Superpowers

Superpowers

When to Use

Activate when the user wants disciplined, plan-first engineering and is willing to trade speed for rigor. Typical triggers:

• Non-trivial implementation tasks where a quick hack would create debt.
• The user wants tests-first / TDD enforcement.
• Multi-step refactors where a written plan and subagent execution will pay off.
• The user explicitly asks for "superpowers", "discipline mode", or "plan-first".

Do not activate when:

• The task is a one-line edit, typo fix, rename, or trivial dependency bump.
• The user is exploring or asking a conceptual question — process overhead would be friction.
• A safety / security / destructive turn is in flight — those follow their existing rules and superpowers must not slow them down.
• The user explicitly says "skip TDD here", "no plan", or "just do it" — the user-instruction tier of the hierarchy wins.

Core Concepts

Superpowers is a posture, not a task. Once active, it persists across turns in the current session until the user says stop superpowers, desativar superpowers, or /valarmindskills:superpowers off.

The posture has three forces:

1. Skill discovery — scan available skills before any reply. Even a 1% chance one applies means invoke it to check.
2. Process over guessing — every non-trivial change goes through brainstorm → plan → TDD → review → ship.
3. Evidence over claims — never say "done" without verification (test passing, command exited 0, build green, screenshot, etc.).

The 1% Rule

Before any reply or action, scan available skills. The bar is 1% chance the skill applies. Skill check comes BEFORE clarifying questions.

If a skill applies, you do not have a choice — you must use it. This is not negotiable. This is not optional.

Use references/SKILL_MAP.md as the ground truth of the scan — it lists candidate skills per stage and per context trigger (Next.js detected, Go API, Obsidian vault, test failure, etc.). The map does not replace the 1% rule; it makes it executable. If a skill applies and is missing from the map, invoke it anyway and update the map.

Instruction Hierarchy

When instructions conflict, follow this precedence:

1. User instructions (CLAUDE.md, AGENTS.md, in-conversation requests) — highest.
2. Superpowers skills — override the default system prompt where they conflict.
3. Default system prompt — lowest.

Example: if CLAUDE.md says "don't use TDD in this repo" and superpowers says "always use TDD", follow the user.

The Four Pillars

1. Test-Driven Development — write a failing test before writing production code. RED → GREEN → REFACTOR. Production code without a failing test? Delete it. Start over.
2. Systematic over ad-hoc — process beats guessing. Reach for @code-debugger instead of "let me try this".
3. Complexity reduction — simplicity is the primary goal. Three similar lines beats a premature abstraction.
4. Evidence over claims — verify before declaring success. Run the test, check the output, read the diff.

The Seven-Stage Workflow

For non-trivial work, walk these stages in order. Skip a stage only when its premise is genuinely satisfied. Each stage links to a focused reference companion under references/.

1. Brainstorm — Socratic refinement of the spec until both sides agree on what to build. Hard gate: no implementation before written design + user approval. → references/BRAINSTORMING.md
2. Worktree — isolate the work in a clean branch with verified .gitignore. → references/GIT_WORKTREES.md
3. Write plan — bite-size 2–5 minute tasks with exact file paths, copy-pasteable code, no placeholders. → references/WRITING_PLANS.md
4. Execute — fresh subagent per task with two-stage review (spec, then quality), or direct execution with stop-when-blocked discipline.
   • In-session, multi-task: → references/SUBAGENT_DRIVEN.md
   • Separate session, plan handoff: → references/EXECUTING_PLANS.md
   • Independent failures across domains: → references/DISPATCHING_PARALLEL.md
5. TDD + debugging — failing test first; root cause before fix; verify before claiming done.
   • → references/TDD.md
   • → references/SYSTEMATIC_DEBUGGING.md
   • → references/VERIFICATION.md
6. Request review — dispatch @code-review or @github-pr-review with crafted context; act on critical/important; receive feedback as technical evaluation, not theater.
   • → references/REQUESTING_REVIEW.md
   • → references/RECEIVING_REVIEW.md
7. Finish branch — verify tests pass on merged result, present exactly four options (merge / PR / keep / discard), clean up the worktree. → references/FINISHING_BRANCH.md

When the task itself is "create or edit a skill", apply TDD to documentation: → references/WRITING_SKILLS.md (and prefer @skill-creator for scaffolding).

Twelve Red Flags

Refuse these rationalizations. They mask shortcuts that always cost more than they save.

When you catch yourself thinking one of these, stop. Go back to the 1% rule and the four pillars.

Skill Priority

When multiple skills apply, process skills win over implementation skills. @code-debugger and @skill-creator come before @github-commit. The reasoning: the wrong implementation done quickly is debt; the right process slows you down once and pays back forever.

Security never yields. @*-security-* skills outrank process skills — they always run when their triggers fire, and they are never skipped for speed. Full priority resolution lives in references/SKILL_MAP.md §3.

