Phase D Batch 09 — Batch Notes

Theme: Claude Code starter templates — "claude-code-starter"-style scaffolds bootstrapping .claude/, agents, rules, commands, skills, hooks

Date analyzed: 2026-05-26

Roster Table

Comparison Matrix: The "*-starter" Namesakes

Four repositories share the "claude-code-starter" naming pattern. What each actually installs:

Observation

Despite sharing a name, these four "starters" share nothing architecturally:

• zbruhnke is a safety/enforcement framework masquerading as a starter
• wyndomb targets knowledge workers, not developers — unique in entire batch
• alexeykrol is a versioned (v6) Russian methodology system with session state
• cloudnative is a wizard-based enterprise installer with marketplace integration

A developer searching "claude-code-starter" on GitHub would receive four completely incompatible frameworks serving different needs.

Intra-Batch Patterns

Pattern 1: Language Localization

Two of the ten frameworks are non-English primary:

• alexeykrol-cc-starter — Russian (CLAUDE.md, agent instructions, skill prompts)
• cloudnative-cc-starter — Japanese (documentation, installer wizard prompts)

Both have high star counts (180, 116) indicating active non-English user communities. Neither English framework equivalent has equivalent explicit localization.

Pattern 2: OpenAI Codex Infiltration

Two of the ten frameworks target OpenAI Codex, not Claude Code:

• codex-starter-acthouse — AGENTS.md + minimal setup, container-based isolation
• codex-second-brain-starter — Full BASB second-brain builder, enterprise Tyler gateway, explicit reasoning effort routing (medium/high)

These are categorically different architectures (AGENTS.md vs CLAUDE.md, .toml sub-agents, OpenAI Responses API). They were included in the batch due to "codex-starter" naming overlap. Both should be tagged with primary_tool: openai-codex and analyzed separately from Claude Code frameworks.

Pattern 3: Scale Extremes

The batch spans an unusual range:

• notque-cc-toolkit — 44 agents, 50+ skills, 77 Python hooks (the most complex single framework in the entire research)
• alex-feel-cc-toolbox — 0 agents, 0 skills, 0 commands, 0 hooks (ships nothing; reads user YAML)
• codex-second-brain-starter — 0 stars, 0 commands, 0 agents but deep content (BASB ontology, reasoning effort routing)

Pattern 4: Memory Architecture Diversity

This batch contains 4 distinct memory architectures:

1. centminmod — Dual-file memory (git-shared CLAUDE-.md + machine-local memory/.md)
2. alexeykrol — Session-scoped SNAPSHOT.md (ephemeral per-session)
3. notque — SQLite with confidence decay and scheduled pruning (0.05 decay/30 days)
4. serpro69 — SQLite BM25 via Capy MCP (no decay, project-committed)

The contrast between notque's decay-based learning database and serpro69's accumulation-only store represents a genuine design fork: does knowledge degrade or compound?

Pattern 5: The Meta-Installer vs Content Dichotomy

alex-feel-cc-toolbox represents an inversion of all other frameworks: instead of shipping skills/agents/commands, it ships an installer that reads a user-defined YAML and fetches/installs whatever the YAML specifies. Configuration inheritance (parent → team → personal, three levels) is unique to this framework.

All other frameworks ship content. alex-feel ships a delivery mechanism.

Pattern 6: OS-Level vs Claude-Level Enforcement

zbruhnke installs a git pre-commit hook at the OS level — Claude cannot bypass it because it runs outside Claude's execution context. This is categorically different from hooks that run inside Claude Code (PreToolUse, PostToolUse) which Claude could theoretically work around.

No other framework in the batch distinguishes between "enforced at OS level" and "enforced at Claude level."

Most Interesting Finds

1. notque's SQLite Learning Database with Confidence Decay

The only framework in the entire research (not just this batch) to implement a learning database with confidence scoring that ages and prunes. Entries decay at 0.05/30 days; entries below 0.3 confidence with <90 days of use are pruned. This is a genuine epistemics system for agent knowledge management. No seed has anything comparable.

