Workflow Skills para Claude Code
bestdan/workflow-skills
Revisão de PRs + planejamento documentado + análise auditável para Claude Code
Ver no GitHubConjunto de skills de engenharia para fluxos de trabalho colaborativos em Claude Code, incluindo revisão de PRs com múltiplos agentes, planejamento multi-etapas durável e pipelines de análise auditáveis. Oferece automação de tarefas de revisão de código, planejamento documentado e loop de tarefas nativo do repositório.
Instalação
npx convez-ai bestdan/workflow-skills1 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
- Revisar pull requests com colaboração de múltiplos agentes (Gemini, Codex, etc)
- Planejar e executar tarefas multi-etapas com documentação persistente
- Realizar análise quantitativa com fact-checking independente
- Capturar e processar trabalho de acompanhamento em loop nativo do repositório
Veredito

Skill robuusta para engenheiros que trabalham com Claude Code e precisam de revisão de código estruturada e planejamento auditável. Requer aprovação de permissões e integração com GitHub, mas oferece automação real de fluxos de trabalho complexos com múltiplos agentes.
README
A tradução para o português deste README ainda está na fila.
workflow-skills
A Claude Code plugin bundling Daniel's general engineering workflow skills: collaborative PR review, durable multi-step planning, auditable quantitative analysis pipelines with independent fact-checking, and a repo-native task loop for capturing and processing follow-up work.
Install
/plugin marketplace add bestdan/workflow-skills
/plugin install workflow-skills@workflow-skills
What's in the box
7 skills, 7 commands, and 1 subagent, organized into four workflows:
PR review
| Skill | Trigger | What it does |
|---|---|---|
| co-review | /co-review [PR# | --local | --remote | --post] [--base <branch>] | Produce your own review of a PR, optionally pull in other local agents (gemini, codex, …) as extra reviewers, reconcile everything against existing GitHub bot/human comments via an independent subagent, auto-fix high-confidence items, surface judgment calls back to you, and — when reviewing your own PR — commit and push the fixes. --local reviews local changes (committed + uncommitted) with no PR; --remote skips local agents for a plain PR review; --post reviews someone else's PR and, after you vet the findings, posts them back as a batched GitHub review instead of editing local files. |
Approve the reviewer commands once. co-review sends the rubric, your extra requests, and the diff to local reviewers on stdin with a fixed prompt argument, so the command never changes — you approve each reviewer once with an exact-match, read-only rule (no broad wildcard). Merge them into the permissions.allow array in ~/.claude/settings.json (or the repo's .claude/settings.json) — don't paste over an existing file. The co-review skill's Permissions section has the exact JSON to copy.
Planning
| Skill | Trigger | What it does |
|---|---|---|
| plan-with-docs | /plan-with-docs, or after approving a plan in plan mode | Write a multi-step implementation plan as markdown files under dev_docs/tasks/<name>_plan/ (one file per PR-sized task, in the canonical task format) instead of printing it inline, then refine the plan through clarifying questions. Default for plans >3 tasks or spanning multiple PRs. |
Task loop
Capture follow-up work with full context during development, then process it automatically. The task skill auto-triggers when you mention deferred work; the commands handle capture, listing, processing, and destination configuration.
