Context Engineering

Context Engineering

Principle expression

Primary: P09 Supporting: P02, P14, P08

Scope

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.

An Agent's larger working environment also contains a task form, workspace, tools, capabilities, effect limits, verification, and a return contract. This Skill owns only the selection, timing, and receiver-facing expression of source information inside that environment; it does not bind the other parts or become their owner.

Do not assume that a familiar filename, three-layer model, hook, skill directory, or memory mechanism exists in another runtime. Investigate the actual target.

Principle source

Use a host Sequence and matching interpretations when the host declares them. Otherwise use this skill's read-only fallback in references/sequence.md. Read only P09, P02, P14, and P08.

Start

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.

Dispatch

• With design, read and follow commands/design.md.
• With audit, read and follow commands/audit.md.
• With verify, read and follow commands/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.

Core method

1. 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.
2. 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.
3. 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.
4. 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. Express the projection for its receiver: omit parent history and runtime maps that do not affect the named action. If any necessary term could change the receiver's action or return, define it at first use by the object it denotes, its boundary, and why it matters now. For a blind-spot pass, deliver the evidence that the ordinary source view omits: the actual user action, observed friction, screenshots or rendered state, stale/reconnect condition, and the source or transcript that grounds it. Do not replace this evidence with a generic instruction to be creative; the delivery path must make the missing judgment possible.
5. Preserve ownership boundaries. Route source content and acceptance to the domain owner, source/lifetime conflicts to artifact-organization, skill expression failures to skill-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.
6. 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. When prompt comprehension is the suspected failure, a fresh Agent may first restate the object, boundary, allowed effects, non-goals, verification, and return from the delivered context alone; a request to define an undeclared term defeats that delivery claim. Keep successful restatement separate from semantic task correctness and acceptance. When a user correction reveals a missed point, verify the revised delivery by asking whether a fresh Agent can name the object, the missed signal, its boundary, and the smallest next probe from the delivered context alone. This tests post-training context delivery; it does not claim that the Agent has acquired the human's tacit experience.
7. 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.

Domain vocabulary

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.

Completion standard

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.