the kit has to package for a student.

The two lanes in the course

Integrated or separate? Both, in this order

One installer sets up both tools, the engine and the kit. From then on the student types plain requests and the kit does the plumbing. Everything below is a command any tool can fire, which is what keeps the "later" option open.

• Start a game. "New endless runner called Sky Dash" scaffolds a Godot 4 project from the genre template, opens the student's project memory, and registers the game with the Studios publisher desk (name, package id, placeholder icon, privacy page).
• One instruction file, two readers. The kit writes the same course rules, week, game and conventions into the file Codex reads and the file Claude Code reads, so both tools know where the student is and what the lanes are.
• Project memory. The persistent brain from the module spec: design decisions, code conventions, bugs, all remembered across 12 weeks. Claude Code owns it; Astra reads it.
• Assets. Sprites, tilesets, 3D models, rigs, music and sound effects through Academy-hosted services, the same image-to-3D, rigging and voice pipeline Todak runs in production. Students never hold raw API keys. Every generation is logged and metered.
• Build and publish. "Build for web" or "push to Android testing" sends the job to the build farm and returns a link. No Xcode, no Android SDK on the student's machine.
• Review. "Review my game" runs the AI reviewer from section 06 and files the report to the classroom system.
• Update. "Ship version 1.0.1" bumps the version, writes the what's-new notes, rebuilds and pushes an update to a game that is already in the store.
• Guard rails. Hooks that log every prompt, remind both tools which week and which game the student is on, and block destructive commands.
• Helpers. A playtester agent that runs the game headless and reports what broke.

Unchanged from v1. The student chooses one genre per slot. Lan can add or remove entries.

The three rules

1. Three families, not three of the same. Action, Puzzle and Strategy, Narrative, Simulation. A student's three games come from at least two different families, and the final game changes camera or dimension from games 1 and 2.
2. If it goes to a phone, it has touch controls. Runners, racers, puzzles and sims are naturally mobile. The mentor steers the capstone choice with that in mind.
3. Every game ships complete. Title screen, how-to-play, win and lose states, sound, and a saved high score or progress. Small and finished beats big and broken.

Decided 19 Sep: every student game is published under the Todak Studios brand, on the Studios accounts for Steam, itch.io, Google Play and Apple. The kit connects to them on the first day, and the student never sees a store key.

What being under Todak Studios means, now decided

• Publisher of record: Todak Studios, under its own brand. No separate "Academy Games" label.
• The student keeps authorship. At enrolment they sign an IP licence that lets Todak Studios publish, market and update the game and use its design and characters. Section 09 lists what the licence must cover.
• If a game earns money, it is split three ways: 50% Todak Studios, 20% Todak Academy, 30% the student. Decided by Neo on 20 Sep (v2.1). The split goes into the IP licence in section 09.
• Naming rule is fixed in the kit before anyone publishes, because a package name can never change after the first upload: com.todakstudios.academy.<game>. RUSH itself is com.todakstudios.policesentrirush.
• Credit. The student's name goes on the store listing and in the game's credits screen.

Things that bite, already known

• Godot web games need two security headers on the hosting site. The showcase sets them and itch.io has a tick-box for it.
• Apple's developer agreements must be accepted or the store keys stop working for days. This blocked the Studios account for months earlier this year. Check before each cohort's capstone week.
• Every store listing needs a privacy policy page before the first submission. The kit generates one per game, the way todakstudios.com hosts RUSH's.
• Google Play's first upload for any new app has to go through the console by hand before the automated path works. The publisher desk batch covers this.

Decided 19 Sep: all three games pass one review gate. An AI reviewer does the legwork and assists three people: the lecturer, Neo for Todak Studios and Lan for Todak. Neo can bypass the gate to publish first. Because games can be published first and fixed after, updating a live game becomes a lesson, not an accident.

Why Neo's bypass is a feature

• Google Play approved RUSH about 20 minutes after submission. Web and itch.io are instant. A blocked-but-working game teaches nothing while it waits for three calendars to line up.
• The App Store takes longer and can reject, so the bypass matters most for web, itch.io and Play. The AI reviewer still runs; the bypass only skips the wait for human approval.
• Every bypass is logged with who used it and why, so the record of the game still shows what the review said.

Updates are part of the course

• Week 12 requires at least one update to a game that is already published: a version bump, what's-new notes, a rebuild and a push through the same lane.
• Students learn what a store update actually involves: Play's staged rollout, Apple's update review, why a package name and signing key are forever, and what a hotfix is.
• The review gate runs on updates too. The report compares the new build against the last one.

