Feature Skills para Claude Code
nigelmcnie/feature-skills
Orquestra o workflow completo de desenvolvimento de features no Claude Code
Ver no GitHubConjunto de skills para Claude Code que automatiza o fluxo completo de desenvolvimento de features: requisitos → planejamento → implementação → review → iteração. Cada skill é especializado em uma fase do workflow, com suporte a captura de contexto, escolha de prioridades e documentação em HTML.
Instalação
npx convez-ai nigelmcnie/feature-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
- Estruturar e documentar requisitos de novas features
- Gerar planos de implementação com fases e PRs
- Automatizar o ciclo de review e feedback iterativo
- Manter rastreamento centralizado de features por projeto
- Capturar ideias e contexto para trabalho futuro
Veredito

Skill bem pensada para equipes que usam Claude Code em workflow estruturado. A arquitetura modular (router + sub-skills) é adequada, mas requer instalação manual e configuração por projeto. Ideal para times que querem formalizar o processo de feature development com IA.
README
A tradução para o português deste README ainda está na fila.
feature-skills
Claude Code skills for a feature-development workflow: requirements → plan → implement → review → iterate.
The skills
| Skill | What it does |
|---|---|
feature | Workflow router; detects state and delegates to the right sub-skill |
feature-context | Capture an idea for later as a context.html |
feature-choice | Suggest which feature to work on next from the tracker |
feature-requirements | Draft and review requirements |
feature-plan | Draft and review an implementation plan |
feature-implement | Implement one phase per MR |
feature-qa | Run the project's quality control checks |
feature-review | Review the merged implementation on main |
feature-iterate | Address review feedback |
Most skills set disable-model-invocation: true and are invoked via slash
commands (/feature-plan, etc.). feature-context and feature-choice
are the exceptions — they're auto-invokable when the user asks to
capture an idea or pick what to work on next.
Install
mkdir -p ~/src/nigelmcnie
cd ~/src/nigelmcnie
git clone git@github.com:nigelmcnie/feature-skills.git
cd feature-skills
./bin/install-symlinks
The install script symlinks each skill dir from this repo into
~/.claude/skills/, where Claude Code discovers them. Re-running the
script is safe.
Typical flow
/feature # router, jumps to the right step
/feature-context <name> # capture an idea for later
/feature-choice # ask what to work on next
/feature-requirements <name> # draft requirements
/feature-plan <name> # plan it
/feature-implement <name> # implement one phase (re-invoke per phase)
/feature-review <name> # review after all phases merged
/feature-iterate <name> # address review feedback
/feature picks up where you left off if you don't know the right
sub-skill to invoke.
Where docs live
Each feature's docs live in a developer-scoped store, separate from the target repo:
~/.claude/feature-docs/<project>/<feature>/
context.html
requirements.html
plan.html
requirements-feedback-N.html # transient, archived after integration
review-feedback-N.html # transient, archived after integration
.feedback-archive/ # synthesis docs archived here
~/.claude/feature-docs/<project>/
features.html # project-level tracker
<project> is derived as basename $(git rev-parse --show-toplevel).
The HTML is canonical. Markdown is generated by bin/feature-html-to-md
when a repo opts in via .feature-workflow.toml (below). Markdown is a
read-only snapshot — no checkboxes, no interactive bits.
Per-repo export config
A .feature-workflow.toml at the repo root tells the skills which
artifacts to export back into the repo, and in what format:
[export]
context = "markdown" # "html" | "markdown" | "none"
requirements = "markdown"
plan = "markdown"
features = "markdown"
markdown→ the skill runsbin/feature-html-to-mdafter writing the HTML, producingdocs/features/<feature>/<artifact>.md(orfeatures.mdat the repo root for the tracker).html→ the HTML is copied verbatim todocs/features/<feature>/.none→ nothing is exported.
A missing config file means none for every artifact (clean repo, no
workflow files committed). Commit the config or gitignore it — the
repo's choice.
Markdown export script
bin/feature-html-to-md <source.html> <target.md> reads
<meta name="feature-doc-type" content="…"> from the source HTML to
dispatch to a per-type converter. Doc types: context, requirements,
plan, features. Synthesis docs are never exported (they're
transient).
The script:
- Strips UI chrome (TOC sidebar, comment widgets, sticky footer, comments rail, scripts, styles).
- Drops
<input type="checkbox">from plan checklists — markdown carries the steps as plain bullets. - Converts
<pre><code class="language-X">to fenced code blocks. - Preserves prose, tables, and links.
Runs as a PEP 723 script via uv run; uv must be on PATH. First
invocation cold-caches beautifulsoup4; subsequent runs are fast.
Design notes
- Cross-skill conventions live in the doc templates at
feature/*-template.html. Structural guidance for each artifact type (context, requirements, plan, features tracker, feedback synthesis) is carried in the HTML template comments; load-bearing principles (Requirements-vs-plan, Indicative notes convention, triage tiers for synthesis docs, etc.) are inlined into each skill'sSKILL.md. - Reviewer subagents are spawned at requirements / plan / review / iterate.
All skills here set
disable-model-invocation: trueso they can use the Agent tool (subagents can't spawn subagents). - Feedback synthesis docs are archived locally to
.feedback-archive/rather than deleted — locally gitignored. - Branch conventions: requirements + plan commit to
main. Implementation branches offmaintofeatures/<feature>-p<N>. Review runs onmainafter all phase MRs land.