Skip to content
/
Phase D Batch 10

Batch 10 — Domain-specific skill packs (Unity, BI, science, AWS)

Batch 10 — Domain-specific skill packs (Unity, BI, science, AWS)

Roster (10)

slug stars distribution cli_binary local_ui orchestration multi_model tier
nowsprinting-cc-unity 37 (archived) claude-plugin no no hierarchical yes (opus for test-designer) A
power-bi-agentic-dev 619 claude-plugin (marketplace) no no hierarchical no A
pbip-demo-agentic-mcp 30 methodology-doc no no sequential no A
reservoir-engineering-claude 28 methodology-doc no no parallel-fan-out no A
matsengrp-plugins 22 claude-plugin no no sequential no A
claude-scientific-skills 26,062 skill-pack no no none no A
aws-skills 294 claude-plugin (marketplace) no no none no A
altimate-code 604 cli-tool yes (altimate) terminal-tui none yes (multi-provider) A
szeider-python-coder 3 cli-tool yes (coder) no parallel-fan-out yes (openrouter) A
ind-kit 22 template-bundle no web-app (Next.js) sequential no A

Intra-batch patterns

All 10 frameworks share a domain-vertical focus — none is a general-purpose agent harness. The dominant pattern is skills-only (no commands): 7 of 10 ship primarily SKILL.md files. The two CLI tools (altimate-code, szeider-python-coder) are the only frameworks with standalone runtimes; both support multi-model via provider routing (altimate via built-in providers, szeider via OpenRouter). Power BI appears twice (power-bi-agentic-dev and pbip-demo-agentic-mcp), reflecting the maturity of Microsoft's PBIP/TMDL open file format as an agentic development surface. Hook sophistication varies widely: power-bi-agentic-dev has the most domain-specific PostToolUse hooks (TMDL syntax, DAX validation, referential integrity), ind-kit has Next.js-specific hooks (use client detection, loading.tsx suggestions), while 5 frameworks have no hooks at all.

Most interesting find

  1. nowsprinting-cc-unity: The TESTABILITY gate — a test-designer subagent (opus model) that returns PASS/WARN/FAIL on the Plan agent's class design before any implementation begins — is the only example in the corpus of a structured feedback loop between design agents with a formal verdict type. It makes TDD structurally mandatory rather than prompt-instructed.

  2. altimate-code: The deterministic SQL intelligence layer (19 anti-pattern rules at 100% F1, column-level lineage at 100% edge-match, PII detection) combined with the bridge-guard.md command documenting actual production regression history (v1.4.0 incidents) makes this the most engineering-disciplined framework in the batch — it builds rule-based tools for problems that should not be probabilistic.

Items written as Tier C

None. All 10 frameworks had sufficient public material for full 11-file analysis.

Cross-references discovered

  • power-bi-agentic-dev and pbip-demo-agentic-mcp share the powerbi-modeling-mcp dependency (Microsoft private preview MCP server); pbip-demo-agentic-mcp is the demo repo, power-bi-agentic-dev is the production plugin.
  • altimate-code is an explicit fork of OpenCode (anomalyco/opencode) — not in this batch but a related framework.
  • reservoir-engineering-claude explicitly adapted its exercise structure from claude-code-for-hydrology (lorenliu13) — not in this batch.
  • claude-scientific-skills adopted the agentskills.io open standard and ships alongside a companion K-Dense BYOK desktop app — the skills are a distribution channel for a commercial product.
  • ind-kit ships .cursor/rules and .windsurf/ in addition to .claude/ — demonstrating that SaaS starter kits are converging on shipping AI context for all major IDE agents simultaneously.