Quality Model

rd-skills quality is the combination of control correctness, professional judgment, bounded context, observable evidence, and maintainable packaging.

Control Correctness

Start simple, inspect current source, and implement once important decisions are clear. Analysis, stable briefs, dependency planning, formal contracts, and independent review are available when the user asks or evidence justifies them. There is no execution-level hierarchy or replacement score/state system. Shared writes remain serial unless the Host supplies isolated workspaces and there is no dependency or shared write surface.

Professional Judgment

Engineering correctness comes first, followed by accurate Professional routing, professional substance, token cost, and execution time. Each assignment uses one Primary Professional and task-needed authorized Layer 3 Skills, with positive and anti-trigger evidence. When independent Review is needed, Main selects its Professional and Layer 3 expertise separately. The reviewer consumes that assignment without rerouting or copying the implementation's Layer 3 selection. Current shared state, consumers, financial invariants, concurrency, transaction, migration, recovery, and security boundaries require enough depth to implement and verify them correctly.

Prefer existing owners and structures. New abstractions or dependencies need a current requirement, real variation, a shared boundary, or repository architecture evidence. Hypothetical future reuse or safety is insufficient.

Evidence Quality

Validation must follow the final material edit. Preserve the validation oracle and distinguish behavioral failures from environment failures. Inspect current owners, tests, and relevant consumers; inherited selectors and truncated searches do not prove absence or completeness. State proof limits explicitly.

Use independent review for a concrete reason, and inspect the current diff or artifact when doing so. Findings retain the defect, evidence, reachable failure, and required action. Fix ordinary defects within the current direction. New evidence can justify redesign or another review; a repair alone does not.

Evidence Privacy

rd-skills has no required product-runtime or private telemetry channel. Ordinary agents use visible assignments, source, diffs, commands, validation, review, repair, and handoff evidence. They do not write private event streams, internal task identities, prompt transcripts, or hidden evidence records.

Evidence artifacts exclude raw prompts, secrets, environment variables, full command logs, personal archives, and user-specific content. Local evaluation report retention follows repository and user policy. A retained report must identify its evidence type and source freshness; persistence does not make stale output current. Benchmarks owns interpretation and proof limits.

Content Quality

Professional and Layer 3 Skills are concise AI instructions, not background articles. Their decision rules, gotchas, execution checklist, stop conditions, output contract, and targeted references must change a current engineering decision.

Claim Quality

Structural validators prove structure. Deterministic and captured fixtures prove only their declared contracts. Default code-generation commands prove only case definition integrity, checked-in harness execution, and starter negative-control behavior; they do not evaluate a generated candidate. These artifacts do not prove real-host startup, wall-clock performance, provider behavior, production accuracy, or installed user experience. Generated catalogs are local discovery assets. The handwritten scorecard records release expectations, not generated status. Repository evidence does not support project rankings, popularity, external-adoption, privacy-guarantee, or competitor-superiority claims beyond the specific evidence it contains.