vast-ai-research/edenlive/edengolem

Claude Code and agent skill by vast-ai-research/edenlive

Trust tier 3
  • community
Source check pending

Checking that the SKILL.md file is still in the repository.

Build or revise standalone Three.js browser game prototypes with a silent desktop keyboard/mouse 3D scene, one continuous round, and reachable win and loss states. Use for bounded playable game requests within those limits; not 2D-only, audio-led, or open-ended games without both terminal outcomes.

Install

npx vskill@latest install vast-ai-research/edenlive/edengolem

SKILL.md

EdenGolem(v0.1)

The final project includes serve_step.py and launch.sh in a copied output directory. It does not promise production art or a full commercial game.

Stage IDs, gate commands, the refinement cap and timeouts live in forge/_shared/constants.py. Data shapes live in docs/specs/. Flag behavior is each tool's --help. A project's live paths and commands come from forge/task_packet.py through forge/next.py. This file routes; grimoire/ teaches the current stage; README.md navigates.

Start or resume

Authoring tools stay in the skill checkout. Generated projects have no forge/, so python forge/... from the output directory will not work. Pass --project-root <output> and take live commands from the task packet; handwritten examples only show the convention.

Choose the workflow before editing:

  • New game: follow intake below.
  • Unfinished game: read <output>/.state.json and request its current packet with python <skill-root>/forge/next.py <output>/.state.json --task-packet.
  • Change a passed game: use revisions, including when the packet reports complete. Start a revision before changing the retained artifact; ordinary stage closeout does not handle revisions.

For a new request, read the intake scope boundary and intake prompt first. If the defining mechanic is outside scope, stop intake, record the blocker, and do not generate a brief.

For an in-scope request, write and validate the brief, then initialize the standalone project (setup). Initialization creates an intake-in-progress state and schema-only evidence; it does not pass intake. Complete the intent review before advancing to scene.

There is no unresolved or skipped stage status. Before initialization, an out-of-scope request has no project state. After initialization, a blocked intake stays in_progress with a recovery note or evidence; never record passed.

Stages

The pipeline has one intake phase and three build stages.

StagePrompt and contractsDeliverableAutomatic checksExplicit review
intakeparse the prompt; scopeA validated brief with an explicit inventory of concrete theme elementsBrief schema validationIntent review comparing the brief to the request
scene (step1)build the static sceneA complete static composition made with asset factories: background, actors, mechanisms, board or interface elementsMetadata + static browser check (canvas CSS matches #app at DPR 2; custom 3D settings name a considered catalog path). Mesh counts are not gated.Experience and performance review of recognition, coverage, composition and geometric simplicity
behavior (step2)build the playable loop; input contractA complete playable loop: applicable controls, camera, actions, navigation, collisions, progression, win/loss, terminal stop, restart and the feedback play requiresReal-input behavior check + closure playthrough. Locomotion, aiming, resources or autonomous agents import a catalog module from the run page or record catalog_skips.Experience and performance review of controls, feedback and the playable loop
acceptance (step3)finish and accept; playthrough contractFinal UI and presentation, integration repairs, full real-input replay and relocated launchAcceptance + relocationExperience and performance review of presentation, outcomes, restart and standalone launch

Read the current prompt and only the capability references it selects. Choose behavior by capabilities (selection, movement, aiming, timing, collision, turn order, targeting, etc.), not by genre labels. A menu game may need no avatar; a creature battle may need a fixed encounter; a tactical board may need grid navigation and turn state. Combine capabilities for hybrids; consult optional patterns when a selected mechanic needs them.

Preserve defining actions and explicit values within the product scope: silent desktop keyboard/mouse play, a 3D modeled scene, one continuous round, and reachable win and loss states. Scope-changing reductions follow intake, not silent substitutions. Prefer fitting catalog factories; missing assets use simple primitives, while the game owns input, state and rules. Stage prompts route detailed scene, camera and UI contracts only when applicable.

Closeout

After stage edits, obtain a fresh task packet and execute its closeout.check command. It includes the required strategy for behavior and acceptance. Inspect evidence, complete the printed review, then execute closeout.review with that review path; review submission does not rerun gates. Individual gates and preflight are diagnostics. See stage closeout for evidence reuse and recovery for bounded failure handling.

The scene check proves static coverage, framing, WebGL canvas CSS fit and catalog-first records for a custom 3D setting. Behavior runs two scopes: the behavior check proves applicable real inputs, persistent effects, and catalog camera/combat/resource reuse from the run page or recorded skips; the closure check proves one representative terminal outcome, a real same-page restart and a fresh round that still accepts input. Acceptance proves the full win/loss/restart sequence with the final UI integrated, plus the copied launcher. A closure pass never substitutes for acceptance; each report names its own scope. Composition and geometric simplicity stay experience review.

Use the shared contracts when editing cards, manifests or state: shared contract, capability routing, pre-ship checklist, and recovery. Keep evidence under each run's evidence/ directory. Report an unevaluated or inapplicable check as such with its reason; never present one as a pass.

Security

  • Security scanPassed, 90 of 100Pattern checks on the skill's files.
  • Model reviewPassed

Quality evaluation

  • Task evaluationNot runNo evaluation yet. Run one in Skill Studio.

Source

Popularity

Repository stars
315
Repository forks
13
7-day trend score
0

Stars and forks are for the whole GitHub repository, not this skill alone. Not refreshed yet.

Works with

Any agent vskill installs to: 56 targets in all. These 39 have a page on this site.

  • Claude Code
  • Cursor
  • GitHub Copilot
  • Windsurf
  • Codex
  • Gemini CLI
  • Cline
  • Amp
  • Roo Code
  • Goose
  • Continue
  • Kilo Code
  • Junie
  • Trae
  • OpenHands
  • Qwen Code
  • AI Agent
  • Tabnine Chat
  • Sourcegraph Cody
  • Amazon Q
  • Firebase Genkit
  • Sweep
  • Devin
  • Phind
  • Mentat
  • Claude Desktop
  • Aider
  • GPT Pilot
  • Plandex
  • bolt.new
  • v0
  • Replit Agent
  • Anse
  • TypingMind
  • BetterChatGPT
  • LibreChat
  • Warp AI
  • Pieces
  • Fig AI
Embed badge
vskill VERIFIED badge
[![vskill VERIFIED](https://verified-skill.com/api/v1/skills/vast-ai-research/edenlive/edengolem/badge)](https://verified-skill.com/skills/vast-ai-research/edenlive/edengolem)

Something wrong with this skill or its page? Tell us.