Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add DY-2026/GameDesignOS --skill paranoia-ai-system-evolvergit clone --depth 1 https://github.com/DY-2026/GameDesignOSWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/dy-2026/gamedesignos/paranoia-ai-system-evolver)<a href="https://agentmods.dev/skills/dy-2026/gamedesignos/paranoia-ai-system-evolver"><img src="https://agentmods.dev/badge/skills/dy-2026/gamedesignos/paranoia-ai-system-evolver/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/dy-2026/gamedesignos/paranoia-ai-system-evolver"><img src="https://agentmods.dev/badge/skills/dy-2026/gamedesignos/paranoia-ai-system-evolver.svg" alt="Reviewed on agentmods" width="80" height="20"></a>- NVIDIA SkillSpector pass
What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00147 | $0.03794 |
| Opus 5 | $0.00073 | $0.01897 |
| Sonnet 5 | $0.00029 | $0.00759 |
| Haiku 4.5 | $0.00015 | $0.00379 |
Grade A, and why
paranoia-ai-system-evolver scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 12d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 173 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Paranoia AI System Evolver
Copyright (c) 2026 Paranoia. Licensed under the MIT License.
核心立场
把 AI 系统演化当成受控系统设计,而不是神秘的自我改良;把信息获取当成决策投资,而不是越多越好的默认动作。
Intent Work Order 先把“帮我写、优化、分析、出图”的指令单翻译成“要改变什么现实、谁验收、不能牺牲什么、AI 可自治到哪里”的作战意图。
WOOP 定义任务意图、验收结果、失败模式和恢复协议。
Decision Object 定义现在到底要决定什么,以及没有新信息时会做什么。
RJR-AI 定义剩余判断权、授权边界和谁能拍板。
VOI 判断哪些信息、检索、追问、实验或 AI 分支值得付出成本。
Scenario VOI Adapter 定义不同使用场景里什么证据才真的会改变行动。
UL(Uncertainty Ladder,不确定性阶梯)控制下一轮释放多少未知,使失败可归因、复杂度逐步增加并经过迁移验证。
OODA 让 agent 用现实反馈刷新地图。
Evals 决定哪些改动值得留下。
Human Gate 防止一次有用突变污染长期系统。
Rollback 让每次提升都可逆。
VOI 的硬规则:真实、新鲜或结构清晰的信息不一定有价值。只有当合理信号可能改变行动、优先级、资源配置或停止条件时,它才具有当前决策价值。
UL 的硬规则:VOI 选最值得消除的未知,UL 控制暴露剂量。每轮默认只释放一个主要不确定性;若失败无法区分主要解释,必须标记 confounded、恢复支架并退回更受控环境,不能继续叠加永久规则。受控样本通过不等于迁移通过,其他维度通过也不自动提升权限或真实后果。
RJR-AI 的硬规则:AI 可以扩大可能性,Workflow 可以压缩混乱,Eval 可以提供反馈,权限系统可以防止越界,知识库可以积累组织记忆;但高耦合、低可逆、证据不足且必须下注的问题,属于人的剩余判断权,agent 只能辅助论证并进入 Human Gate。
何时使用
用于改动这些层:
- prompt、system instruction、memory、RAG、tool routing、workflow、schema、eval set、docs 或 Codex skill;
- agent feedback loop、trace format、release gate 与 rollback policy;
- AI 工作单、任务单、需求单、prompt brief 从“命令 AI 做动作”升级为“给 AI 一个清晰作战意图并让它在边界内循环”;
- 需要把项目整体流程、workflow run、产出质量、验收、复盘和候选规则沉淀纳入治理检查;
- 需要 model compression、causal mediator、WOOP harness protocol 或 total description cost 降低的 AI engineering 结构;
- 需要判断某次搜索、追问、读记忆、日志分析、实验或更多 AI 对话是否值得;
- 出现 FOMO、信息过载、分支爆炸、研究替代行动或高结构低价值输出时。
不要用它来合理化失控的模型权重改动、静默长期记忆写入、未经批准的全局 skill 安装,或没有 Human Gate 的生产影响行为。它也不是通用热点总结器;没有决策对象时,只允许有预算的探索或明确的信息消费。
快速流程
- 定义任务和被改动的系统层:
prompt、memory、RAG、tool routing、workflow、eval、schema、docs或skill。 - 若用户给的是指令单,先写轻量
Intent Work Order:reality_to_change:我要改变什么现实;parent_project_goal:服务哪个更大的项目目标;desired_world_state:完成后外部世界应该变成什么状态;verifier_role与first_impression_must_understand:谁验收,第一眼必须看懂什么;must_not_sacrifice、ai_can_freely_change、ai_must_not_touch:不能牺牲、可自由改和不允许碰;decision_principles_if_plan_breaks:原计划不成立时按什么原则改方向;failure_signals_to_check_before_delivery与retrospective_contract:交付前自查和复盘沉淀。
- 写轻量
WOOP Task Card:Wish / Intent Spec:目标、输出物、范围与停止条件;Outcome / Evaluation Rubric:验收标准与决策收益;Obstacle / Failure Pattern:目标漂移、过度信任、上下文污染、工具滥用、FOMO 调研、选项爆炸、虚假确定性等内在失败模式;Plan / If-Then Protocol:触发条件、判断者、恢复动作、重试、交还人或 rollback。