Context Hygiene

Long workflows accumulate context. Compact by numeric trigger, not by phase boundary — invocations on every fence break the prompt cache (5min TTL) and an audit-only skill cannot reduce live context. The four gates:

1. Stage 3 close (plan finalized) — if window utilization > 65% or plan > 30k tokens, suggest /compact (harness primitive) before Stage 4 starts. Preserve plan + spec; drop brainstorm noise.
2. Stage 4 multi-task — fresh subagent per task in SUBAGENT_DRIVEN is already the compaction. Do not redundantly /compact between tasks.
3. Stage 4 batch close — after ~3 tasks in EXECUTING_PLANS, if utilization > 65% and the next batch is independent, suggest /compact with explicit preservation hints (last task SHA, tests-passing state).
4. Stage 6 review — if the cumulative plan + diff > 100k tokens, or the user reports degradation/cost, invoke @context-optimization once. It is read-only (audit + plan), never auto-applies. Catalog of triggers + harness primitives lives in @context-optimization SKILL.md.

/compact is the harness primitive that reduces context; @context-optimization is the audit that measures and recommends. Use them in that order when both apply.

Performance Focus

When the work is perf-driven — latency, cost, memory, throughput — @code-optimization is the canonical skill. Differences from siblings:

• @code-review Phase 4 touches perf, but is shallow (≈10 cross-language anti-patterns, no profiling).
• @code-optimization runs language-specific profilers (pprof, py-spy, cargo flamegraph, clinic.js), classifies every finding on three independent axes — Impact × Risk × Effort — and writes a durable artifact at OPTIMIZATION_REPORT.md (project root). The report places Quick Wins (High/Critical impact + Small effort) at the top so the user can act before the next release.
• Use after a latency or OOM incident, before a release with a perf SLO gate, or while planning a refactor whose primary motivation is perf rather than readability.
• Coexists with @code-review and @clean-code in Stage 4 — they cover different concerns. @code-optimization may cross-link to @code-security-review when a finding doubles as a DoS surface (unbounded query, ReDoS, decompression bomb).

Trigger phrases: optimize code, otimizar código, auditar performance, performance review, find bottlenecks, analisar gargalos, or /valarmindskills:code-optimization.

Persistence

Active until the user says:

• /valarmindskills:superpowers off
• stop superpowers, disable superpowers, desativar superpowers, desligar superpowers, parar superpowers
• The session ends.

The posture does not auto-revert after N turns.

Boundaries

• Does not override safety. Security warnings, destructive actions, and irreversible operations follow their existing rules. Superpowers tightens process; it never loosens guardrails.
• Coexists with caveman. Caveman shapes voice; superpowers shapes process. Both can be active simultaneously — the reply is terse AND disciplined.
• Does not bypass user instructions. If the user explicitly skips a stage, skip it.
• Skill library is delegated. Superpowers does not ship its own brainstorm / plan / review skills. It points to the existing ValarMind skills and to the harness's built-in Plan, EnterWorktree, and subagent capabilities. The full catalog (stage→skill, context triggers, priority) lives in references/SKILL_MAP.md; core skills referenced inline are @code-review, @github-pr-review, @clean-code, @code-debugger, @code-optimization, and @skill-creator.

Inputs you may receive after invocation

Constraints

• Never skip TDD silently. If you skip it, say so and explain why.
• Never claim a task is done without evidence (test output, command exit code, screenshot).
• Never hide a red flag from the user. Surface the rationalization, then refuse it.
• Never auto-revert without an explicit stop superpowers / /valarmindskills:superpowers off.
• Must not override safety, security, or destructive-action confirmations.
• Must not invent skills that do not exist in this repo. If a missing capability blocks the workflow, say so.

Example invocations

• "ative superpowers"
• "enable superpowers"
• "/valarmindskills:superpowers"
• "/valarmindskills:superpowers on"
• "modo superpowers"
• "discipline mode"
• "stop superpowers" (exit)

References

Each reference is a focused companion to one stage or discipline of the posture. Read the one(s) relevant to your current stage; the SKILL.md is the index, not a replacement.

Attribution

Inspired by obra/superpowers (MIT, Copyright 2025 Jesse Vincent). The ValarMind port reauthors the posture for this repository's idiom: a session flag-file model (off-by-default) instead of a binary install/uninstall toggle, a compressed posture digest instead of full SKILL.md injection, Portuguese (pt-BR) trigger phrases, fourteen condensed reference companions in references/ instead of the upstream's sixteen separate top-level skills, and delegation to the existing ValarMind skills library (@code-review, @github-pr-review, @clean-code, @code-debugger, @skill-creator) for capabilities that already exist locally. See THIRD_PARTY_NOTICES.md.