Skip to content
/

Research phases

A five-phase research project

The corpus was built in four discovery phases (A, B, D, E) followed by a synthesis pass. Each phase doubled the corpus and discovered structural patterns the previous phase had no vocabulary for.

A

Phase A — Seed corpus

11 frameworks
5 archetypes 11 seeds

Framed the space as spec-driven development: skills-behavioral, mirror-cmd-skill, MCP-anchored, markdown-scaffold, closed-IDE. The 11 seeds became the taxonomy anchors.

Key finding

The field could be described with 5 archetypes. All seeds shipped installable methodology.

B

Phase B — Tier A frameworks

319 frameworks
15 archetypes 33 batches

Uncovered an entire operational layer the seeds had no vocabulary for: multi-agent orchestrators, PR-lifecycle daemons, governance frameworks, UI passthroughs, memory engines, observability layers. Batch synthesis notes in _index/batch-NN-notes.md.

Key finding

The framework set was not just about spec-driven development. An orchestration/governance layer existed independently.

D

Phase D — Tier B frameworks

323 frameworks
21 archetypes (+6 new) 32 batches

Three new races surfaced that didn't exist in Phase A's taxonomy: (1) cross-vendor model routing — 20 frameworks routing Claude vs GPT vs Gemini; (2) token-efficient symbolic encoding — 14 frameworks, led by rtk-token-killer at 54k★; (3) LLM-evaluated hooks — 4-5 frameworks where the hook itself is an LLM call. Phase D also surfaced claude-code-system-prompts (Piebald AI) — revealing Anthropic's own Claude Code already implements 8-stage multi-agent review internally.

Key finding

The community is unknowingly reinventing capabilities that already exist inside Claude Code.

E

Phase E — Synthesis

25 rubric dimensions Cross-corpus

Cross-corpus synthesis identifying: top-15 deep-dive frameworks, 5-framework reference baseline, 7-framework defensible set, 45 one-of-a-kind architectural primitives, and the headline finding that the community has split into five concurrent races.

Key finding

The Phase C 5-framework synthesis baseline (superpowers + openspec + AO + forge-lucasduys + entroly) was validated. Phase D added Trellis and gh-aw-githubnext to yield a 7-framework defensible set.