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。
浏览器与自动化

evaluating-gameplay-balance

Evaluates and improves gameplay balance from telemetry in any engine. Use when comparing monotonous vs exploratory play, diagnosing death/spawn/scoring/input issues, setting difficulty, timer, or pacing targets with human-limited simulated players, or proposing structural balance fixes instead of numeric tuning.

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 evaluating-gameplay-balance

技能说明


name: evaluating-gameplay-balance description: "Evaluates and improves gameplay balance from telemetry in any engine. Use when comparing monotonous vs exploratory play, diagnosing death/spawn/scoring/input issues, setting difficulty, timer, or pacing targets with human-limited simulated players, or proposing structural balance fixes instead of numeric tuning."

Use this skill to analyze whether a game rewards skillful play and to propose structural improvements.

Engine-neutral contract:

  • Produce comparable runs for monotonous policies and exploratory policies.
  • Define the public input schema and visible-state features available to policies.
  • Keep seeds, tick rate, max duration, policy definitions, search budget, and aggregation logic comparable across runs.
  • Record score, elapsed time, end state, and telemetry for death, spawn, scoring, and input behavior.
  • Compute exploratory_ratio = exploratory.best.score / monotonous.max_score.
  • Treat the ratio as a quality detector, not an optimization target.
  • Declare an execution profile for every non-baseline policy as well as its visible-state features: oracle (hidden information allowed; upper bound), precise (visible state, near-perfect timing and attention), or human-limited (visible state plus the human limits in references/simulation-harness.md). Compute the ratio from precise policies (search-based runners included, unless labeled oracle); it detects exploits, not difficulty.
  • Set difficulty, timers, pacing, and progression targets only from human-limited runs, and label every conclusion with the profile it rests on. Without a human-limited run, do not state a difficulty or pacing verdict — report it as not yet measured.
  • A claim that a strategy dominates, or that a trade-off is balanced, names the profiles it was tested on. If the ranking flips between profiles, that alone is not a balance defect — report it as skill-dependent, but still apply the experience guardrails to each profile's best strategy (a human-limited optimum of idling or waiting is a defect).

Non-deterministic harnesses:

The contract above assumes a seeded forward model. Some rigs cannot provide one — a real browser driven over a wall-clock window, an engine whose RNG you do not control, a policy comparison run against live rendering. Seeding is still the first choice; when it is genuinely unavailable, the following replaces it, and skipping it produces confident conclusions from noise:

  • Report a min–max band over n ≥ 3 runs per policy, never a point estimate. A single run of a long-tailed distribution proves nothing about either policy.
  • To claim a band has moved, require n ≥ 5 and no overlap with the recorded prior band. A sample that overlaps the previous band is not a finding, however plausible the mechanism that would explain it.
  • Record the accepted band and its known tails in the project's spec, so later runs compare against history instead of rediscovering the same spread. If a policy is known to land near a bad-looking value in some fraction of runs (e.g. "~20 % of idle runs reach 0.18 of skilled score, by design, because the field can collapse on its own"), that tail is documented and settled — not re-litigated every time it appears.
  • Retraction is a valid and expected output. A conclusion published from a small sample that a larger sample does not support should be withdrawn in place: keep the mechanism if it was real, and state that its effect size was indistinguishable from noise. Observed case: a band shift published at n=4, with a coherent causal story attached, was retracted at n=6 as the same spread the project had already recorded twice.

Experience guardrails:

  • Reject changes that degrade play experience even if metrics improve.
  • Score only in-game causal events; do not award points for raw input facts.
  • Game-over should be tied to hazards or world-state collapse.
  • Do not add hidden behavior that only helps or hurts test agents.
  • Avoid numeric-only tuning, branch-only fixes, and added randomness as the primary answer.
  • Treat hands-on play reports as calibration data for the human-limited profile. When a report conflicts with simulated results, first check the harness for defects, then adjust the profile toward the report, re-run, and record the adjustment; do not dismiss the report with bot data alone.
  • Do not tune a parameter with a policy whose behavior is defined by that parameter (e.g. "wait while the timer is above 80 %"); the measurement moves with the tuning.

Workflow:

  1. Locate an existing simulation harness. If none exists, design one using the harness contract before judging balance.
  2. Choose the comparison protocol:
    • If seeding is available, run comparable monotonous and exploratory policies with the same deterministic seeds.
    • If seeding is genuinely unavailable, use the non-deterministic band protocol above and keep run count, wall-clock window, policy definitions, and sampling cadence comparable.
  3. Validate that the report includes run configuration plus death, spawn, scoring, and input telemetry.
  4. If telemetry is incomplete, report the gap and request instrumentation/rerun instead of judging balance from score alone.
  5. Sanity-check non-trivial policies before analyzing the game (see Simulated-player sanity checks in references/simulation-harness.md). A failed check is a harness defect: fix the policy, re-run, and report the defect instead of drawing conclusions from it.
  6. Analyze death, spawn, scoring, and input patterns.
  7. Identify root causes in rules or generation logic.
  8. Propose at least three candidate fixes with expected impact, risk, and complexity.
  9. Re-test with the same policies, budgets, aggregation, and comparison protocol after implementation. Preserve the same seeds for deterministic runs; preserve the same run count and sampling setup for non-deterministic runs.

Project checker triage, when the report includes ratio.diagnostic:

  • Bef
阅读完整来源 ↗