modelchk

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Size the run before you spend it: how strong a model, and how hard it should think.

modelchk

is read-only and advisory. From one assessment it sizes the two dials that set a model's per-run cognitive spend — capability tier and reasoning effort. It does not choose, route, switch, pin, spawn, set, or require any concrete model or level.

在投入成本前预估运行规格：确定模型的能力强度，以及它需要进行的推理深度。

modelchk

是只读且仅提供建议的工具。通过一次评估，它可以确定设置模型单次运行认知成本的两个维度——能力层级和推理强度。它不会选择、路由、切换、固定、生成、设置或要求任何具体的模型或等级。

Goal

目标

From a single risk-and-complexity read, recommend two coordinates.

Capability tier — the cheapest sufficient class of mind:

```
fast
```
for local, mechanical, reversible work with cheap, complete verification.
```
standard
```
for ordinary repo-grounded reasoning, multi-step drafting, normal coding, and conventional documentation or skill work.
```
frontier
```
for architecture, high ambiguity, safety/security/privacy/data-loss risk, release-critical review, cross-domain scope, or work where one wrong assumption wastes a large run.

Reasoning effort — how hard that mind should deliberate.

modelchk

recommends the effort intent; resolving it to the active model's actual level — like choosing the model itself — is the executor's step, not this skill's. From least to most deliberation:

```
glance
```
— minimal deliberation; take the direct path. (Resolves to the model's floor.)
```
measured
```
— ordinary, everyday deliberation. (The model's default, or the middle of its ladder when no default is named.)
```
thorough
```
— deliberate extra: work the alternatives and check the assumptions. (Above the everyday setting, short of the top.)
```
exhaustive
```
— maximal deliberation; exhaust the search and re-check the work. (The model's ceiling.)

The two axes are orthogonal — a bounded-but-fiddly task can be

fast

thorough

, a quick expert call

frontier

glance

— yet in most work they move together, parting only when a cheap task needs hard thinking or a strong model needs only a quick call. Effort buys deliberation, never capability, and more of it is not more correct.

通过一次风险与复杂度评估，给出两个维度的建议。

能力层级——成本最低且足够的模型类别：

```
fast
```
（快速）：适用于本地、机械性、可逆且可低成本完全验证的工作。
```
standard
```
（标准）：适用于基于代码库的常规推理、多步骤起草、普通编码，以及常规文档撰写或技能类工作。
```
frontier
```
（前沿）：适用于架构设计、高度模糊性任务、安全/隐私/数据丢失风险场景、发布关键审查、跨领域范围工作，或一旦假设错误会浪费大量算力的任务。

推理强度——模型需要进行的思考深度。

modelchk

仅建议推理的意图；将其转化为实际模型的具体等级（比如选择模型本身）是执行者的工作，而非本工具的职责。从最低到最高推理程度：

```
glance
```
（粗略）：最小程度的思考；采用直接路径。（对应模型的最低等级。）
```
measured
```
（适度）：日常常规思考。（模型的默认等级，或当无默认等级时对应等级阶梯的中间值。）
```
thorough
```
（深入）：额外思考：考虑替代方案并检查假设。（高于日常设置，但未达最高等级。）
```
exhaustive
```
（穷尽）：最大程度的思考；穷尽所有搜索并重新检查工作。（模型的最高等级。）

这两个维度是相互独立的——一个有边界但繁琐的任务可以是

fast

thorough

，一个快速的专家咨询可以是

frontier

glance

——但在大多数工作中它们是同步的，仅当低成本任务需要深度思考，或强模型只需快速响应时才会分离。推理强度决定思考深度，而非能力，更高的强度并不意味着更正确的结果。

Workflow

工作流程

Frame the exact work unit being sized: task, artifact, review, rerun, or plan.
Score risk and complexity once — this single read feeds both coordinates:
- file, module, or ownership boundary crossing;
- reversibility and blast radius;
- safety, security, privacy, publishing, or data-loss risk;
- novelty, ambiguity, and long-context synthesis load;
- need for external research, adversarial review, or careful release sequencing;
- cost of a wrong answer.
Read off the capability tier: the cheapest class whose ceiling covers the work's judgment and risk.
Read off the reasoning effort: default it to track the tier (
```
fast
```
→
```
glance
```
,
```
standard
```
→
```
measured
```
,
```
frontier
```
→
```
thorough
```
, reserving
```
exhaustive
```
for the hardest, highest-stakes work), then deviate where deliberation-hunger and capability-need part — raise it for ambiguity, long multi-step reasoning, or adversarial self-check on an otherwise cheap task; lower it for a bounded task under a strong model.
Report both coordinates, one shared rationale,
```
move up if...
```
and
```
move down if...
```
triggers for each dial, and the proof surface — the verification the work still needs regardless of tier or effort.
Stop.