| Command / Skill | Trigger | What it does |
|---|---|---|
/add-task | /add-task [description] | Capture follow-up work and deliver it via the configured handler. Defaults to the repo-pr handler: dispatches a remote agent to commit the task as a markdown file on a branch from main and open a PR, with zero local impact. |
/do-tasks | /do-tasks [slug | --all | -n N] [--remote|--local] | The single execute verb. Resolves the handler and turns dependency-ready tasks into PRs. For repo-pr: default = highest-ranked task; --all / -n N batch dispatch (each to its own cloud VM) bounded by wip_limit; --local runs in-session. For linear: claims and executes one issue in the current session (single/foreground, pre-claim wip_limit gate). For gh-issue: the same single/foreground claim+execute over the gh CLI. For jira: the same single/foreground claim+execute over the Atlassian MCP (needs ready_status set). |
/list-tasks | /list-tasks [status] | Show a table of all tasks in dev_docs/tasks/ with status, priority, size, dependency blockers, tags, and expiry. Filter by status (unclaimed, claimed, blocked, expired, all). |
/push-plan | /push-plan <name> [--ready-only] | Push a vetted local plan (dev_docs/tasks/<name>_plan/) to the configured tracker. repo-pr is a no-op (the files are the destination); linear creates one issue per task under a project, in dependency order, translating is_blocked_by to native blockers. Create-missing-only — records tracker ids back into the files, so re-runs don't duplicate. jira/gh-issue not supported yet. |
/task-config | /task-config [handler] | Configure where /add-task delivers (writes dev_docs/tasks/.task-config.yml). Handlers: repo-pr (default, markdown PR), gh-issue (GitHub Issue), jira (Jira work item via Atlassian MCP), linear (Linear issue via Linear MCP). Verifies prerequisites before writing the config. |
/promote-tasks | /promote-tasks [dry-run | apply] | Score new tasks against the confidence check and promote them to ready or needs_refinement. |
/doctor | /doctor [--fix] | Diagnose the task-loop setup — config validity, handler prerequisites, legacy dirs, schema drift, and hygiene — reporting PASS/WARN/FAIL per check. Read-only by default; --fix applies the safe mechanical repairs (run the migration, prune expired, fill defaulted fields) and leaves judgment calls as warnings. |
| break-down-task | a task is too big for one PR, /promote-tasks flags scope exceeds size 5, or "split / slice / break down this task" | Take an existing over-sized task, find natural shear points (vertical/horizontal slices, preparatory refactor, interface-then-impl), and on a HIGH-confidence split replace it with PR-sized component tasks chained by is_blocked_by. The resolution of the promote-tasks size gate; sibling of plan-with-docs (which slices a fresh idea). |
task skill | mentions of "task", "todo", "follow-up", "we should come back to this", or any of the /add-task / /do-tasks commands | Auto-trigger skill: provides the task file format, capture workflow, handler abstraction, and processing logic. Loaded in-context whenever Claude sees deferred-work language. |
Auditable analysis
| Skill | Trigger | What it does |
|---|---|---|
| analysis-pipeline | /analysis-pipeline, or auto when building report-producing analyses | Structure quantitative work as model → template → fill pipelines. Enforces input provenance, separation of narrative and computation, and reproducible output. Includes a working example/ to copy from. |
| review-facts | /review-facts | Independent fact-check of a completed analysis. Spawns the bundled fact-reviewer subagent (fresh context, read-only tools) to verify links resolve, cited values match sources, output is reproducible, narrative numbers trace to the model, units/formulas are correct, and the recommendation matches the data. |
| analysis-conventions | Auto when writing notebooks or analysis scripts | Coding conventions for analysis work: when to use plain scripts vs marimo, uv run for execution, scratchpad patterns, notebook structure. Not user-invocable — loads in context. |
Bundled subagent
- fact-reviewer — Read-only auditor used by
/review-facts. Runs in a fresh context withRead,Glob,Grep,Bash, andWebFetch. Reports findings; never edits the analysis.
Why auditable analysis matters
When you ask Claude to analyze something — compare vendors, project costs, evaluate options — the default behavior is to do the math in its head and write the answer directly into prose:
Based on the pricing, Vendor A costs $2,510/year, Vendor B costs $2,627/year, and Vendor C costs $2,197/year. I recommend Vendor C, saving $430/year.
This looks helpful but is fragile: you can't tell where the numbers came from, you can't update them when inputs change, and you can't verify the math. The analysis-pipeline and review-facts skills push Claude to structure analyses as reproducible pipelines instead — every number traces back to a named, sourced input, and a separate agent audits the result before it ships.
Example
skills/analysis-pipeline/example/ contains a complete working pipeline (vendor cost comparison) you can run:
cd skills/analysis-pipeline/example
uv run model.py # compute all values -> model_output.json
uv run fill_templates.py # fill template placeholders -> memo.filled.md
Then audit it:
/review-facts .
Staying up to date
Third-party marketplaces have auto-update disabled by default. Enable auto-update in the /plugin UI (Marketplaces tab), or update manually:
/plugin marketplace update bestdan/workflow-skills
/reload-plugins
Development
Run the full quality gate locally with just check (formatting, manifest
validation, and repo-native structural checks). See CONTRIBUTING.md
for the dev loop, the "adding a skill" checklist, and the opt-in behavioral evals.
License
MIT