| name | crow-pet-skill-architect |
| description | Choose the execution architecture for new or changing skills before implementation: inline, thin references, deterministic scripts, composed pipeline, compiled workflow, or manual side-effect workflow. |
crow.pet Skill Architect
Use this skill when crow.pet is asked to create, adapt, compose, or harden a skill or repeated workflow.
Workflow
- Read the task record, user ask, and linked
SPEC.MD when present.
- Decide whether the request is a simple skill, large reference workflow, deterministic repeated operation, high-output multi-stage workflow, or side-effect workflow.
- Draft a skill architecture plan from
../../templates/skill-architecture.md.
- Choose one execution mode:
inline, thin-references, deterministic-script, composed-pipeline, compiled-workflow, or manual-side-effect.
- Use semantic skill organization checks when the request concerns skill shape, naming, audit, scaffold, or folder promotion.
- Declare isolation, handoff strategy, dynamic context, context budget, cleanup policy, and capability needs.
- For composed pipelines and compiled workflows, define stages and compact handoffs from
../../templates/stage-handoff.md.
- Route semantic needs through Capability Broker.
- Route side effects through Policy Gate.
- Create a workflow definition from
../../templates/workflow-definition.md when repeatability, fragility, parallel fan-out, budgets, retries, pause/skip, or idempotency justify compilation.
- Map portable execution semantics through
../../adapters/patterns/execution-architecture.md and ../../adapters/patterns/compiled-workflow.md when a harness-specific feature is needed.
- Hand the approved architecture to
crow-pet-spec-writer, crow-pet-work-queue, or the selected builder skill.
- Evaluate the result for correctness, portability, semantic organization, and context efficiency.
Progressive Disclosure
Load ../../templates/skill-architecture.md only when drafting or checking an architecture plan.
Load ../../templates/stage-handoff.md only when the chosen mode is composed-pipeline or compiled-workflow.
Load ../../templates/workflow-definition.md only when the chosen mode is compiled-workflow.
Load workflows/semantic-skill-organization.md only when creating, adapting, auditing, renaming, or splitting a skill or repeated workflow.
Load ../../adapters/patterns/execution-architecture.md when mapping portable execution semantics to a target harness.
Load ../../adapters/patterns/compiled-workflow.md when mapping a reusable workflow definition to a target runner or work item instance.
Load ../../adapters/patterns/capability-broker.md only when semantic capabilities need concrete adapter bindings.
Load ../crow-pet-policy-gate/SKILL.md when side effects, external services, publication, destructive actions, or sensitive context are involved.
Load contract.md and rubric.md only for acceptance or evaluation.
Dependency Edges
delegates_to: crow-pet-spec-writer, crow-pet-work-queue, selected builder skill, or selected target adapter.uses_adapter: adapters/patterns/execution-architecture.md, adapters/patterns/compiled-workflow.md, adapters/patterns/capability-broker.md, and selected harness or target adapters.loads_context: task record, linked SPEC.MD, existing skill files, and context capsule only when task-relevant memory affects the architecture.evaluates_with: contract.md, rubric.md, templates/skill-architecture.md, templates/workflow-definition.md, workflows/semantic-skill-organization.md, and stage handoff evidence when used.
Rules
- Do not make every skill a pipeline.
- Keep one skill to one durable purpose; split only when semantic migration triggers fire.
- Prefer the smallest execution architecture that satisfies the workflow.
- Keep canonical skills portable; map runtime-specific fields through adapters.
- Use scripts for deterministic repeated work when they reduce context and ambiguity.
- Use composed pipelines only for large, noisy, repeated, or multi-stage workflows.
- Use compiled workflows only when the workflow is repeatable, fragile, high-volume, long-running, budget-sensitive, or needs retries, pause/skip, idempotency, or runtime visibility.
- Preserve the invariant: parallel read, sequenced write.
- Use manual-side-effect mode for workflows that publish, spend, deploy, message externally, or mutate high-risk systems.
- Stage outputs should be compact references or summaries, not raw dumps.
- Avoid nested delegated-agent structures unless the user explicitly asks for them and the selected harness supports them.
- Dynamic context must be distilled before injection; never treat a whole file tree, transcript, or raw scrape as a stage input.
- Do not import target marketplace rules, plugin manifests, or install-specific conventions into canonical crow.pet skills.
