菜单
Use when refining an existing skill document so it better supports capable current and future models. Especially useful when a skill has become verbose, procedural, template-heavy, example-heavy, or overfit to current model weaknesses.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
Do not use unless the user explicitly requests it. Dispatch each requested task to a new Codex session in a newly allocated dedicated worktree, then coordinate rebasing and fast-forward integration into local main without implementing, reviewing, or archiving the task in the coordinator session.
Use only when explicitly requested by the user or, during an explicit debugging session, after other relevant checks such as typecheck, lint, and tests have been exhausted.
Use only when the user explicitly requests this skill.
Use when building greenfield React SPA UI with Lisse (`@lisse/react` or `@lisse/core`) for Figma-style smooth/squircle corners, choosing `SmoothCorners` vs `useSmoothCorners`, configuring corner smoothing, borders, gradient borders, shadows, styling hooks, auto-effects, client-side path generation, performance, or limitations.
Use React Aria and React Stately hooks to build accessible custom React components, including choosing the right hook, wiring refs/state/props, preserving ARIA and keyboard behavior, excluding React Aria i18n and SSR utilities, and checking hook-specific guidance in bundled category docs.
Design, review, and optimize PostgreSQL and SQLite schemas, queries, migrations, transactions, indexes, and data-access code, especially when SQL should remain portable between both engines.
| name | model-forward-skill-refinement |
| description | Use when refining an existing skill document so it better supports capable current and future models. Especially useful when a skill has become verbose, procedural, template-heavy, example-heavy, or overfit to current model weaknesses. |
| disable-model-invocation | true |
Refine skill documents so they define durable behavior, not temporary scaffolding.
A model-forward skill trusts the agent to understand ordinary requirements, infer local context, and exercise judgment. It encodes only the constraints, protocols, artifacts, and failure modes that should remain important as models improve.
A skill should define what must be true, not teach the model how to think.
Prefer:
Avoid:
Keep an instruction only when it materially reduces the risk of incorrect behavior, ambiguity, scope drift, unsafe action, poor reviewability, or coordination failure.
Remove or compress instructions that are merely true, inferable, repeated, procedural, speculative, or useful only for weaker current models.
A shorter skill is not automatically better. The target is durable clarity: the fewest rules that preserve the intended behavior.
Prefer constraints over procedures.
Better:
Do not edit the approved task contract after implementation begins unless the supervisor changes requirements.
Worse:
After each step, check whether the task contract should be updated, then decide whether to ask...
Prefer judgment tests over inventories.
Better:
Treat ambiguity as high-risk when a wrong assumption could materially alter behavior, scope, safety, compatibility, reviewability, or rollback.
Worse:
High-risk ambiguity includes API behavior, UI behavior, CLI behavior, schema behavior, database behavior...
Prefer artifact semantics over templates.
Better:
The file is a locked pre-implementation contract, not a log.
Worse:
Include sections for assumptions, open questions, verification, progress, and discoveries.
Prefer communication protocols only when they improve coordination.
Useful protocols make review, approval, correction, or reference easier. Avoid requiring the agent to prove routine compliance or restate artifacts unnecessarily.
Before finalizing a refined skill, ask:
Keep the sentence only when the answer justifies its cost.
Treat this as a menu, not a required structure:
---
name: <skill-name>
description: <when to use this skill>
---
# <Skill Title>
<Purpose statement>
## Core Rules
<Non-negotiable invariants>
## Standards
<Durable judgment principles>
## Required Protocols
<Human-agent, artifact, or tool protocols that must be followed>
## Artifact Shape
<Only when the skill creates or modifies a durable artifact>
## Omit / Avoid
<Only when negative constraints prevent common failure modes>
Omit any section that does not serve the skill.
A model-forward refinement succeeds when the skill becomes easier for a capable agent to obey without losing the behavior that matters.
The result should trust ordinary reasoning while preserving the boundaries, gates, and protocols that prevent predictable failure.