Praxis - Plugin Claude Code para Desenvolvimento
moiri-gamboni/praxis
Workflow estruturado de dev com agents paralelos, TDD e revisão automática
Ver no GitHubPlugin para Claude Code que fornece skills estruturadas para todo ciclo de desenvolvimento: ideação, design, prototipagem, implementação e revisão de código. Inclui agentes especializados para debugging sistemático, TDD, revisão multi-dimensional e workflows Git automatizados, com commands como /design, /implement, /review e /ship.
Instalação
npx convez-ai moiri-gamboni/praxis1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Ciclo completo de desenvolvimento estruturado (ideate → design → implement → review → ship)
- Debugging sistemático e test-driven development com validação automática
- Revisão de código multi-dimensional com agentes especializados em paralelo
- Prototipagem rápida com decisões arquiteturais documentadas
Veredito

Ferramenta robusta e bem-pensada que estrutura workflows complexos de desenvolvimento. Recomendada para equipes que buscam rigor metodológico e automação de processos críticos (TDD, code review, git); exige conhecimento técnico aprofundado e integração com Claude Code.
README
A tradução para o português deste README ainda está na fila.
Praxis
A Claude Code plugin for software development. Skills teach Claude structured approaches to ideation, design, prototyping, implementation, and code review. They invoke as slash commands (/design, /implement, /review, /ship, etc.) and auto-trigger when relevant. Specialized agents handle parallel exploration, adversarial design review, multi-dimensional code review, and plan-document review.
Praxis incorporates material from superpowers, Anthropic's claude-plugins-official (feature-dev, pr-review-toolkit, commit-commands, frontend-design), and Claude Code's built-in skills, picking the best version of each component and filling gaps between them.
Installation
# Add the marketplace
/plugin marketplace add moiri-gamboni/praxis
# Install the plugin
/plugin install praxis@praxis-marketplace
Updates are applied automatically, or manually with /plugin marketplace update.
Skills
Per Claude Code's unified skills/commands architecture, every skill is also a slash command. Auto-activating skills (no manual trigger needed) and slash-invoked skills are listed together.
| Skill | Trigger / Use |
|---|---|
| ideate | Auto-activates when in problem-space exploration: figuring out what to build, whether to build it, what existing solutions cover the problem |
| systematic-debugging | Auto-activates on bugs, test failures, unexpected behavior |
| test-driven-development | Auto-activates when implementing a feature or bugfix |
| verification-before-completion | Auto-activates when about to claim work is complete or passing |
| receiving-code-review | Auto-activates when receiving code review or red-team feedback |
| frontend-design | Auto-activates when building web components, pages, or applications |
/prototype <what> | Build the first working version fast but sound: understand (codebase + prior art via parallel subagents) → judicious tech choices in a decision record → confirm gate → thin end-to-end slice built to an extensible standard → independent verification → handoff. For MVPs, POCs, spikes, demos, timed tasks. Also auto-activates. |
/design <feature> | Architecture design: 4-dim shared exploration → competing architects → synthesis matrix → red-team fleet. Reads plans/<slug>-ideation.md if present. Writes plans/<slug>.md. |
/implement <task or plan> | Parallel work orchestration: decompose, spawn sub-agents in worktrees (each with TDD + review + simplify gates), then integrate with spec-reviewer + verification-before-completion. |
/review [git-range] | Multi-wave code review by logical units: per-unit deep review with full reviewer fleet → cross-unit boundary review → verification pass on Critical findings. |
/simplify [scope] | Simplification pass on recently modified code. |
/ship [merge] [test cmd] | PR-first shipping. State-driven default: opens or updates a PR. With merge arg: explicit local merge after acceptance prompt. |
/clean-gone | Delete local branches whose remote counterpart is gone, plus their worktrees. Auto-fires after /ship merge and at the end of /implement. |
Skills chain naturally: each suggests a next step based on context. The intended pipeline is ideate → /design → /implement → /review → /ship; /prototype is the fast-path alternative when the goal is a working MVP or spike rather than a production-bar build.
Agents
| Agent | What it does | Invoked by |
|---|---|---|
| code-explorer | Traces execution paths, maps architecture, documents dependencies. Dimensional invocation in /design Phase 1.2 (architectural fit / touchpoints / risks-deps / constraints). | /design Phase 1.2, ideate, manual |
| code-architect | Designs implementation blueprints. Dispatched in parallel with different philosophies (minimal / clean / pragmatic) from /design Phase 1.3. | /design Phase 1.3 |
| red-team | Adversarially challenges design. Dispatched as a fleet in /design Phase 1.5 (architectural soundness / failure modes / operational / hidden complexity / scope-assumptions / conditional security / conditional doc-currency). Per-finding confidence scores + anti-complexity bias. | /design Phase 1.5 |
| plan-doc-reviewer | Reads plan + ideation file, returns Approved or Issues. Independent second pair of eyes on plan completeness, spec alignment, buildability. Calibrated to flag only real implementation problems. | /design Phase 3 self-review |
| code-reviewer | Reviews code against project guidelines and plans, with confidence scoring (>= 80 threshold) and anti-complexity bias on proposed fixes. | /review Wave 1, Wave 2 cross-unit |
| spec-reviewer | Verifies implementation matches a specification (which can be a /design plan file). Skeptical, independent reading. | /implement Phase 4 (when plan file present), /review Wave 1, manual |
| code-simplifier | Simplifies code while preserving functionality. Direct-modify in /simplify standalone, advisory in /review. | /simplify, /review Wave 1, /implement workers + Phase 4 |
| implementer | Single-unit worker for /implement. Bakes the worker procedure (skill loop, push, log, audit) into its system prompt so compliance is structural. Does not further delegate. | /implement Phase 2, manual single-unit builds |
| comment-analyzer | Checks comment accuracy and long-term maintainability. Anti-complexity: removing bad comments preferred over adding obvious ones. | /review Wave 1 |
| test-analyzer | Reviews test coverage quality, prioritizing behavioral coverage. Each proposed test must articulate failure scenario + likelihood + consequence. | /review Wave 1 |
| silent-failure-hunter | Finds swallowed errors and inadequate error handling. Each finding must articulate the actual failure mode, not abstract concern. | /review Wave 1 |
| type-analyzer | Evaluates type design: encapsulation, invariants, enforcement. Each improvement must articulate specific bug class prevented vs cost. | /review Wave 1, Wave 2 cross-unit |
All agents run on Opus. Reviewer agents use a per-finding confidence threshold (>= 80) and reject fabricating findings to look thorough.
