ADR-0112 — Two-tier skills: a framework-free universal catalog, a Suspec-coupled kit

Context

ADR-0111 drew the kit/catalog line as "Suspec-concept builders + core-loop methods → kit; styles / stances / depth → catalog." That made the catalog a catch-all and left the catalog's skills shot through with Suspec concepts — most reference spec / task packet / review packet / AC, and adversarial-review even bundled a type: review packet template. A skill that names a Suspec artifact cannot be installed and used in a non-Suspec repo, which defeats the whole point of an installable catalog. A research pass (deep-research, 105 agents, 19 confirmed claims) also established the adopted-skill anatomy: a skill earns use by changing a default (not restating one), fitting one purpose, triggering off its description, explaining the why over all-caps imperatives, staying <500 lines with progressive disclosure, and accreting a Gotchas section — over-specification measurably hurts ([[AGENTSMD-HARM]]).

Decision

Partition skills by coupling, as a binary — there is no middle tier.

1. The universality test. A skill is universal iff its body and its bundled files name zero Suspec concepts — no spec, task / task packet, review / review packet, finding, acceptance criterion / AC, ## Execution, type: spec|task|review, status board, or SPEC-/TASK-/REVIEW- id, and no citation of a Suspec ADR (an ADR reference is itself a coupling smell). The moment it names one, it is a Suspec skill.

2. Placement is the binary. Framework-free → the catalog (suspec-skills): the universal disciplines and stances, installable into any repo (npx skills) with zero Suspec knowledge. Suspec-coupled → the kit (.agents/skills/): every skill that operates a Suspec concept — the artifact builders, the core-loop methods, and the write-* task-implementation depth.

3. The kit is an always-core + an opt-in Suspec-depth group. The minimal core (the loop + the always-needed authoring) ships first; the write-* kind-specialised implementers are opt-in depth an adopter prunes or ignores (progressive disclosure keeps their always-on cost ~one description each). This is ADR-0105 stretch-collapse applied to skills: depth is summoned, not forced.

Level: convention.

Consequences

• The catalog becomes genuinely universal — a developer with no Suspec install gets the ultimate versions of the fundamental disciplines (review, the stances, evidence, concision, flaky-test).
• The kit holds every Suspec-coupled skill in one home (no homeless "Suspec-coupled depth").
• Moves: write-feature/-fix/-refactor/-rewrite/-migration/-performance/-testing/-documentation relocate catalog → kit; the catalog's long-form implement-task folds into the kit's. The universal set (adversarial-review, persona-skeptic/-challenger/-surveyor, empirical-proof, concise-output, fix-flaky-test) is stripped framework-free — adversarial-review loses its type:review template (its Suspec review-packet building already lives in the kit's review-output).
• The anatomy upgrades (Gotchas, explain-why, behavior-changing) apply to both tiers.
• Migration tail: the relocation is mechanical; no artifact format changes.

Affected obligations / constraints

• Refines: ADR-0111 (catalog = universal-only, not the catch-all; Suspec-coupled depth is kit-side) and ADR-0064 (kit = always-core + opt-in Suspec-depth group). Reaffirms: ADR-0105 (depth is a dial), ADR-0093, single-sourcing. Grounded by: [[AGENTSMD-HARM]] (over-specification harm) + the deep-research skill-anatomy synthesis.
• Does NOT change: the artifact formats, the core loop, the verdict model, or the checks contract. Accepted ADRs 0064/0111 are refined here, not edited in place (Nygard immutability).