SuperAPI AI Directory

html-game-parallel-builder

Coding & development

View repository

Updated 11-13 16:18

About this skill

Web Game Parallel Build

Overview

This skill exists to cut wall-clock time. It builds a TypeScript browser game from modular src/ files using parallel coding subagents with batched integration checkpoints, so an entire playable game (or major fix) can ship inside a live podcast, classroom demo, or livestream window. The audience watches the game come up at http://localhost:<port>/ from run_web_server.sh during the show; that local preview is the immediate live target. GitHub Pages and the optional single-file export are post-show release paths, not the artifact the audience watches you build.

Without the wall-time constraint this skill would not exist -- a serial agent with TDD produces cleaner output for the same total cost. The whole point is to hit the deadline.

The skill hits the deadline AT the quality bar, not at the cost of it. Every rule that looks like rigor (typed cross-module contracts, smoke tests between batches, exact tsc --noEmit output in subagent reports, pause-and-invoke for type design) is there because it saves time at the end: 30 seconds of contract discipline per agent prevents 30 minutes of integration debugging in a 14-module pile at the boundary. Cutting any of these rules to "save time" moves the same time into integration debt and blows the window.

Mechanism: parallel coding subagents coordinated by an orchestrator around shared contracts and batched integration checkpoints. The orchestrator does NOT write game code (except Batch 1 foundation). It gathers requirements, writes type contracts, dispatches coding agents in parallel, runs smoke tests, and fixes integration issues. Type design lives in typescript-engineer; this skill does not restate any TypeScript rules.

Required upstream skills

Invoke these; do not hand-roll their work here.

  • typescript-engineer (this repo) owns all type design. Route to typescript-engineer/references/game-type-patterns.md for branded ids, save-file boundary, GameEvent unions, ECS shapes, and as const satisfies config tables; typescript-engineer/references/modular-type-design.md for feature-area ownership; typescript-engineer/references/strict-mode-flags.md for strict posture.
  • delegate-manager-to-subagents owns manager dispatch: read-only manager review, subagent ownership boundaries, evidence-first handoff. Invoke before dispatching any coding subagent.
  • parallel-plan owns the up-front decision of how many lanes to split into and when one agent should hand off rather than carry everything. Invoke once before Batch 1; revisit if a batch becomes lopsided.

What this skill owns

The toolchain scaffold (TypeScript config, ESLint config, npm scripts, check_codebase.sh gate, setup scripts, gitignore) is inherited byte-identical from starter-repo-template/templates/typescript/ and is NOT owned here. The skill ships those files under templates/ only so an orchestrator can scaffold a new game repo in one step; the canonical owner is the starter template, and edits to those files must land in starter first. See references/BUILD_ARTIFACTS.md#starter-parity for the parity inventory.

What the skill DOES own (the game-specific overlay on top of the starter scaffold):

  • The preassigned workstream layout for browser games.
  • The src/ module decomposition.
  • The build identity split (GitHub Pages default vs. single-file export) -- summary below; detail in references/BUILD_ARTIFACTS.md.
  • Web-platform gotchas (canvas + innerHTML, branded ids leaking to DOM, type-only imports under verbatimModuleSyntax) -- detail in references/SMOKE_AND_GOTCHAS.md.
  • Playwright smoke testing between batches -- detail in references/SMOKE_AND_GOTCHAS.md.
  • Game-specific shipped artifacts under templates/ (export_single_file.sh, src_index.html, src_layout.md, agent_prompt_template.md, playwright_smoke_test.md, deploy_pages_workflow.yml) -- inventory in references/BUILD_ARTIFACTS.md.

Build identity

Default to a GitHub Pages-ready TypeScript browser game that builds into dist/. Use the single-file HTML export only when the user explicitly asks for a portable file. The three driver scripts have distinct, non-overlapping jobs: run_web_server.sh for the live preview, build_github_pages.sh for the canonical dist/ build, and export_single_file.sh for the optional dist-single/ artifact. Do not blend them. Full per-script constraints and the shipped-artifact inventory in references/BUILD_ARTIFACTS.md.

Terminology

Canonical definitions live in references/DEFINITIONS.md. Naming constraints and legacy handling live in references/NAMING_GUARDRAILS.md. Capacity and sizing targets live in references/CAPACITY_AND_SIZING.md.

Use Milestone / Workstream / Work package only in planning docs, not in durable code identifiers. Use Stage / Pass / Step for durable gameplay pipeline or algorithm steps in code identifiers. If a repository already has phaseN_* artifacts, treat that as legacy naming and do not introduce new planning-phase names into code.

Preassigned workstreams

Use the A0-E ownership map in references/BATCH_DISPATCH.md unless project requirements force a different split. Read it before assigning files. Lane count, scope splits, and merges go through parallel-plan before Batch 1.

When to use

  • Building a TypeScript browser game in a time-pressured live context (podcast, classroom demo, livestream) where wall-clock time matters more than strict serial discipline.
  • Multiple subagents will write .ts, CSS, and HTML concurrently.
  • The app has interacting modules (data generation, display, user interaction).
  • Canvas rendering is involved (gel visualizations, charts, game graphics).
  • The release target is GitHub Pages (default) or a portable single HTML file (optional export).

When NOT to use

  • Simple single-file edits.
  • Server-side applications.
  • Multi-page apps with heavy build tooling beyond tsc + esbuild (Vite, Webpack, Next, etc.).
  • Tasks where one subagent can finish in under 10 minutes; the parallelization overhead is not worth it.
  • Off-stage work where wall-clock minutes are not a constraint -- a serial agent with TDD will produce cleaner output for the same total cost.

