Batch 03 Notes
Theme: Spec-Driven (BMAD-aware + cc-sdd cross-runtime variants) — opinionated multi-phase methodologies targeting multiple AI runtimes simultaneously.
Roster
| # | slug | display_name | stars | distribution_type | status | key_differentiator |
|---|---|---|---|---|---|---|
| 1 | specpulse | SpecPulse | 385 | cli-tool | dormant (Nov 2025) | 8-platform CLI-first scaffold; CLI creates file structure deterministically, AI fills content into pre-existing paths — prevents directory hallucination |
| 2 | spec-kitty | Spec Kitty | 1263 | cli-tool | active | Highest stars in batch; delta-style specs (code-as-truth, spec-as-CR); git worktree per feature; retrospective mutates workflow; local Kanban dashboard; 17-agent target support; 30 contributors |
| 3 | specsmith | SpecSmith | 0 | methodology-doc | active | Pure methodology, zero tooling; minimum form = outcome + test_gates YAML only; Holdfast game as case study (377 tests, 4 milestones) |
| 4 | specs-workflow-mcp | Specs Workflow MCP | 127 | mcp-server | dormant (Aug 2025) | Single MCP tool with 5 action types; .workflow-confirmations.json persists state across sessions; AI is stateless — MCP server holds all workflow state |
| 5 | spec-driven-agentic-development | Spec-Driven Agentic Dev | 172 | standalone-repo | dormant (Dec 2025) | Parallel Explore agents (1-3 based on complexity triage) dispatched before spec writing; EARS-format requirements; /spec:create auto-triages complexity |
| 6 | cc-spec-driven | cc-spec-driven | 27 | claude-plugin | dormant (Jan 2026) | Only framework in batch using LLM-evaluated hooks (not shell scripts); Stop hook returns {"decision":"block"} to prevent completion; CR lifecycle with _deps.yaml bidirectional dependency graph |
| 7 | specd-sdd | SpecD | 9 | npm-package | active | Context compiled not discovered; customizable schema.yaml artifact workflow; multi-language code-graph (@specd/code-graph); TypeScript monorepo with 13 packages |
| 8 | moai-adk | MoAI-ADK | 1030 | cli-tool | active | Second highest stars; 20 hook events; 30 skills; 7+6 agents; auto-TDD/DDD based on coverage; V3R4 self-evolving harness; MX-tag in-code annotation; /batch for parallel worktrees |
| 9 | iac-spec-kit | IBM IaC Spec Kit | 64 | cli-tool | active | Explicit SPEC-KIT fork specialized for IaC; generic-to-specific vocabulary split (spec: "managed database" → plan: "AWS RDS"); IBM organizational contributor; no hooks, no agents |
| 10 | ai-dlc-specs-md | specs.md (AI-DLC) | 156 | npm-package | active | 3 selectable flows (Simple/FIRE/AI-DLC) in one installer; adaptive checkpoints (0-2 based on complexity); live Ink TUI dashboard with chokidar file watching; 11-tool interactive installer |
Intra-Batch Patterns
1. Two dominant philosophies for spec authority. The batch splits cleanly between "spec is truth" (specpulse, iac-spec-kit, ai-dlc-specs-md, specs-workflow-mcp — spec documents drive implementation) and "code is truth, spec is delta" (spec-kitty — code is the canonical source, specs are change requests against it). This is the most fundamental architectural divergence in spec-driven development: spec-first vs. code-as-truth. Only spec-kitty in the entire batch takes the code-as-truth position.
2. Multiple AI runtime targeting is the defining characteristic. Every framework in this batch except specsmith explicitly targets 2+ AI runtimes. The range is 2 (specsmith targets none/any) to 17 (spec-kitty). The average target count is ~8 platforms. This cross-platform ambition distinguishes this batch from single-tool specialist batches. The installer approach varies: specpulse CLI deploys identically across 8 platforms; ai-dlc-specs-md has a 4-step wizard with auto-detection; iac-spec-kit deploys to Claude Code, Cursor, and Copilot simultaneously.
