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。
直接命令不会经过审查 Prompt;运行前请先检查来源。
npx skills add https://github.com/nWave-ai/nWave-experimental --skill nw-cross-cutting-context-residency命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地?可下载 SkillsMP 当前能够提供的文件。
基于 SOC 职业分类
正在显示 SKILL.md
| name | nw-cross-cutting-context-residency |
| description | KNOWLEDGE — use when choosing resident, map-only, or on-demand context. |
| user-invocable | false |
| disable-model-invocation | true |
Composition: Load with nw-cross-cutting-invariants; no forced sequence; return to its routing table for other concerns.
context:pay-on-demand-not-every-session — resident context is paid every session (STANDING)Whoever needs it loads it; whoever does not need it must not pay for it in every session. Context injected at session start is a per-session tax on every reader, including the readers who will never touch that subject. Detail belongs where it is needed, at the moment it is needed — not resident.
Three residency classes; classify before you inject anything:
Density is a second axis, independent of residency. Normative content cannot be GENERATED from code, but it can be COMPRESSED: doctrine is usually rules — subject · condition · action — written as narrative. A table says the same thing in a fraction of the bytes and reads better. Do not confuse "not derivable" with "not compressible"; measuring the first tells you nothing about the second.
The failure this clause prevents: a projection that ADDS instead of REPLACING. Retiring a duplicate is a saving; generating a new view beside a surviving hand-written one is a new divergence surface AND a bigger payload.
context:resident-doctrine-can-be-compensation-debt — resident prose that patches bad guidance is debt (STANDING)When a body of resident context exists to compensate for guidance that fails to arrive at the point of need, that context is not doctrine — it is debt, and it is the SECOND cost of a defect whose first cost is already being paid at the gate.
The chain: a gate refuses reactively instead of guiding at the authoring surface (the GDP-2 emission corollary) → readers keep getting caught → someone compensates by injecting standing prose into every session → now everyone pays, every session, for a message that should have arrived once, in place, to the one person who needed it.
So, before adding anything to a resident payload, ask: is the real gap a MESSAGE that fails to guide? If yes, fix the message; the resident need dissolves rather than being served. And the converse is the practical lever: as guidance moves to the point of need, resident doctrine can thin without losing anything — the two workstreams compose, and the thinning is only safe in that order.
Do not thin first. Removing resident prose while the corresponding guidance still arrives too late removes the compensation and keeps the defect.