The process

digraph process {
    rankdir=TB;

    "1. Gather UI preferences" [shape=box];
    "2. Define minimum viable scope" [shape=box];
    "3. Write type contracts (src/types/*.ts)" [shape=box];
    "4. Configure build (templates -> src/, dist/, .github/)" [shape=box];
    "5. Batch 1: Foundation (sequential)" [shape=box];
    "tsc + smoke 1" [shape=diamond];
    "6. Batch 2: Infrastructure (parallel)" [shape=box];
    "tsc + smoke 2" [shape=diamond];
    "7. Batch 3: Core features (parallel)" [shape=box];
    "tsc + smoke 3" [shape=diamond];
    "8. Batch 4: Advanced features (parallel)" [shape=box];
    "tsc + full playthrough" [shape=diamond];
    "Fix?" [shape=diamond];
    "Fix integration bugs" [shape=box];
    "Done" [shape=doublecircle];

    "1. Gather UI preferences" -> "2. Define minimum viable scope";
    "2. Define minimum viable scope" -> "3. Write type contracts (src/types/*.ts)";
    "3. Write type contracts (src/types/*.ts)" -> "4. Configure build (templates -> src/, dist/, .github/)";
    "4. Configure build (templates -> src/, dist/, .github/)" -> "5. Batch 1: Foundation (sequential)";
    "5. Batch 1: Foundation (sequential)" -> "tsc + smoke 1";
    "tsc + smoke 1" -> "6. Batch 2: Infrastructure (parallel)" [label="pass"];
    "6. Batch 2: Infrastructure (parallel)" -> "tsc + smoke 2";
    "tsc + smoke 2" -> "7. Batch 3: Core features (parallel)" [label="pass"];
    "7. Batch 3: Core features (parallel)" -> "tsc + smoke 3";
    "tsc + smoke 3" -> "8. Batch 4: Advanced features (parallel)" [label="pass"];
    "8. Batch 4: Advanced features (parallel)" -> "tsc + full playthrough";
    "tsc + full playthrough" -> "Fix?";
    "Fix?" -> "Fix integration bugs" [label="errors"];
    "Fix integration bugs" -> "tsc + full playthrough";
    "Fix?" -> "Done" [label="clean"];
}

Step list

Each step's full body lives in references/STEP_DETAILS.md. Steps 1-4 and Step 6 expand there; Step 5 has its own dispatch tables in references/BATCH_DISPATCH.md.

  1. Gather UI preferences (~2 min). Buttons vs dropdowns, theme, mobile support, references. Ask before writing code.
  2. Define minimum viable scope. 1-2 mechanics, 1-2 content types, fixed difficulty, no save/load in first pass. Defer features that come after the core loop.
  3. Write type contracts (~5-10 min, sequential). Real .ts files under src/types/, imported with import type, extensionless paths. npx tsc --noEmit -p tsconfig.json passes before any batch is green. See references/STEP_DETAILS.md.
  4. Configure the build (~2-3 min, sequential). Copy templates, set executable bits, place tsconfig, install the GitHub Actions workflow. See references/STEP_DETAILS.md.
  5. Batched agent dispatch. Four batches: Batch 1 foundation (sequential), Batches 2-4 parallel. Per-batch agent files, type imports, prompt-must-include block, and total wall-clock numbers in references/BATCH_DISPATCH.md. Manager dispatch follows delegate-manager-to-subagents.
  6. Integration fix loop. After the final smoke, if Playwright finds errors: console messages, snapshot, identify the broken module, dispatch a small fix agent, rebuild, re-test. See references/STEP_DETAILS.md.

Smoke testing

Run a Playwright smoke after each batch. Recipe and the pre-smoke tsc gate live in references/SMOKE_AND_GOTCHAS.md; that file also carries the web-platform gotchas table read between batches when something breaks.

When a smoke test fails, fix the failing module before proceeding to the next batch. If the failure is a contract violation, update src/types/ and re-run tsc before re-dispatching agents -- all subsequent agents need the corrected contract. This is load-bearing: proceeding past a failing smoke compounds bugs across batches.

Critical guardrails

The full guardrails set, the bottleneck rationale, the rationalization table, and the common-mistakes prose live in references/QUALITY_GUARDRAILS.md. The load-bearing rules below stay inline because the skill fails without them:

  • No coding agents dispatched before contracts are written.
  • No batch declared green without npx tsc --noEmit passing.
  • No skipping the smoke test between batches.
  • No as cast outside a brand constructor or save-file type guard; route the problem to typescript-engineer.
  • The manager does not write game code except Batch 1.
  • A coding subagent that needs a cross-module type pauses, invokes typescript-engineer for the typed stub, and waits before resuming -- it does not redeclare the shape locally to keep moving. Full delegation contract in references/BATCH_DISPATCH.md.
  • Every subagent report quotes the exact npx tsc --noEmit command and its exact success line (exit 0, no diagnostic output); "all green" without that evidence is treated as a false-green claim and re-dispatched.

Agent prompt template

Use templates/agent_prompt_template.md for every coding subagent. Read references/BATCH_DISPATCH.md before dispatch for the required prompt fields, evidence contract, and type-design pause-and-delegate procedure. Game-shape questions route to typescript-engineer/references/game-type-patterns.md.

Subagent dispatch

Dispatch a fresh subagent for each atomic task. Reusing a subagent across tasks carries stale context, encourages drift, and weakens independent judgment. SendMessage is for status only; do not use it to chain follow-on editing work onto a teammate that has already finished its assigned task. See docs/REPO_STYLE.md.