3. Hook sophistication as a differentiator. Three frameworks use hooks but with radically different designs: cc-spec-driven uses LLM-evaluated prompt hooks (no shell scripts — all hook logic is an AI reasoning step), moai-adk uses 20-event shell script hooks (most event coverage in batch), and specs-workflow-mcp uses MCP as a hook substitute (tool calls trigger state transitions). The majority (specpulse, spec-kitty, specsmith, spec-driven-agentic-development, iac-spec-kit, specd-sdd, ai-dlc-specs-md) use no hooks at all.
4. Stars do not correlate with sophistication. spec-kitty (1263 stars) is genuinely the most feature-rich — but specd-sdd (9 stars) has deeper technical architecture (TypeScript monorepo, multi-language code graph, customizable schema) than several high-star frameworks. specsmith (0 stars) has the clearest pedagogical value. The gap between discovery/marketing success and technical depth is widest in this batch.
5. Dormancy is a real risk. 4 of 10 frameworks (specpulse, specs-workflow-mcp, spec-driven-agentic-development, cc-spec-driven) are dormant despite having 27-385 stars. All went dormant 6-18 months after their last activity burst. Active maintainership correlates with: npm package distribution (ai-dlc-specs-md, specd-sdd both npm+active), organizational backing (iac-spec-kit by IBM), or high contributor count (spec-kitty, 30 contributors).
Most Interesting Finds
cc-spec-driven's LLM-evaluated hooks — the only framework in the entire batch (and likely the broader corpus) where hook execution is itself an AI reasoning step rather than a shell script. The Stop hook sends the current session context to an LLM prompt that reasons about whether workflow steps are complete and returns {"decision":"block"} or {"decision":"continue"}. The PostToolUse hook similarly evaluates whether the file operation just completed maintained spec-to-implementation consistency. This is qualitatively different from grep-based shell hooks — it catches semantic violations, not syntactic ones.
spec-kitty's retrospective-as-workflow-mutator — after each feature cycle completes, the /spec-kitty:retrospective skill analyzes what worked and what didn't and writes back to the workflow's own rules files. The framework literally modifies its own operating instructions. This self-evolution is conceptually related to moai-adk's V3R4 harness evolution but implemented at the spec-workflow level rather than the agent-harness level.
ai-dlc-specs-md's adaptive checkpoint count — the FIRE flow's runtime determination of 0/1/2 checkpoints based on complexity assessment is the cleanest expression of the principle that approval overhead should scale with risk. No other framework in this batch (or the seed set) makes checkpoint count a runtime variable.
Tier C Items
None. All 10 frameworks are deployable tooling or methodology documents with direct practitioner utility. specsmith (0 stars, methodology-only) is borderline but has a real spec template and verifiable test gate format — it is not merely documentation of another tool.
Cross-References
- spec-kitty is the delta-philosophy counterpart to ai-dlc-specs-md's spec-is-truth philosophy — read both together to understand the two dominant positions on spec authority.
- moai-adk and spec-kitty are the two frameworks with retrospective/self-evolution features; moai-adk evolves the harness (V3R4), spec-kitty evolves the workflow rules.
- iac-spec-kit explicitly self-identifies as a fork of SPEC-KIT (the Batch 03 seed reference); comparing its
templates/to SPEC-KIT's reveals exactly what IaC specialization added (principles.md, generic vocabulary, cloud plan separation). - cc-spec-driven's LLM-evaluated hooks are architecturally related to the prompt-hook pattern studied in batch 17 (Hooks/TDD/governance); cross-reference for hook design comparisons.
- specd-sdd's "context compiled not discovered" philosophy is the strongest articulation of the deterministic-context pattern; compare to moai-adk's MX-tag annotation which solves the same problem differently (ambient tags vs. pre-assembled context packages).
- specs-workflow-mcp is the only MCP-server-as-workflow-enforcer in this batch; compare to MCP server approaches in batch 01 (openspec-mcp) for the two dominant designs (tool-per-action vs. single-tool-with-action-param).
- spec-driven-agentic-development's parallel Explore agents pattern is the closest thing in this batch to true multi-agent parallelism; compare to batch 08 (parallel swarm orchestrators) for production-grade implementations of the same idea.