---

type: adr id: adr-0064 status: accepted created: 2026-06-11 updated: 2026-06-11

ADR-0064 — Minimal starter kit: core copy surface, advanced tier, focused guides

Context

Adoption collapses under ceremony: the adopter audit counted 47–49 installed files before the first line of project content (O-006, R-001), judged several broad personas to be internalized stances adding noise, and found the skill-loading story underdocumented (O-007) — while singling out the surveyor stance as the standout (O-008) and the audit template as the fastest path to value (O-001). Vendor guidance favors few, focused skills over broad personas.

Decision

1. Core copy surface = 12 files. templates/{spec, task, review, finding, status, intake, inventory, change-plan}.md + agent/AGENTS.md + three focused agent guides: agent/write-spec/, agent/implement-task/, agent/review-output/ (SKILL.md carrier — auto-discoverable by agent CLIs, plainly readable by humans). Whole kit excluding advanced/ ≤ 24 files.
2. The three guides absorb the cross-cutting disciplines. write-spec carries the architect stance (intent not implementation; survey before inventing) and the dropped-from-sources practice; implement-task carries scope discipline, evidence-or-it-didn't-happen, and adversarial self-review before handoff; review-output carries refute-by-default, the evidence rules, and the finding-saving close. No standalone persona files in the core kit.
3. Advanced tier (starter-kit/advanced/, all optional): templates {audit, bug, research, adr, rfc, prd, threat-model}.md; reference cards sol-reference.md (the notation) and checks-reference.md; guides {write-audit, write-research, persona-surveyor, write-bug-report, write-prd, write-rfc, write-change-plan, write-inventory, spec-check, save-findings, split-work}. The audit template is flagged "recommended first taste for brownfield teams". persona-surveyor stays a standalone guide — its evidentiary discipline (three named instances; observation vs claim) is the one stance that does not fold cleanly.
4. Adoption is manual-first: a copy checklist (the 12 files) leads; the agent-assisted prompt is the second path; a future corpus init is the third.

Alternatives considered

Alternative	Why weaker
Ship everything, label tiers in docs only	The install footprint is the friction; labels don't shrink it
Plain .md guides instead of SKILL.md dirs	Loses agent-CLI auto-discovery; SKILL.md reads fine as plain markdown
Fold surveyor too	Its rules are load-bearing for research quality and don't belong to any single write-guide

Consequences

Positive: five-minute adoption; every core file earns its place. Negative: advanced material is a second copy step. Neutral: guides install beside user skills (pass--free names cannot collide).

Status

Accepted. Partially supersedes ADR-0019, ADR-0002, ADR-0009 (persona shipping model); refines ADR-0042, ADR-0036, ADR-0047, ADR-0056.

Propagation

starter-kit tree, ADOPTING, docs/03/10, .agents dev subset, code-skills library index.

Addendum (2026-06-11): the advanced tier additionally ships adversarial-review — a deep hostile re-review guide (with a session task template) for agent branches that warrant more than the review packet. It pairs with a dev-tier copy at .agents/skills/adversarial-review/; the pair is registered in the propagation matrix.

Ledger note (2026-06-12): adoption framing (the 12-file copy checklist) superseded by ADR-0069 — the kit is copied whole as a workspace; the core/advanced tiering and the guide set survive unchanged.