Install superpowers plugin for structured development workflows¶
- Date: 2026-04-09
- Status: Accepted
- Tags:
tooling,claude,workflow
Context¶
During onsite onboarding, Everett — an accomplished engineer deeply versed in AI tooling — recommended the superpowers plugin alongside rlm-claude-code (ADR-0022) as part of his standard Claude Code setup. Superpowers enforces structured workflows for planning, implementation, debugging, and code review rather than letting Claude freeform its way through complex tasks.
The core problem: Claude Code is capable but unstructured. Without guardrails, it skips planning, jumps to implementation, misses edge cases, and claims work is done without verification. Superpowers addresses this by injecting discipline at each phase of development.
Everett's Recommended Workflow (AIDLC)¶
Everett's daily pattern chains superpowers and RLM together in a deliberate pipeline:
-
Brainstorm —
/superpowers:brainstormto explore the problem space, surface missing requirements, and produce a structured spec. Example: identifying that an identity-service was built from the admin side but completely lacked the participant journey (self-registration, program enrollment, participant roles). -
Research via RLM — Superpowers recognizes when the problem needs deeper context and offers to dispatch the
rlm-claude-code:rlm-orchestrator. The orchestrator performs broad codebase research (53+ tool uses, ~100K tokens) to map existing apps, services, and architecture before planning begins. -
Plan — Based on brainstorm spec + RLM research findings, superpowers generates a concrete implementation plan with prioritized phases and review checkpoints.
-
Execute — Subagent-driven or inline execution of the plan, with verification gates before claiming completion.
-
Deliver —
/deliveror/buildto ship — finishing the branch, running verification, and opening a PR.
The key insight: brainstorm feeds RLM research, research feeds planning, planning feeds execution. Each phase's output is the next phase's input. Skipping a phase degrades everything downstream.
Key Capabilities¶
- Brainstorming — mandatory design exploration before creative work; produces a spec
- Plan writing — converts specs into step-by-step plans with review checkpoints
- Execution modes — inline or subagent-driven, with parallel agent dispatch for independent tasks
- Systematic debugging — structured root cause analysis before proposing fixes
- TDD — test-first workflow enforced before implementation code
- Verification — requires running actual commands and confirming output before success claims
- Code review — requesting and receiving review with technical rigor over performative agreement
- Git worktrees — isolated feature branches with smart directory selection
Decision¶
Install the superpowers plugin from the official Claude Code plugin marketplace at user scope. Use the default configuration (skills auto-activate based on context). Adopt Everett's AIDLC workflow as the recommended pattern for non-trivial feature work.
Installation method: Official Claude Code plugin marketplace (claude plugin marketplace add claude-plugins-official, then install superpowers).
Version at adoption: 5.0.7
Workflow integration with RLM (ADR-0022):
- Start with
/superpowers:brainstormto explore and spec - Dispatch
rlm-claude-code:rlm-orchestratorfor deep codebase research when the problem spans multiple services/files - Use superpowers plan-writing to structure findings into executable steps
- Execute via subagent-driven development or inline, with verification gates
- Finish with
/superpowers:finishing-a-development-branchor custom/git-ship
Alternatives Considered¶
| Option | Why Rejected |
|---|---|
| Custom skills replicating the same workflows | Significant maintenance burden; superpowers is actively developed (5.0.x release cadence) with community contributions and cross-platform support |
| No structured workflow enforcement | Claude skips steps, claims completion without verification, and produces lower-quality output on complex tasks |
| Only use brainstorming + planning (subset) | The skills work as a cohesive system — verification and review close the loop that planning opens; RLM integration needs the full chain |
| Use RLM alone without superpowers | RLM handles context decomposition but doesn't enforce workflow discipline — you still need the brainstorm → plan → verify pipeline |
Consequences¶
Positive:
- Enforces plan-before-code discipline across all Claude Code sessions
- Brainstorming catches missing requirements early (e.g., the participant journey gap Everett surfaced)
- RLM orchestrator integration enables deep research before planning — plans are grounded in actual codebase state
- Verification step prevents false "done" claims — evidence before assertions
- Subagent-driven development enables parallel task execution
- Cross-platform: Claude Code, Cursor, Windsurf, Codex, OpenCode
Negative / Trade-offs:
- Adds overhead to simple tasks (mitigated by skipping skills for trivial operations)
- Session-start hook adds bootstrap context to every session
- Dependency on third-party plugin with active release cadence
- Some skills overlap with existing custom workflows (e.g., git-ship vs finishing-a-development-branch)
- Full AIDLC pipeline can be token-heavy (~100K+ for research phase) — appropriate for features, overkill for bug fixes
Ongoing obligations:
- Track upstream releases for breaking changes or new skills
- Review skill overlap with custom skills periodically — consolidate where appropriate
- Monitor token usage on the brainstorm → RLM research → plan pipeline
- Document AIDLC workflow pattern for team reference
ADRs are immutable. If this decision changes, create a new ADR with status "Superseded by" pointing here, and update this record's status.