跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
设计与 UI

gating-intent-legibility

Measures whether a player who was never told the rules can read a game's intent off the screen, by sampling scenes from a recorded run and having an isolated agent that has not seen the design or source name the goal, the options, and the risk. Use when a game runs and its mechanics are verified but it is unknown whether the screen communicates what to aim for without a tutorial or HUD text, or as a cheap early gate that can fail a design for decision collapse before a telemetry sweep. Not for judging fun or difficulty, not for verifying that a mechanic matches its spec, and not for confirming that varied decisions survive continued optimal play.

abagames · 已有静态检查记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add abagames/agentic-gamedev-skills --skill gating-intent-legibility

技能说明


name: gating-intent-legibility description: "Measures whether a player who was never told the rules can read a game's intent off the screen, by sampling scenes from a recorded run and having an isolated agent that has not seen the design or source name the goal, the options, and the risk. Use when a game runs and its mechanics are verified but it is unknown whether the screen communicates what to aim for without a tutorial or HUD text, or as a cheap early gate that can fail a design for decision collapse before a telemetry sweep. Not for judging fun or difficulty, not for verifying that a mechanic matches its spec, and not for confirming that varied decisions survive continued optimal play."

Gating Intent Legibility

Purpose

Measure what a game's screen transmits to a player who was never told the rules. Nobody on the project can judge this — designer, implementer, and any agent that read the source all know the answer before they look.

The method is blind restoration applied to a new pair of layers: the artifact is a set of sampled gameplay scenes, the layer to reconstruct is the player's intent, and the withheld key is the design's stated core loop. Because an agent shown any screen invents a fluent goal, recovery is not scored alone — each grader must also predict a frame it has not seen, which is then revealed, so the game supplies the ground truth.

One capture set yields per-scene intent (goal, options, risk) and cross-scene divergence (do different situations call for different actions). Divergence is a property of the set of verdicts; no single scene carries it.

When to Use

  • A build runs, its mechanics are verified, and it is unknown whether a first-time player can tell what to do.
  • A "visually self-explanatory" design claim, or a protagonist/danger/reward legibility claim, has only ever been checked by its author.
  • Before adding a tutorial or explanatory HUD text.
  • As a cheap falsifier before a balance sweep: if every situation calls for the same action, telemetry will only confirm it more slowly.

When Not to Use

  • Confirming that decision variety survives continued play — no image set supports that claim. Use evaluating-gameplay-balance.
  • Verifying a mechanic against its spec — probing-web-game-mechanics.
  • Judging fun, taste, difficulty, or beauty — a blind grader rubber-stamps these, and taste is outside what any firewall gate can measure.
  • The build produces only title or attract frames.
  • No isolated grader can be spawned. Label any such run "self-audit", not a gate verdict.

Capture Contract

The instrument needs rendered frames; searchers run against models that draw nothing. Any engine, three capabilities:

  • C1 — frame capture at chosen ticks (a tick, not a wall-clock moment).
  • C2 — deterministic replay of an exported input track into the rendering build, same states at the same ticks.
  • C3 — state fingerprint at those ticks (score, entity count, position, phase) to confirm C2 held.

Check C1–C3 against the project; do not assume a tool provides them because it is adjacent. Headless tooling that renders nothing cannot satisfy C1 however reproducible it is. For a disposable prototype built to be measured rather than shipped, prefer a route where C1 and C3 already exist — prerequisite capture work is a poor investment in something to be discarded. Route selection, not an engine endorsement.

Verify C2 by replaying twice and comparing C3 fingerprints against the search run. If C1 is missing or replay diverges: do not reconstruct frames from simulated state or approximate the track by hand — mislabelled frames produce a confident wrong verdict. Record directly against the rendering build instead, treat the run as naive-track, and report collapse as not measurable.

Read references/capture-capability.md when the capture path is not established, when deciding how C1–C3 are satisfied on a given engine, or when a replay check has failed.

Procedure

  1. Freeze the withheld key before any capture is graded: intended goal in one sentence, the option set per state, the intended risk/reward pairing, and the intended action-class list.

  2. Drive the run with something that is not the judge — scripted track, separate exploratory policy, or human. Record input source, seed, and length. The driver is a search instrument, never an evaluator: no driver fitness score enters any verdict.

  3. Run a coverage precheck. Inspect phases, entity-count regimes, resource levels, and difficulty band. A run that never left one regime is inadmissible-sample, not collapse.

  4. Choose the track and respect what it licenses.

    • Skilled (search-driven or expert human) — widest coverage; the only track that may carry collapse. Says little about first contact.
    • Naive (first-contact stand-in that still clears the precheck) — per-scene legibility and the entry-point comparison. Never collapse: low divergence is confounded with the driver's own repertoire.
    • Degenerate (idle, hold-only, spam) — licenses nothing; inadmissible-sample by construction.

    Both tracks on the same build, same Δ, same framing, yields the entry-point verdict.

  5. Sample, do not hand-pick. Fixed-interval sampling from a recorded run (default: a tick count covering ≈30 s of play at normal speed, recorded and expressed in ticks), or enumeration of a state space declared before results were seen. Reject scene-by-scene curation and any set re-picked after a disappointing result. N ≥ 8 scenes per run.

  6. Build each scene as a triplet. t-Δback, t, t+Δfwd. Only t-Δback and t are ever shown in stage A; t+Δfwd is the withheld oracle. Two lead-in frames are the minimum that recovers motion — displacement, hence direction and speed. Use three lead-in frames when the game's decisions turn on acceleration — gra

阅读完整来源 ↗