Skip to content
/

An opinionated atlas of frameworks

Choosing between Superpowers, SpecKit, OpenSpec, BMAD, and Taskmaster?

Start here. The atlas reads 654 entries348 frameworks built on top of Claude Code, Cursor, Codex, and others, plus 305 supporting tools they sit on, against a 25-dimension rubric — and explains, in plain language, what each one actually does differently.

§ focus

The five most-asked-about.

Each ships installable methodology — commands, skills, hooks, MCP, personas. Each card reads best when and skip if from its METRICS, so the design choice is up front, not in the docs.

Superpowers
superpowers

Iron-law SKILL.md + SessionStart hook. The behavioral floor everyone else builds against.

best when Skills should be Iron Laws with pre-built rationalization tables, not suggestions — and they should trigger automatically via session hooks without any user action.
skip if Skipping brainstorming for 'simple' projects
Spec Kit
spec-kit

GitHub's official toolkit. Spec → Plan → Tasks pipeline with 9 commands + 18 before/after hooks.

best when Specifications should be the source of truth from which code is continuously regenerated — not documentation that trails implementation.
skip if Vibe coding (prompting without structured specs)
OpenSpec
openspec

Spec-as-source-of-truth across 29 AI tools. Delta-diff specs, mirrored cmd+skill model.

best when Planning artifacts should be editable schema-driven files that any team can customise without waiting for a framework release.
skip if Rigid phase gates that block progress
BMAD-METHOD
bmad-method

Six named persona subagents — analyst, architect, dev, pm, sm, ux. Genuine multi-agent 'Party Mode'.

best when AI should act as a Socratic facilitator that brings out the user's vision — not an autonomous author — and genuine independent subagents produce better multi-perspective output than a single LLM roleplaying multiple characters.
skip if AI doing the thinking for the user and producing average results
Taskmaster AI
taskmaster-ai

PRD → tasks.json → autonomous expansion. The analyze-complexity stage writes the next stage's prompts.

best when The AI agent should be stateless across sessions; all project memory belongs in structured files (tasks.json), not in conversation history or model weights.
skip if Manually editing tasks.json
653
entries total
348
frameworks
305
supporting tools
21
archetypes

§ collisions

Why the field feels confusing.

Same name, wildly different things. The corpus has six independent projects called Conductor, eight or more oh-my-* repos with opposite purposes, and a long museum of similar collisions.

→ The full Naming Collision Museum

The Stack · synthesis baseline

Five frameworks, one defensible reference design

If forced to pick five frameworks to merge into a flawless reference design — this is the Phase E answer. Each layer solves exactly one dimension; the five together cover all 25 rubric dimensions. Phase D adds two more for sub-agent context injection and security governance.

What should I actually build? →

Source: _index/PHASE-E-EXECUTIVE-SUMMARY.md §5

This is the same 5-framework baseline from Phase C — Phase D did not surface a framework that replaces any on its own dimension. The defensible 7-framework set adds Trellis (LLM-evaluated hooks at the sub-agent boundary) and gh-aw-githubnext (security/governance reference). entroly is a memory tool (category: other) but is the only substrate-level memory framework in the corpus — included here with an explicit label.

§ entrances

Three ways in.

Different readers want different entry points. The Console for sorting; the archetypes for grouping; the tour for narrative.

§ spotlight

Twelve worth a deep look.

One per archetype family — curated for architectural distinctness, not stars. Stars are an SEO signal in this corpus; these twelve are the ones where the engineering depth matches.

Show top-by-stars instead (cautionary)