The classroom is the Academy's high-spec Windows PCs. The one Mac in the loop is the build farm.

Why the office, not the classroom

• It joins the fleet. The agents that published RUSH and ApaNakMakan already run there: tr-home drives the store consoles, EdgeXpert runs the watchers that message Neo when a store changes state, the NAS holds builds and assets. The Mac mini plugs into that instead of standing alone in a lab.
• The Studios keys stay in the office. The farm holds Apple and Google signing material. That belongs on a machine Todak Studios controls physically, not in a shared classroom.
• Students never touch it. They push code to GitHub. The farm builds every platform and the link appears in the showcase. Nothing about the farm's location changes the student's day.
• Until it arrives, the kit dry run uses tr-home for Android and web builds. Only the iPhone builds wait for the mini.

Three layers as in v1, now covering both tools. Nothing exotic to build; the work is the collector, the database and the lecturer's view.

Layer 1 · Claude Code's own telemetry

Claude Code can export a stream of events to any server we name. The kit installer switches it on. With the content flags set, each event carries the full prompt text, the full reply, every tool call with file paths, the model used, tokens and cost, lines added and removed, commits, the session id and the student's login email.

CLAUDE_CODE_ENABLE_TELEMETRY=1
OTEL_LOG_USER_PROMPTS=1
OTEL_LOG_ASSISTANT_RESPONSES=1
OTEL_LOG_TOOL_DETAILS=1
OTEL_EXPORTER_OTLP_ENDPOINT=https://telemetry.course.neotodak.com

Layer 1b · Astra's transcripts

Codex keeps a full transcript of every session on the machine it runs on. Our fleet already copies those transcripts into the Todak brain, which is how RUSH's Astra sessions are on record today. The kit's collector does the same for each student, tagged with the same login and session ids.

Layer 2 · the kit's hooks

On every prompt and at the end of every turn, the kit writes a row to the course database with the context the tools do not know: which cohort, which game, which week, the current commit. The same hook reminds both tools where the student is in the course, so the answers stay on the curriculum.

Layer 3 · Git

Every student game lives in the Academy's GitHub organisation. Commit history is the tamper-proof timeline. The lecturer view links prompt to commit to build to review report to published game.

What the lecturer sees

The Academy's classroom system already has students, classes, enrolments, progress and AI grading. We add a prompt-history table and a Prompt Journey page per student. The existing AI grading scores prompt quality weekly, so the student gets a growth chart at graduation and we get proof for marketing that the course teaches prompting, not just games.

Rules that come with it

• Consent at enrolment under Malaysian data protection law, as one of the three documents in section 09.
• Students see their own journey. Lecturers see their class. Exports and deletion on request, within what the NDA and the IP licence allow.
• A student can technically switch recording off on their own laptop. The answer is a course rule, not a lock: work with recording switched off on purpose does not count as evidence, and the Git history stays either way. A recording fault is different: the Workbench shows "saved" or "waiting to upload" at all times, and a fault gets recovery, never a penalty (v2.2, on Astra's advice). If cohort 1 shows this being gamed, the "later" option in section 02 puts the backbone on our server and makes the record unavoidable.

Decided 19 Sep: three documents, signed at registration, before the first lesson. What the students are about to see is not public knowledge, and some of them will come from the game industry. This section lists what each document must cover. A lawyer drafts the wording.

• Under-18 applicants need a parent or guardian to sign all three. Decide whether cohort 1 accepts them at all.
• Sign once, in one sitting, on the registration form, not spread across the term. The classroom system stores the signed copies against the student record.
• Plain-language summary on the first page of each document, so an 18-year-old knows what they are agreeing to.

The module spec's cohort figures stay as the base: about RM5,100 for 15 learners and RM6,700 for 20, with Claude Pro seats as most of the bill. Two lines change from v1: the Astra seat is new, and the Mac mini is now approved.

Neo's item 5: build a working kit and test it before anyone else does. The order below puts the risk first. Cohort 1 is not scheduled until the gate at the bottom is passed.

