# Game Engine Production

> Create scoped games, prepare game or engine video reviews, or use game worlds as virtual sets for original video production. Use for PlayMCP and engine-based production requests; qualify each engine and capture route separately.

- Skill: `markoblogo/game-engine-production` (Agent Skill)
- Install (CLI): `npx get-skill add markoblogo/game-engine-production`
- Raw SKILL.md: https://api.skillmd.com/api/skills/markoblogo/game-engine-production/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: markoblogo (https://skillmd.com/u/markoblogo)
- Updated: 2026-11-18
- Page: https://skillmd.com/skills/markoblogo/game-engine-production

---


# Game-engine production

Choose the requested mode: playable game, game/engine review, or engine as a
virtual set for a film about another subject. Read the corresponding scope and
production contract in [the workflow](../../docs/workflows/game-engines.md).

For PlayMCP, read the connected service's current guides and resume from the
matching case or schema-2 checkpoint: [current Jekyll case](../../docs/experiments/jekyll-toki-pona.md),
[Coffee Showdown](../../docs/experiments/playmcp-coffee-showdown.md) or
[SIRTET](../../docs/experiments/sirtet.md) or
[Möbius Walk](../../docs/experiments/dream-walk.md). Use recorded
project/site/revision identities rather than starting a duplicate project. Match
review evidence to the current build and storyboard. Keep approval quotes
verbatim; changes requested by the owner are not approval of an unseen result.
Respect existing task authorisation and remaining budget without asking again
for actions already covered.

## Scope-changing decisions

- Game: verify controls through actual play and inspect each promised ending; for
  intentionally endless play, test movement, pause, sound and repeated cycles without inventing a finale.
  A passing page-format check does not establish understandable mechanics.
- Review: demonstrate the exact function being described. Keep commentary and
  real UI/gameplay provenance aligned; do not replace an unavailable demo with
  an invented success claim.
- Virtual set: test world persistence, camera/action repeatability and capture
  on a short take before planning a full film. PlayMCP game evidence does not
  qualify a different engine or a general reusable-world pipeline.

## Retained PlayMCP lessons

Character cutout can remove a held mug. Inspect the object and grip after every
asset transformation. Human hands and alien forepaws need different anatomy.
For authored aiming, verify shoulder pivot, barrel axis, muzzle and target
geometry together; fixed body scale does not prove a correct firing pose.

Synchronise the storyboard after owner-directed changes before requesting a
review. A stale camera or ending promise can produce irrelevant findings.
Keep service blocking items, creative suggestions and owner preferences separate.

When making a film, capture the real running build with recorded inputs and
sound provenance. Keep scripted input distinct from human play. Verify capture
and montage independently of game QA; neither generated sprite clips nor a
successful hosted game prove an exported narrated film. Apply the workflow's
rights and delivery checks to the exact output. Keep media and credentials
outside the repository.

## First retained capture adapter

For this exact PlayMCP Canvas game,the [film record](../../experiments/2026-10-17/disco-coffee-game-overview.md)
retains captureStream plus MediaRecorder with the final safetyTrim audio bus,
scripted tap/hold logs and both outcomes. Check recording offsets and frame counts
after conversion:initial WebM packet warnings and a one-frame comparison shortfall
were caught in final QA. Bake static cards once rather than recompositing every
frame when only the gameplay is moving. Scope this adapter to the recorded game;
do not infer another engine exposes the same Canvas or audio graph.

For vertical game recuts,read the [Shorts workflow](../../docs/workflows/shorts.md).
The first two-outcome pair tests manual portrait closeups,simultaneous context
panels and extracted opening-frame covers. Inspect the final consequence after
reframing:the live actor's crop can miss exploded fragments. Technical QA does
not establish mobile UI coverage,automatic crop tracking or audience virality.

## Reverse-duel and filmed-world lessons

The [SIRTET case](../../docs/experiments/sirtet.md) adds scoped evidence:

- For competing controls trapped in immediate cancellation, try explicit control
  turns. Per-input immediate cancellation
  can trap held keys in one-cell ping-pong. Drive repeats from the game loop;
  keep gravity independent, protect the human burst and make repair/rescue finite.
  Pin owner-accepted mechanics before changing art. A personal playtest is not
  a measured population win rate.
- Separate objectives from timeouts: an all-clear, top-out and song expiry can
  have different outcomes. Replay must reset the platform's finished receipt as
  well as board/audio state. When the documented API lacks reset, verify a fresh
  document path that clears stale seek and preserves mute; test actual click and
  platform Play after both terminal branches.
- Continuous filmed characters need decoded playback, not currentTime changes
  every render frame. Verify frame counters and visible pose changes separately
  from artistic liveliness. Pause/seek/loop/ending ownership must not compete for
  one playhead. Keep background films muted; verify the intended song is sole audio.
- For an authorised intro trim, preserve the master, measure the cut boundary,
  shift the authored grid/words, shorten the game clock, and refresh storyboard
  and asset provenance. A new analyzer estimate must not silently rebalance the
  owner's accepted game.