- 在获取更多信息前定义
Decision Object:- 决策问题、owner、deadline;
- 真实可选项;
current_default_action,即没有新信息时的行动;- stakes、reversibility 与
boundary_status: undefined | far | near | locked。
- 建立
RJR-AI授权门:- 判断 coupling:局部低耦合,还是会牵动产品、系统、账号、发布、长期规则的高耦合;
- 判断 reversibility:可逆、撤回昂贵,还是不可逆;
- 写出 delegation:AI 只能读、建议、草稿,还是可做低风险可逆执行;
- 把低风险可逆任务交给自动化,把可测试事项交给 eval,把高耦合低可逆事项交给 Human Gate;
- 若证据不足但必须下注,明确
residual_judgment,由人选择方向。
- 建立 VOI 决策门:
- 只保留会影响选项排序的不确定性;
- 每轮最多提出 3 个
candidate_information_actions候选信息行动; - 为可能信号预注册
posterior_update与action_if_seen; - 若所有信号都不会改变行动,停止调研或标记为
model_learning/information_consumption; - 用 EVPI 作为价值上界,用 EVSI 判断具体样本、实验或探针;
- 扣除获取、延迟、注意力、隐私、污染和实施风险成本;
- 选择净价值最高的最小探针,并写停止规则。
- 选择
Scenario VOI Adapter,按具体使用场景定义有效证据:skill_evolution:看真实 trace、行为 eval、负迁移、rollback,而不是一次漂亮案例;game_direction:看玩家承诺、核心循环、题材解释规则、生产风险和最小原型信号;experience_diagnosis:看 evidence_id、issue priority、修复动作和下一轮验证是否改变;source_curation:看材料是否改变入库、分类、沉淀或拒绝,而不是只看内容新鲜;content_decision:看选题、角度、标题承诺、论证主线和发布判断是否改变;platform_fact:看当前一手来源、实际平台状态、兼容策略和时效边界;high_risk_action:看是否降低不可逆错误,并默认进入 Human Gate;ai_branch_management:看分支是否改变下一探针,不能改变行动的分支应归档或关闭。
- 建立
UL (Uncertainty Ladder),为下一轮验证设计可归因的环境:- 声明目标能力与当前阶段:
UL-L0 | UL-L1 | UL-L2 | UL-L3 | UL-L4 | UL-L5; - 记录输入新颖度、上下文歧义、工具环境、协作、权限与后果、验收歧义的暴露向量;
- 写出本轮
released_this_round、held_constant、仍保留的支架和失败后果预算; - 每轮默认只释放一个主要变量,预注册 pass / fail / confounded / stop 信号;
- 失败后用消融、对照或反事实定位
primary_bottleneck,只针对主瓶颈修改; - 同层复现后才增加复杂度,最后用近/中迁移和负迁移样本限制适用范围;
- 权限、发布、资金、长期记忆和真实用户影响不随阶段自动晋升,仍走 RJR-AI / Human Gate。
- 声明目标能力与当前阶段:
- 显式写出 operating model:
- compression:什么短模型能解释多数真实案例;
- causality:哪些 mediator 把输入连接到结果;
- control points:agent、workflow 或 human 能干预哪个 mediator;
- cost:core model、routing、state、validation、exception、recovery 的成本在哪里累积。
- 维护紧凑 OODA 状态:
- Observe:目标、上下文、证据、惊讶信号、触发的 Obstacle;
- Orient:当前框架、用户模型、领域模型、决策边界、不确定性地图;
- Decide:选择动作、拒绝动作、VOI 理由与停止条件;
- Act:artifact、tool call、最小探针或 test;
- Evaluate:用 Outcome 打分,记录先验—信号—后验—行动变化。
- 分离 task OODA 和 meta OODA。任务循环完成当前工作;元循环只提出未来系统可考虑的
candidate改动。 - 每个演化改动保持
candidate,直到证据、行为 eval、必要审批和 rollback 都存在。 - 当目标层是
skill,回放代表性任务,检查是否减少低 VOI 分支、是否保留具体负反馈、是否出现更啰嗦、更慢或误触发的负迁移;同时检查受控通过是否能迁移到陌生样本。 - 满足任一条件即停止继续获取信息:行动对合理信号已稳健、边际 VOI 不高于边际成本、样本门达到、deadline 到达、剩余不确定性不改变行动,或 Human Gate 已承诺执行。
What ships with it
57 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
- agents/openai.yaml 1.4 KB
- evals/behavior_evals.json 366 B
- evals/evals.json 9.4 KB
- evals/negative_cases.md 5.3 KB
- evals/rubric.yaml 6.3 KB
- evals/synthetic_outputs.json 6.1 KB
- evals/uncertainty-ladder-cases.en.md 1.7 KB
- evals/uncertainty-ladder-cases.md 2.7 KB
- evals/voi-decision-gate-cases.en.md 6.6 KB
- evals/voi-decision-gate-cases.md 9.6 KB
- examples/ul-state.example.json 2.8 KB
- quick_validate.py 18 KB runs code
- README.en.md 6.9 KB
- README.md 6.9 KB
- README.zh-CN.md 6.9 KB
- references/eval-versioning-playbook.en.md 5.8 KB
- references/eval-versioning-playbook.md 396 B
- references/eval-versioning-playbook.zh-CN.md 6.0 KB
- references/evolution-loop-playbook.en.md 8.9 KB
- references/evolution-loop-playbook.md 919 B
- references/evolution-loop-playbook.zh-CN.md 11 KB
- references/intent-engineering-work-order.en.md 1.5 KB