1. Game Kit v1, two lanes · 2 weeks. The installer, the shared instruction file for both tools, genre templates for every menu entry, the memory wiring, the backbone commands (new game, assets, build, publish, review, update), the logging hooks.
2. Reference game 1 through the kit · 1 week, overlapping. One of us plays the student: Astra designs and builds a game-1-sized 2D game, Claude Code scaffolds, builds and publishes it to the web and itch.io under Todak Studios. This is the test of the one unproven part, Astra writing Godot gameplay with a beginner.
3. Review gate · 1 week. The AI reviewer, the one-page report into the classroom system, the approve and bypass buttons for the lecturer, Neo and Lan, and the bypass log.
4. Prompt collector and Prompt Journey · 1.5 weeks. Collector for both tools, prompt table in the classroom database, the lecturer page, the weekly prompt score.
5. Studios publisher desk · 1 week. The manifest queue, the batch checklist for creating store records, the naming rule, the per-game privacy page.
6. Build farm · 1 week after the Mac mini arrives at the office. Until then, tr-home does Android and web builds for the dry run.
7. Reference game 2 and RUSH as reference 3 · 2 weeks. A game-2-sized project through the kit to Play internal testing, and RUSH re-run through the kit's publish and update commands so the capstone path is the RUSH path.
8. Showcase gallery · 1 week. Every student's three games playable on course.neotodak.com, with the required web headers.
9. Paperwork to the lawyer · start now. Section 09 is the brief. Needed back before the enrolment form goes out.
10. Windows lab image and home installer · 3 days. One image for the classroom PCs, one script for laptops at home.

The seven questions from v1, with Neo's answers of 19 September, and the small items that are still open.

1. 1Claude Code with the kit, own platform only later. Decided 19 Sep Yes, and Codex Astra joins as the second tool. Build the kit first and test it on ourselves.
2. 2Two small games then one final game inside 12 weeks. Confirmed 8 Sep
3. 3Genre menu and the three rules. Standing Lan adds or removes entries.
4. 4Graduation and publishing rule. Decided 19 Sep All three games pass the review gate: AI reviewer assisting the lecturer, Neo and Lan. Neo can bypass to publish first. Updating a published game is taught.
5. 5Publisher label. Decided 19 Sep Todak Studios brand. IP licence signed at enrolment; the student gets a share of any profit.
6. 6Mac mini build farm. Decided 19 Sep Neo buys it and sets it up at the Kenwingston office with the NAS, tr-home and EdgeXpert.
7. 7Consent and who sees the records. Decided 19 Sep Consent, IP licence and NDA, all signed at registration.

Still open, all small

1. aProfit share: 50% Studios, 20% Academy, 30% student. Decided 20 Sep
2. bWho drafts the three documents, and whether cohort 1 accepts under-18s. Lan
3. cAstra seats: one per student for all 12 weeks, or from week 2 with a pooled set for the 3D weeks. Cost model
4. dMac mini memory size, and its name on the fleet. Neo
5. eWho plays the student in the dry run, and when the kit build starts. Neo

• v2.5 · 21 Sep 2026. Student-facing change: one box, two AIs behind it (a router in the kit). The lanes in section 02 are unchanged behind the box. See Classroom v2.1.
• v2.4 · 20 Sep 2026. Epic Games Store and Roblox added to the publishing table, from the store audit in Stores v1.
• v2.3 · 20 Sep 2026. Home moved off Neo's personal estate: domain, hosting, database, build farm and AI services now per Home v1. Section 07's fleet placement is superseded.
• v2.2 · 20 Sep 2026. Astra's review adopted: beginner trials added to the gate before cohort 1; kit timeline restated as ten to twelve weeks to a pilot; recording faults get recovery, not penalty; the review gate decides whether a game ships, the walkthroughs and viva decide whether a student passes.
• v2.1 · 20 Sep 2026. Profit split decided: 50% Todak Studios, 20% Todak Academy, 30% the student. Mac mini price updated to the 2026 lineup. Teaching format, the student's screen and the infrastructure bill now live in Classroom v1.
• v2 · 19 Sep 2026. Neo's answers to the seven v1 decisions folded in. Codex Astra added as the second tool with the lanes proven on police sentri : RUSH; publisher fixed as the Todak Studios brand with an IP licence and profit share; one review gate for all three games with Neo's bypass and an update lesson; Mac mini approved and placed at the Kenwingston office on the fleet network; consent, IP licence and NDA at registration; build list reordered to build and test the kit on ourselves first.
• v1.1 · 8 Sep 2026. Added the stance "Not prompt-only. Not vibe coders." with the proposed ways of holding that line.
• v1 · 8 Sep 2026. First working plan. Genres as a menu; publishing wired to Studios accounts from day one; classroom is Windows PCs; games 1 and 2 small, game 3 final.