源文件边界
本页为人类读者重排
skills/context-engineering/SKILL.md中的关键判断，不替代可安装的完整 prompt；事实、修改权和未展示的运行说明仍属于源文件。
安装与日常使用
Section titled “安装与日常使用”npx skills add lidessen/rossovia --skill context-engineering安装后直接向 agent 描述目标；只有在需要强制选择时才显式点名 skill 或内部 operation。
它负责的判断
Section titled “它负责的判断”Own one judgment: for a named agent action, which source information must reach the agent, when must it become available, and what is the smallest runtime-evidenced delivery path that preserves source authority?
Context engineering owns delivery, not content. A project source, accepted design, or domain skill supplies meaning; this skill decides how the selected agent can discover the relevant part at the moment it changes action. Loaded instructions, summaries, indexes, and retrieved fragments remain projections unless separately governed as sources.
Do not assume that a familiar filename, three-layer model, hook, skill directory, or memory mechanism exists in another runtime. Investigate the actual target.
Primary: P09 Supporting: P02, P14, P08
- Name the action and source. Identify the concrete agent judgment that is failing and the authoritative project source that could change it. Separate missing content from failed delivery, and source facts from projections.
- Map the actual runtime. Establish which instruction surfaces, skill metadata, on-demand resources, retrieval paths, and context-injection mechanisms this agent can really use. Mark documented capability, observed task loading, and unknown behavior separately. Use delivery capability map.
- Choose delivery timing. Decide whether the information must be available before scoped action, when a method or task is activated, or only when a concrete detail becomes relevant. These are timing questions, not mandatory L1/L2/L3 containers. Map them to the runtime’s native surfaces.
- Select the smallest path. Prefer a discoverable pointer to a named source over copied doctrine. Compress only what the target decision can safely lose. Keep hard constraints reachable at the moment they govern action; remove detail that does not change that action.
- Preserve ownership boundaries. Route source content and acceptance to the
domain owner, source/lifetime conflicts to
artifact-organization, skill expression failures toskill-engineering, and vendor-specific mechanics to a current project adapter. This skill may recommend a delivery projection; it does not become the source or runtime. - Verify delivery and use. Run a real task through the ordinary entry surface, observe the delivery path where possible, and check the named action. Use a boundary task to expose needless loading. Add a controlled baseline only when claiming comparative improvement, not merely to prove delivery.
- Externalize proportionally. Return a conversation result for a bounded audit. Retain a delivery map or evaluation record only when a later actor must reproduce, approve, or revise the placement decision.
交给 Agent 使用
Section titled “交给 Agent 使用”Target agent, runtime, and decision:Observed context failure or placement question:Authoritative source and status:Current delivery path and evidence:Moment the information must affect action:Smallest suspected change:Observation that would disconfirm it:If the information itself is missing, ambiguous, or unaccepted, route to its domain owner. Delivery cannot repair absent doctrine or decide what is true.
- With
design, read and followcommands/design.md. - With
audit, read and followcommands/audit.md. - With
verify, read and followcommands/verify.md. - With no argument, inspect the actual failure and choose the smallest matching path. Default to read-only audit when a change was not requested.
Read concepts when context, source, delivery projection, or delivery timing are being conflated. Do not turn its timing
model into a required directory or file hierarchy.
The result is ready when it names the target action, authoritative source, actual runtime capabilities, selected delivery timing and path, source boundary, and disconfirming observation. Claim delivery only from task-specific evidence; claim improvement only from a comparable evaluation. Unknown runtime behavior remains explicit rather than being filled by convention.
阅读完整技能
Section titled “阅读完整技能”本页有意省略 prompt 的内部装配说明。需要安装、审查或修改这个技能时，请回到 完整源文件。