Skip to content
Back to skills

Play Time Models

ASecurity

Genex's experience with games that call a language model while somebody plays. None of this is required: the platform facts (the player pays and approves, text or JSON only, no author key, the codes) are in `genex__skill {"name":"genex-tool-llm"}` and `genex__skill {"name":"genex-llm-in-games"}`, and the game's design is its own. Use what fits, ignore what does not.

  • 798 stars
  • 0 votes
  • 0 copies
  • 0 views
  • Added November 5, 2026
ai-agentsgoshell

Works with

  • cli

Security analysis

A100/100

Scanned November 5, 2026

npx -y skills add genex-games/genex-desktop --skill play-time-models --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Play Time Models?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Play Time Models
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/genex-games-play-time-models/badge)](https://www.skillsdirectory.com/skills/genex-games-play-time-models)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
# Models at play time (optional advice)

Genex's experience with games that call a language model while somebody plays. None of this is
required: the platform facts (the player pays and approves, text or JSON only, no author key, the
codes) are in `genex__skill {"name":"genex-tool-llm"}` and `genex__skill {"name":"genex-llm-in-games"}`,
and the game's design is its own. Use what fits, ignore what does not.

## When the feature cannot be a model call

- When the person does not want a play-time model, or the lane is off on the stand, an authored
  version often serves as well: a written dialogue tree, a table of lines, a rule-based director.
  Most good games are built that way.
- When somebody asks for something the lane cannot return (a 3D model typed into existence, paid by
  the player), JSON parameters, then render, is the shape that works: the model returns a
  structured description and the game builds it from assets and code it already ships (a creature
  assembled from generated parts, a room laid out from prefab ids, a palette, a stat block, a line
  delivered from pre-generated voice clips).
- Saying plainly in the handoff that the lane was off, or that the feature runs on its authored
  fallback, keeps the person's picture of the game accurate.

## Keeping the feature playable

- A feature built behind a graceful `unavailable` state (the NPC uses its authored lines, the quest
  falls back to the written one) never promises the player something that 404s.
- In the Genex app (`getShell() === "app"`) the lane has no approval surface; leaving the feature's
  entry point out and playing the authored fallback reads better than an error.
- Guests hold no wallet: showing the feature as sign-in-to-use, rather than hiding it, tells them
  what signing in gives them.
- Offering the featured models, and the full catalog only when someone asks, keeps the picker short.

## When a standing budget ends

| code | what tends to work |
| --- | --- |
| `grant_limit_reached` | authored behaviour returns; a button offers to re-request |
| `grant_stopped` | accept it silently and keep playing |
| `grant_expired` | as stopped; re-request on the next deliberate click |
| `waiting_for_plan` | wait out the stated time |
| `grant_insufficient_funds` | pause the thinking NPCs, say so once, stay playable |

## Keeping the burn low

- Batching several NPCs into one call that returns all their decisions, caching a decision until the
  situation that caused it changes, and picking the cheapest model that passes the game's own check
  all lower what a player spends.
- A call fired on a timer the player cannot see is spend they did not ask for.
- When the game hits `grant_concurrency` or `grant_rate_limited`, batching and caching usually fix
  it better than raising the limits.

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

View all in ai-agents →

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.