obra/superpowers brainstorming vs ed3d3-plugins/ed3d-plan-and-execute

A concrete comparison of the original Superpowers brainstorming/planning workflow and the Ed Ropple plan-and-execute fork derived from it.

• Upstream: https://github.com/obra/superpowers
• Fork: https://github.com/ed3dai/ed3d-plugins (the ed3d-plan-and-execute plugin)

TL;DR

obra/superpowers is a flat set of skills: brainstorming → writing-plans → executing-plans (or subagent-driven-development). Each skill is self-contained and conversational.

ed3d3-plugins/ed3d-plan-and-execute is a fork-and-fan-out: it takes the same backbone but explodes it into three orchestrator commands (/start-design-plan, /start-implementation-plan, /execute-implementation-plan), 14 skills, and 4 dedicated subagents, with mandatory /clear between each phase.

Concrete differences

1. Skill surface area

	superpowers	ed3d-plan-and-execute
Skills in flow	3 (brainstorming, writing-plans, executing-plans)	14 (adds: starting-a-{design,implementation}-plan, writing-design-plans, writing-implementation-plans, asking-clarifying-questions, using-plan-and-execute, executing-an-implementation-plan, requesting-code-review, etc.)
Subagents	none baked in (uses dispatching-parallel-agents skill)	4 (task-implementor-fast, code-reviewer, task-bug-fixer, test-analyst) plus depends on 3 other plugins (ed3d-research-agents, ed3d-house-style, ed3d-extending-claude)
Slash commands	none — skills only	5 commands wrap the skills
brainstorming/SKILL.md length	164 lines	366 lines
executing-plans/SKILL.md length	70 lines	562 lines

2. What brainstorming itself owns

Superpowers brainstorming is end-to-end. Its checklist runs all 9 steps including writing the design doc to docs/superpowers/specs/YYYY-MM-DD-<topic>-design.md, spec self-review, user-review gate, and handoff to writing-plans. It also includes a "Visual Companion" (browser-based mockups) feature that ed3d does not.

ed3d brainstorming is user-invocable: false. It owns only three phases: Understanding → Exploration → Design Presentation, all held in conversation context. Writing the doc is moved to a separate writing-design-plans skill, and the whole flow is orchestrated by starting-a-design-plan, which adds two phases superpowers doesn't have: a dedicated Clarification phase (invokes asking-clarifying-questions) and a Definition of Done confirmation phase that creates a skeleton design doc before brainstorming begins, to "capture DoD at peak resolution."

3. Research strategy

Superpowers brainstorming does its own context-gathering ("Check out the current project state first").

ed3d brainstorming forbids this: "DO NOT perform deep research yourself. Delegate to specialized agents." It mandates dispatching codebase-investigator for "how is X implemented?" questions and internet-researcher (or WebSearch/WebFetch) for external docs. The rationale given: "You'll consume context and may hallucinate."

4. User-interaction primitive

Superpowers prefers "multiple choice... but open-ended is fine too" — natural dialogue.

ed3d mandates the AskUserQuestion tool for any 2–4 option choice, with structured trade-offs. Open-ended is reserved for design-validation feedback. It explicitly warns against "trap answers" (options that look plausible but are dangerous).

5. Context isolation

Superpowers chains skills in the same conversation.

ed3d requires /clear between phases — the docs include a copy-this-command-before-clearing-context pattern:

/ed3d-plan-and-execute:start-implementation-plan @docs/design-plans/...md .

The rationale: "Long conversations accumulate context that degrades quality."

6. Design vs. plan philosophy

The key conceptual addition in ed3d is the design plan / implementation plan split:

• Design plan: "archival… can be checked into git and referenced months later… does NOT include implementation code."
• Implementation plan: "just-in-time… verifies current codebase state matches design assumptions. Generates fresh, executable code based on actual state. May diverge from design if codebase has changed."

writing-implementation-plans explicitly says: "Design plans may be days or weeks old. Your code must work with the codebase as it exists NOW." It dispatches codebase-investigator to re-verify every assumption per phase before writing tasks.

Superpowers writing-plans collapses this into one document and assumes the spec was written recently.

7. File-system layout

	superpowers	ed3d
Spec/design	docs/superpowers/specs/YYYY-MM-DD-<topic>-design.md	docs/design-plans/YYYY-MM-DD-<slug>.md
Plan	single file docs/superpowers/plans/YYYY-MM-DD-<feature>.md	directory docs/implementation-plans/YYYY-MM-DD-<feature>/phase_##.md + test-requirements.md
Phase cap	none explicit	hard cap of 8; skill refuses to write more
Acceptance criteria	implicit in spec	explicit, ID-tagged (<slug>.AC1.1), mapped to tests
Project customization	none baked in	.ed3d/design-plan-guidance.md, .ed3d/implementation-plan-guidance.md injected at specific phases

8. Execution phase

Superpowers executing-plans (70 lines): load plan → TodoWrite → execute steps → hand off to finishing-a-development-branch. Recommends subagent-driven-development if subagents are available.

ed3d executing-an-implementation-plan (562 lines) hard-codes a multi-agent fix loop per task:

1. task-implementor-fast implements with TDD and commits
2. code-reviewer checks tests/build/lint/quality
3. task-bug-fixer resolves Critical/Important/Minor issues
4. Re-review until zero issues
5. After all tasks: project-claude-librarian updates CLAUDE.md, test-analyst validates coverage against acceptance criteria, generates human test plan

9. Prompt-engineering style

Superpowers uses one <HARD-GATE> block and a brief "Anti-Pattern" callout, then flows into normal prose.

ed3d skills each include a "Common Rationalizations — STOP" table (8–11 rows) and frequent YOU MUST formatting. Lots of repetition across skills designed to defeat skipping.

10. What ed3d removed

Notable omissions from superpowers when ed3d forked:

• Visual Companion (browser-based mockups) — gone
• Sub-project decomposition guidance inside brainstorming — moved out (decomposition is now "thinking work you do before the plugin")
• The <SUBAGENT-STOP> early-exit in using-superpowers — ed3d's equivalent using-plan-and-execute doesn't have it
• dispatching-parallel-agents skill — ed3d uses dedicated named subagents instead

The core thesis difference

Superpowers treats brainstorming as one human-in-the-loop conversation that produces a spec, and trusts a single executor (or subagent dispatcher) to ship it.

ed3d treats the workflow as a pipeline of role-separated subagents, each with its own clean context, with the human as gatekeeper at phase boundaries. The verbosity inflation isn't redundancy — it's the cost of (a) every skill having to re-establish protocol because phases run in fresh /clear'd contexts, and (b) making each subagent role's rules independently auditable.