- references/intent-engineering-work-order.md 228 B
- references/intent-engineering-work-order.zh-CN.md 7.4 KB
- references/model-compression-playbook.en.md 7.9 KB
- references/model-compression-playbook.md 1.7 KB
- references/model-compression-playbook.zh-CN.md 7.4 KB
- references/project-workflow-governance.en.md 1.5 KB
- references/project-workflow-governance.md 1.4 KB
- references/project-workflow-governance.zh-CN.md 5.3 KB
- references/uncertainty-ladder-protocol.en.md 4.0 KB
- references/uncertainty-ladder-protocol.md 867 B
- references/uncertainty-ladder-protocol.zh-CN.md 8.9 KB
- references/value-of-information-playbook.en.md 16 KB
- references/value-of-information-playbook.md 431 B
- references/value-of-information-playbook.zh-CN.md 17 KB
- references/woop-harness-protocol.en.md 5.0 KB
- references/woop-harness-protocol.md 206 B
- references/woop-harness-protocol.zh-CN.md 6.3 KB
- templates/evolution_proposal.en.md 5.2 KB
- templates/evolution_proposal.md 5.1 KB
- templates/evolution_proposal.zh-CN.md 5.1 KB
- templates/intent_work_order.en.md 1.4 KB
- templates/intent_work_order.md 2.5 KB
- templates/intent_work_order.zh-CN.md 2.6 KB
- templates/ooda_voi_state.en.md 3.1 KB
- templates/ooda_voi_state.md 3.1 KB
- templates/ooda_voi_state.zh-CN.md 3.1 KB
- templates/uncertainty_ladder_state.en.md 1.4 KB
- templates/uncertainty_ladder_state.md 552 B
- templates/uncertainty_ladder_state.zh-CN.md 1.5 KB
- templates/voi_decision_gate.en.md 2.7 KB
- templates/voi_decision_gate.md 2.7 KB
- templates/voi_decision_gate.zh-CN.md 2.7 KB
- templates/workflow_governance_review.en.md 1.2 KB
- templates/workflow_governance_review.md 1.2 KB
- templates/workflow_governance_review.zh-CN.md 2.1 KB
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 12d ago First seen · 173 lines · 147 tokens per session scan A 6d8f9bc33fe3
paranoia-ai-system-evolver is a skill published in the GitHub repository DY-2026/GameDesignOS (387 stars, last pushed 25d ago), licensed MIT. It adds 147 tokens to every session and 3,794 once invoked, about $0.0007 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other skills, from other repositories
dspy
DSPy: declarative LM programs, auto-optimize prompts, RAG.
ai-engineering-toolkit
6 production-ready AI engineering workflows: prompt evaluation (8-dimension scoring), context budget planning, RAG pipeline design, agent security audit (65-point checklist), eval harness building, and product sense coaching.
llm-app-patterns
Production-ready patterns for building LLM applications. Covers RAG pipelines, agent architectures, prompt IDEs, and LLMOps monitoring. Use when designing AI applications, implementing RAG, building agents, or setting up LLM observability.
llm-security
Use for authorized security assessment of LLM applications and AI agents, including prompt injection, tool abuse, RAG exposure, memory poisoning, and model supply-chain risks.
llm-ops
LLM Operations -- RAG, embeddings, vector databases, fine-tuning, prompt engineering avancado, custos de LLM, evals de qualidade e arquiteturas de IA para producao.
ai-product
Every product will be AI-powered. The question is whether you'll build it right or ship a demo that falls apart in production.