Publication, contest entry, automatic service share clips and authored YouTube
video delivery are separate states. Record the returned game identity and then
check the challenge entry list. Honour owner-requested sequencing: a future film
proposal does not authorize starting its capture/render. Retain technical QA,
reviewer opinions and owner acceptance separately, including disputed flags.

For a player-paced book game, native review needs its own adapter: initialise
the URL bot as well as responding to later setBot calls, and restore seek state
before large assets load. Preserve an explicit pause when readiness completes.
Use a documented legal-input replay for feed previews; do not label it human play
or impose its timing on the reader. Book/card words are game content; remove
authored control instructions and HUD for the platform contract. Reviewers may
still propose rhythm mechanics for an untimed educational game: preserve the
owner's chosen genre and record that mismatch rather than silently replacing it.

## Narrative flow and native chrome

Test the actual public game page through its native Start button, then provide
only the inputs the design asks the person for. Do not silently advance reading
or feedback with test-only Enter presses: that can hide a stalled user journey.
For narrated sequences, transition after actual audio completion and a brief tail;
use the paused game/audio clock for waits, and keep answer choices untimed when
requested. Ignore stray clicks during intro/reading/reactions, and finish only
at the promised terminal outcome. A format pass is not end-to-end flow evidence. Update brief design answers as well as storyboard timings after owner corrections; stale design metadata can contradict the tested runtime and produce a false timing blocker.

Test pointer cards against the containing platform's controls, not only the raw
host URL. In the Jekyll case the parent scrub bar intercepted bottom-row answers;
a64 CSS-pixel bottom reserve for normal scenes keeps targets clear. Film/clean
capture keeps full framing. Verify with ordinary pointer input, no forced clicks,
and both outcomes; preserve owner corrections separately from acceptance.

## Capture and editing follow-up

The [SIRTET videos](../../experiments/2026-10-18/sirtet-videos.md) retain three lessons:
- Verify a complete real-engine outcome before promising footage. A terminal fixture
  tests rendering, not gameplay. If a local demonstration changes a seed, disclose
  the sequence change on-screen and in copy; retain rules hashes and legal input
  replay. Never imply it is a victory on the public seed or alter the live game for a shot.
- Conversion is not native editor qualification. Compare layer spans, transforms and
  packaged media hashes; record inserted backgrounds/name changes. A raster title
  layer is not editable type. Keep downstream rendering separate from conversion.
- Run ASR on clean narration before music. Compare the known script, retain raw
  substitutions/zero-length intervals and separate corrections. Word timestamps
  are estimates; manually bound repairs to the source line and inspect encoded
  text against the actual all-clear, rather than trusting an SRT alone.

## Resume production memory

Resume the case's schema-2 [production project](../video-production-router/references/projects.md)
before reading historical logs or creating another game. Keep owner decisions,
revision-bound QA, rejected voice probes and explicit waits in its journal.
A manual production project records an existing game/virtual set without claiming
an automatic router provider. Old state files are retained migration sources,
not competing current instructions; do not restart paid work to rebuild context.

Attach scene/capture plans and reusable materials to that same project's
[production composition](../video-production-router/references/production.md).
Label real versus scripted gameplay and generated performances; plan portrait
cuts separately. A `bundle --draft` may contain only pending-work reports when
footage/finals are absent. It does not authorise capture, rendering or hosting.

## Replay and vocabulary checks

A terminal state is not a complete ending test. Exercise both outcomes through the actual public player, wait without another input for the promised exit/reset, verify one native result with the correct win flag, then start a zero-score round. A face playing to its last frame is not evidence of a usable finale.

For randomised educational rounds, retain a source-bound bank and contextual glosses, deal once per round, shuffle answer positions, and test with fixed seeds. Preserve narrative groups and the winning threshold. Prefer words actually answered recently rather than reserving unshown words; unavailable browser storage must fall back safely. Keep review/demo history separate from human history. Distinguish full source-paragraph vocabulary from words visible in the short reading excerpt. New speech prompts need their own text/hash/decode evidence; do not rerun rejected-language probes.

## Ambient capture and publication records

For a long background video, use the [bounded game-canvas recipe](../../docs/workflows/ambient-game-video.md). Retain filming-only timing changes and repeated-cycle disclosure. Version changed scripts when a host caches them. Native format pass, blind style advice, actual viewing/listening and owner approval are different evidence; never create score/targets or fake approval to satisfy an unrelated rubric.

Deliver publication copy as direct TXT/Markdown files when requested, not inside ZIPs. Later owner-supplied video links are separate publication observations pinned to the retained master; they do not prove metrics, independent listening or uploaded-byte identity. Commit only sanitised receipts and asset hashes, never raw account output, balances, signed URLs, credentials or media.