Reference: /Users/yigitkonur/research/spec-driven-dev/spec-system/notque-cc-toolkit/06-memory-context.md

2. serpro69's Cross-Provider Code Generation

klaude-plugin/ → kodex-plugin/ via Go tooling (cmd/generate-kodex/) is the cleanest solution to the multi-provider problem in the batch. Other frameworks treat Claude Code and Codex as separate targets. serpro69 maintains a single canonical source and generates the Codex output — including /kk: → $kk: prefix rewriting and .md → .toml agent conversion. CI enforces freshness.

Reference: /Users/yigitkonur/research/spec-driven-dev/spec-system/serpro69-claude-toolbox/03-components.md

3. cloudnative's Plugin Marketplace Integration

cloudnative-cc-starter is the only framework in the batch (and possibly in the entire research) to integrate with the Anthropic official plugin marketplace (config/plugins.json pointing to anthropics/claude-plugins-official and anthropics/skills). This anticipates an ecosystem that other frameworks build around static files — if the marketplace matures, cloudnative's architecture looks forward-compatible.

Reference: /Users/yigitkonur/research/spec-driven-dev/spec-system/cloudnative-cc-starter/01-distribution.md

4. codex-second-brain's Explicit Reasoning Effort Routing

The BASB second-brain builder explicitly routes different task types to different reasoning effort levels (medium vs. high) via the OpenAI Responses API reasoning effort parameter. This is the clearest example of effort-as-architecture in the batch — not "set it to max," but map task types to appropriate compute levels.

Reference: /Users/yigitkonur/research/spec-driven-dev/spec-system/codex-second-brain-starter/07-orchestration.md

5. zbruhnke's Skill Auto-Router via UserPromptSubmit Hook

skill-eval.js runs on every user prompt, parses the message against skill-rules.json rules, and automatically invokes the appropriate skill. This transforms skills from user-invoked commands into auto-dispatched behaviors — the user never needs to know the skill names. No seed does this; it's a qualitatively different UX for skill systems.

Reference: /Users/yigitkonur/research/spec-driven-dev/spec-system/zbruhnke-cc-starter/03-components.md

Tier C Items

None. All 10 frameworks have sufficient material for a full 11-file analysis:

• codex-starter-acthouse and codex-second-brain-starter are minimal and target OpenAI Codex, but they're architecturally distinct enough to be informative
• alex-feel-cc-toolbox ships no content but the meta-installer architecture is well-documented in the repo

Cross-References Discovered

Architecturally Related (Within Batch)

• notque + serpro69 — both use SQLite as persistent knowledge store, but fundamentally different approaches (notque: confidence decay + pruning; serpro69: BM25 accumulation via MCP)
• centminmod + serpro69 — both dual-source (centminmod: CLAUDE-.md + memory/.md; serpro69: docs/wip/ + knowledge.db)
• codex-starter-acthouse + codex-second-brain-starter — both OpenAI Codex targets, should be compared as a sub-cluster
• alexeykrol + cloudnative — both have structured installer workflows and versioned methodology systems
• alex-feel + cloudnative — both install globally to ~/.claude/ and provide multi-environment support

Cross-Batch References

• notque's AFK/headless mode — similar concept to autonomous execution modes in batch 01's single-author SDD frameworks; cross-reference when aggregating orchestration patterns
• serpro69's isolated review sub-agents — similar to batch 29's agent SDK frameworks' reviewer patterns; cross-reference multi-model review architecture
• alex-feel's meta-installer YAML — closest peer is kiro's project configuration system (mentioned in alex-feel's METRICS.yaml differs_from_seeds); cross-reference environment-as-code paradigm
• cloudnative's plugin marketplace integration — forward-reference to any batch analyzing Anthropic's official plugin ecosystem

Seed Relationships

• centminmod most closely parallels the ccmemory seed (memory architecture) but adds multi-provider delegation
• serpro69 most closely parallels spec-driver seed (structured pipeline) but adds language profiles + BM25 memory
• notque has no close seed parallel — closest is taskmaster-ai (orchestration scale) but notque's hook depth is unmatched
• alex-feel most closely parallels kiro (declarative project config) but targets installation rather than task spec