Context Engineering and Skills

Deciding what the model sees

Context engineering is the discipline of deciding what goes into the model's context window: instructions, examples, retrieved data, and memory, in what order and at what length. It has replaced one-off prompt writing as the core skill of the field. In practice: cutting an irrelevant document from context often improves both cost and accuracy at the same time.

The System prompt is the standing instruction block the model receives before any user input: role, rules, format requirements. It is resent, and rebilled, on every call. In practice: a bloated system prompt quietly multiplies the cost of every single run.

Retrieval fetches relevant documents at request time and places them in context, so the model can answer from data it was never trained on. In practice: a support bot that quotes your current help-center articles is using retrieval.

From prompts to structured capability files

A prompt is disposable context: written once, pasted somewhere, drifting immediately. A structured capability file declares the same intent durably: inputs, outputs, steps, and routing in a format a machine can validate. Portability is the payoff: define the capability once and run it on any compliant runtime without a rewrite. In practice: a portable skill moves between agent platforms the way a container image moves between clouds.

Measure skill activity without mixing evidence

Skill usage can describe different things, so keep its counts separate:

• Declared counts saved workflow nodes that reference a skill. It describes configuration, not execution.
• Estimated activity applies graph reach probability and expected visits to recorded workflow runs. It is an estimate, not a runtime trace.
• Reported executions come from a runtime or trusted integration and are not combined with estimates.
• Retrievals count successful approved-skill fetches; they are not execution events.

The Skill Usage view keeps these evidence types distinct while showing workflow impact, revision spread, and prioritized recommendations. Usage events contain metadata rather than skill content or prompts. Unlinked workflow references stay unlinked rather than being guessed from a similar name.

The Cromus Lens

A SKILL.md file is context engineering made durable and portable. The SOP Compiler turns prose procedures into a structured capability contract with inputs, outputs, and steps. The open SKILL.md, ETHOS.md, MEMORY.md, and GRAPH.md specs are the contract between authored context and any compliant runtime. Skill Usage adds a way to examine where skills are referenced and how activity is estimated or reported, without treating estimates as runtime traces. The contrast in one line: a prompt is disposable context; a skill is governed, versioned, scored context.

The panel below is the same readability pre-check used elsewhere in Cromus, framed here as context quality in 10 seconds: paste a procedure and see whether its structure is ready to become a skill.