DotLightSkillset

A lightweight, curated Claude Code skillset for .NET developers.

Plugin identifier: dotlight-skillset

Combines five upstream MIT skill libraries into one opinionated bundle, with workflow overrides that fix the rough edges of "pure TDD" agent loops — and, since 0.6.0, a minimalism layer that keeps generated code reviewable by humans and a designed (not test-accreted) persistence layer.

What it bundles

Workflow (19 skills) — customized fork of obra/superpowers v6.1.1 plus adapted skills from mattpocock/skills (grill-me, design-an-interface, grill-with-docs), hsmejky/skills (improve-architecture, caveman), and DietrichGebert/ponytail (lazy-senior-dev): brainstorming → writing-plans → executing-plans → TDD → code review → finishing-branch, plus worktrees, systematic-debugging, parallel agents, skill authoring, doc-aware grilling, Ousterhout-style architecture review, explicit-trigger ultra-terse mode, and the lazy-senior-dev minimalism reflex (reuse before write, stdlib before dependency, the minimal version of the agreed design).
.NET patterns (29 skills) — curated fork of Aaronontheweb/dotnet-skills v1.4.1: C# standards, Minimal API design, DI, configuration, serialization, Aspire (4 skills — configuration, service-defaults, integration-testing, mcp-first for runtime debugging), OpenTelemetry (v1.4.1 rewrite + 3 reference files), Testcontainers, Playwright (Blazor + CI caching), ILSpy decompilation, quality gates (slopwatch + CRAP + database-review). Plus dotlight originals: rider-mcp-first (Rider MCP semantic ops over filesystem Grep/Read — 50–90 % token savings on .NET exploration), nhibernate-patterns (NHibernate as first-class runtime — most dotlight projects run NH, not EF), database-design-conventions (lifecycle-pattern catalog incl. ActiveFrom/ActiveTo temporal versioning, one-schema-one-dialect, normalization gate), and database-review (schema quality gate against test-shaped schemas and dialect drift).
Specialized .NET agents (3) — dotnet-performance-analyst, dotnet-benchmark-designer, dotnet-concurrency-specialist. Use the Agent tool with subagent_type matching the agent name.

Superpowers drives the process, dotnet-skills supply the patterns, agents do focused diagnostic work.

Why this exists

Out-of-the-box agent workflows have four habits that hurt .NET projects:

TDD-first domain discovery. "Minimal code to pass" applied literally produces entities without relationships and invariants.
Text-based Socratic dialogue. Multiple-choice questions as plain-text lists, even in clients that render AskUserQuestion as clickable choice cards.
Over-generation. Agents produce enormous, speculative diffs that no senior developer can sustainably review — humans are lazy in the good way; agents are not.
Test-shaped schemas. A DB layer accreted from TDD cycles is tailor-made for the tests, not designed for persistence.

Fixed in modified SKILL.md files:

brainstorming — prefers AskUserQuestion over text multi-choice, caps questions at 5-8, enforces Domain Model as the first design section, and always proposes the minimal version first (reviewability is a design constraint).
writing-plans — requires a ## Domain Model section AND a ## Persistence Model section (may be None — no schema changes) derived from the design. If missing, loops back. Minimalism Gate: mandatory ## Non-Goals, every file names its forcing requirement, reverse task-to-requirement coverage. Default exec sub-skill is executing-plans, not subagent-driven-development.
test-driven-development — domain model must exist before first RED-GREEN-REFACTOR; schema changes are gated on the plan's Persistence Model; tests go at pre-agreed seams. Calls out "test-cheating" (satisfying tests by violating invariants) and tautological tests as the top LLM-TDD failure modes.
lazy-senior-dev + requesting-code-review — the ladder (does it need to exist → reuse → stdlib → platform → one line → minimum code) governs implementation of the agreed design; review flags diffs over a human review budget, single-implementation interfaces, and unrequested features as spec-compliance findings.

What's deliberately excluded