A skill may contain cognition

Engineers sometimes treat skills as "the non-AI part": deterministic glue between model calls. That habit pushes every judgment into either a mega-prompt or a faux agent. The middle layer exists precisely so known procedures can orchestrate cognition responsibly.

A skill may call five cognitive units, branch on their results, retry a step, and write to the world. It remains one skill when the capability was known before the run. It stops being honest architecture when several open judgments were merged into one fake edge because the prompt grew long.

Helpful context: What is a skill defines the perform responsibility. What is a cognitive unit owns one judgment per call. Not a composite cognitive units blocks multi-judgment mega-edges.

Cognition inside skills is expected

Cognition inside skills is expected, not exceptional. When a skill diagram has no cognitive unit calls, the question is whether judgment was smuggled into prompts, hidden inside code comments, or falsely promoted to an agent loop.

Invoice field interpretation cognitive units run sequentially inside an extraction skill. Each call owns one wrong-answerable reading. The skill owns ordering, validation gates, persistence seams, and recovery when a field abstains. Meal composition calls GenerateCandidates, ComposeWeek, and critique-related cognitive units before returning a plan artifact. Procedure-level evaluation covers seams between those calls alongside leaf scores on each template.

The skill boundary is responsibility, not absence of models. Perform work may be rich with judgment and still contain no goal pursuit. A skill that calls five cognitive units in sequence, branching on abstention at each step, is architecturally richer than many agent demos, yet it remains perform work because the overall procedure was declared before the run and no step must choose whether the goal has been sufficiently achieved.

Invoice extraction skill

Skill: Extract structured invoicecognitive unit: Interpret headercognitive unit: Match line itemscognitive unit: Judge materialityCode: Validate + persist

Many cognitive units calls. One perform responsibility.

Skill versus composite cognitive units

The distinction is not "one call versus many calls." It is "one judgment versus many judgments." If each step is a separable open judgment, the steps belong as cognitive units inside the skill, named and suite-bound, with substitution available from libraries.

Declaring "ExtractInvoice" as one cognitive unit because the prompt is long hides cost, ownership, substitution, and failure locality. Operators cannot tell which reading failed when materiality was fine but line matching was wrong. If the path must still be discovered case by case and evidence can change what work happens next, the territory is agent, not skill packaging.

A skill composes judgments. It does not merge them into a larger judgment.

The cognitive unit · Ch. 16

Meal Companion and invoice examples

The Compose Weekly Plan skill calls domain-library cognitive units and deterministic validators. GenerateCandidates proposes options under constraints. ComposeWeek assembles a coherent provisional week. AssessMealPracticality may appear inside candidate filtering. Deterministic code enforces allergy rules the models cannot override. The Meal Planning Agent owns acceptance of the full week. cognitive units answer point questions. The skill performs a packaged capability. The agent pursues the goal and chooses transitions after critique. When critique finds that Tuesday is impractical while variety is fine, local repair does not require regenerating the entire week through one composite edge. Named cognitive units and a skill boundary make that repair architecturally possible.

The cognitive unit treats invoice extraction the same way. Header interpretation, line matching, discrepancy reading, and materiality judgment are separate cognitive units with separate failure modes. The skill performs extraction as a known capability: retrieve sources, invoke interpreters in order, calculate variance, judge materiality, validate, persist. Teams sometimes resist the split because "the model can do all of that in one shot." One shot is a demo strategy, not a responsibility map. Production systems need pointable failure, substitution, and cost attribution.

Procedure evaluation and failure modes

Leaf cognitive unit suites answer whether each judgment performed at its published standard. Skill-level evaluation answers whether the capability performed correctly across seams. Did validation run after composition every time? Did persistence wait for deterministic approval gates? Did recovery trigger when a cognitive unit abstained? Did the skill respect authority grants on acting steps? A skill can fail seam tests while every cognitive unit passes. That outcome is normal and desirable because it means the architecture can see procedure failures that leaf scores hide.

Skills may branch. Retrying a cognitive unit call when abstention returns missing fields, or taking the validation-failure path when deterministic gates reject output, does not automatically create agency if the branch set was declared before the run and procedure evaluation covers each path. Hidden branches inside imperative glue become agent-shaped scripts without stop design.

Common mistakes follow predictable patterns. Hiding judgment in skill code means a maintainer must read imperative code to discover what was decided, when the judgment probably belongs in a cognitive unit. Calling the skill an agent because it contains models confuses the boundary: models inside a known procedure do not create pursuit. Omitting deterministic gates is equally damaging, because cognition inside a skill does not remove the need for exact checks like allergy validation and schema enforcement.

When a cognitive unit abstains, the skill declares recovery: pause and ask for missing context, retry with enriched inputs, or route to human review when terrain requires it. Abstention is a valid result, not a skill crash, when architecture expects it. Meal composition might abstain when week representation is incomplete. Extraction might abstain when scan quality is too poor for field interpretation. Recovery paths belong in procedure evaluation.

Maturity and sharing

When agent exploration discovers a stable branch tree, the mature move is to promote it into a skill rather than leaving tacit knowledge in loop code. The promotion preserves pursuit at the agent layer while giving perform work a reusable name. See when a trajectory becomes a skill.

Cost attribution benefits from the same structure. Finance teams asking which model spend belongs to which product decision can get answers when cognitive units inside skills are named. Seam-level retries show up as procedure cost, not as mysterious agent overhead. When every judgment hides inside one prompt, cost attribution becomes political. Named cognitive units turn spend into engineering facts tied to suite ownership. Instrumentation at the skill layer logs cognitive unit call order, abstentions, and seam failures. Dashboards show capability health alongside leaf accuracy.

When another team imports a skill, they inherit perform behaviour without inheriting the calling agent's goal. Procedure docs read like runbooks: expected order, abstention handling, acting grants, and known seam failures. Engineers onboarding to the skill do not need to read the agent loop to understand perform behaviour. Skills travel well when pursuit stays outside them. That portability is the reason organisations invest in skill packaging: a sealed extraction skill can serve three different products with three different pursuit policies and three different authority regimes without any modification to its internal cognitive unit calls or seam tests.

Branching inside a Skill is legitimate when each branch still maps to a named judgment with its own eval hook.

What to do next

cognitive unit calls inside skills deserve explicit design. Each open judgment stays on its own named edge, composed in procedure order. Seam evaluation at the skill level separates from leaf cognitive unit suites. Composite edges created for demo convenience dissolve under this discipline.

See two different boundaries, skill versus capability, and where the cognitive unit ends and the agent begins.

Read next: Not a composite cognitive units.