明确要预估的具体工作单元：任务、工件、审查、重运行或计划。
一次性评估风险与复杂度——这一评估结果将用于两个维度的确定：
- 是否跨越文件、模块或所有权边界；
- 可逆性和影响范围；
- 安全、隐私、发布或数据丢失风险；
- 新颖性、模糊性和长上下文综合负载；
- 是否需要外部研究、对抗性审查或谨慎的发布顺序；
- 错误答案的成本。
确定能力层级：选择能力上限能覆盖任务判断与风险的成本最低类别。
确定推理强度：默认与能力层级匹配（
```
fast
```
→
```
glance
```
，
```
standard
```
→
```
measured
```
，
```
frontier
```
→
```
thorough
```
，
```
exhaustive
```
仅保留给最困难、最高风险的工作），然后根据对思考的需求和能力需求的差异进行调整——如果任务模糊、需要多步骤推理，或在低成本任务中需要对抗性自我检查，则提高推理强度；如果任务有明确边界且使用强模型，则降低推理强度。
报告两个维度的建议、一个共同的理由、每个维度的「升级触发条件」和「降级触发条件」，以及验证依据——无论层级或强度如何，任务仍需进行的验证工作。
结束流程。

Rules

规则

Size the run's cognitive spend, nothing else. The two dials are which mind (capability tier) and how hard it thinks (reasoning effort). Orchestration dials — context budget, fan-out width, tool-permission scope — are a different reflex and stay out; sharing this one risk read does not pull them in.
Neutral language only: tier is
```
fast
```
/
```
standard
```
/
```
frontier
```
; effort is
```
glance
```
/
```
measured
```
/
```
thorough
```
/
```
exhaustive
```
, each an intent defined by a position on the active model's ladder — floor, default, above-default, ceiling — never a named vendor level.
When resolved, effort binds to positions, not levels: a model lacking an interior level collapses the rung to the nearest it offers, so the intent always maps to a real setting and never resolves out of range.
The default is the cheapest sufficient tier and the effort that meets the work, not the strongest of either.
Effort is deliberation budget, not capability and not answer length. Never raise it to buy capability — that is the tier's job.
No routing authority. This skill recommends two coordinates; the user, harness, or executor decides what runs and sets the actual level.
Do not name concrete model products, vendors, or versions in durable mechanism text.
Risk beats size: a one-file high-risk change can want
```
frontier
```
; a broad mechanical rename can stay
```
fast
```
+
```
thorough
```
when verification is complete.
Verification is separate. A stronger tier or a higher effort never replaces tests, review, command output, manual QA, or other proof surface.
Do not execute the task being sized, change configuration, call another model, or alter provider settings.
Do not override an explicit user tier or effort choice. Report the mismatch if one is visible.

仅预估运行的认知成本，不涉及其他内容。两个维度是「使用哪类模型」（能力层级）和「思考深度」（推理强度）。编排维度——上下文预算、分支宽度、工具权限范围——属于其他范畴，不在此工具的处理范围内；共享本次风险评估结果并不涉及这些维度。
仅使用中立语言：层级为
```
fast
```
/
```
standard
```
/
```
frontier
```
；强度为
```
glance
```
/
```
measured
```
/
```
thorough
```
/
```
exhaustive
```
，每个术语代表一种意图，对应实际模型等级阶梯上的位置——最低、默认、高于默认、最高——而非特定厂商的等级。
当转化为实际设置时，推理强度绑定的是位置而非具体等级：如果模型没有中间等级，则映射到最近的可用等级，确保意图始终对应实际存在的设置，不会超出范围。
默认选择成本最低且足够的层级，以及满足任务需求的推理强度，而非最强的层级或最高的强度。
推理强度是思考预算，而非能力或答案长度。绝不能通过提高强度来获取能力——这是能力层级的职责。
无路由权限。本工具仅提供两个维度的建议；用户、管理工具或执行者决定运行内容并设置实际等级。
在持久机制文本中不得提及具体的模型产品、厂商或版本。
风险优先于规模：单个文件的高风险变更可能需要
```
frontier
```
层级；范围广泛的机械性重命名在验证完成后可以保持
```
fast
```
+
```
thorough
```
。
验证是独立的。更强的层级或更高的强度永远无法替代测试、审查、命令输出、人工QA或其他验证依据。
不得执行待预估的任务、更改配置、调用其他模型或修改提供商设置。
不得覆盖用户明确指定的层级或强度选择。如果存在不匹配情况，需报告该差异。

Output

输出

text

recommended_tier: fast|standard|frontier
recommended_effort: glance|measured|thorough|exhaustive
rationale: <one sentence, covering both dials>
move_up_if: <signals that would justify a stronger tier or higher effort>
move_down_if: <signals that would justify a cheaper tier or lower effort>
proof_surface: <verification still required, independent of tier and effort>

text

recommended_tier: fast|standard|frontier
recommended_effort: glance|measured|thorough|exhaustive
rationale: <一句话，涵盖两个维度>
move_up_if: <证明需要更强层级或更高强度的信号>
move_down_if: <证明需要更低成本层级或更低强度的信号>
proof_surface: <无论层级和强度如何，仍需进行的验证工作>

Verification

验证

Before finishing, confirm the report:

names exactly one tier (

fast

standard

frontier

) and one effort (

glance

measured

thorough

exhaustive

);

states the cheapest sufficient tier and the effort that meets the work, not the strongest of either;
gives move-up and move-down triggers covering both dials;
names the proof surface;
makes no routing, switching, provider, vendor, product, version, or concrete-level claim, and names no orchestration dial beyond the two cognitive-spend ones.

完成报告前，确认报告：

明确指定一个层级（
```
fast
```
/
```
standard
```
/
```
frontier
```
）和一个强度（
```
glance
```
/
```
measured
```
/
```
thorough
```
/
```
exhaustive
```
）；
选择的是成本最低且足够的层级，以及满足任务需求的推理强度，而非最强的层级或最高的强度；
给出涵盖两个维度的升级和降级触发条件；
指定验证依据；
未提及路由、切换、提供商、厂商、产品、版本或具体等级的声明，且未涉及除两个认知成本维度外的编排维度。