Design
Ideation produces a file /design reads. ideate covers problem-space exploration (whether to build, what already exists, build-vs-buy via prior-art search). It writes plans/<slug>-ideation.md with problem statement, prior art investigated, alternatives considered at concept level, and key constraints. /design reads this as foundational context, gating on its presence ("run ideate first or proceed without prior-art search?").
Two-axis design space in /design. Phase 1.2 dispatches 4 code-explorer agents along evaluation dimensions (architectural fit, touchpoints, risks/dependencies, constraints) — exploration topics. Phase 1.3 dispatches 2-3 code-architect agents along approaches (minimal-changes, clean, pragmatic) — strategic stances. Phase 1.4 synthesis builds a matrix (dimensions × approaches), classifies decisions as Sensitivity Points (one-dimension impact) or Tradeoff Points (multi-dimension opposing), produces one merged plan with structured Alternatives Considered.
Adversarial review fleet. Phase 1.5 dispatches red-team agents in parallel along attack angles (architectural soundness, failure modes, operational concerns, hidden complexity, scope/assumptions; conditional security; conditional documentation currency with web-tool verification of named libraries). Per-finding confidence scores. Optional verification pass on Critical findings. Resolution Log enforces explicit Fixed/Rejected/Deferred per finding — silent acceptance not allowed.
Plan-doc reviewer. After Phase 3 self-review, plan-doc-reviewer reads the plan + ideation file independently and returns Approved or Issues. Calibrated to only flag real implementation problems.
Hybrid file-writing. Explorer and architect agents write detailed outputs to plans/<slug>/.workspace/exploration/<dimension>.md and plans/<slug>/.workspace/architects/<approach>.md, returning summary + path. Coordinator context stays light.
Parallel implementation. /implement decomposes work into independent units, spawns workers in isolated worktrees (each using TDD + review + simplify + verification gates). Workers write structured logs to <workspace>/workers/<unit>.md. Team lead merges incrementally, runs cross-cutting /review + /simplify + spec-reviewer against the plan + verification-before-completion. Calls /clean-gone to sweep stale state at the end.
Multi-wave code review. /review identifies logical code-path units (not files), dispatches the full reviewer fleet per unit (scaled to complexity), runs a cross-unit boundary review, then a verification pass that re-runs each Critical finding through a fresh second agent. Confirmed/Disputed labels; never auto-drop disputed findings.
PR-first shipping. /ship is state-driven: on main fast-paths to PR creation; on feature branch opens a PR or pushes update to existing one with summary confirmation. /ship merge runs an explicit local merge with typed acceptance.
Example Workflows
Large feature (parallel)
- Conversational ideation (ideate skill activates) — produces
plans/<slug>-ideation.md /designreads ideation, runs shared exploration + architect approaches + red-team fleet, writesplans/<slug>.md/implement plans/<slug>.mdspawns sub-agents in worktrees: each does TDD + review + simplify + verification + commit + push + log- Coordinator merges incrementally, then cross-cutting
/review+/simplify+spec-revieweragainst plan +verification-before-completion, opens PR - PR feedback iteration via manual edits +
/shipto push updates
Bug fix (solo)
- systematic-debugging skill activates automatically on the bug report
- test-driven-development skill activates (write failing test for the bug)
- Fix the bug; verification-before-completion confirms tests pass
/shipto open a PR
Before opening a PR
/review(multi-wave by logical units)- Fix Critical confirmed findings
/simplify(cross-cutting)/shipto open PR
Upstream Tracking
Praxis tracks changes to its source plugins and can automatically incorporate improvements via PR.
# Sync upstream and analyze changes
scripts/analyze-upstream.sh
# Non-interactive (for cron)
scripts/analyze-upstream.sh --auto
# Sync only (no analysis)
scripts/sync-upstream.sh
The upstream branch stores verbatim copies of the source plugins. upstream.json maps each praxis file to its source(s) with adaptation level and tracks per-source commit hashes for incremental analysis.
License
AGPL-3.0. See LICENSE and NOTICE for details.
Incorporates MIT-licensed material from superpowers by Jesse Vincent. Components from Anthropic's claude-plugins-official: frontend-design is Apache-2.0 licensed; other components had no license specified as of 2026-02-16.