Research Skills para Claude Code
philadamson93/research-skills
Slash commands e lifecycle hooks para fluxos de pesquisa e auditoria em Claude Code
Ver no GitHubConjunto de skills e hooks em Shell para Claude Code que automatiza fluxos de pesquisa, auditoria de código e gestão de dados sensíveis. Inclui comandos para revisão de planos, versionamento seguro com validação de PHI, handoffs entre agentes e debates estruturados de arquitetura.
Instalação
npx convez-ai philadamson93/research-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
- Auditoria de planos de implementação antes de execução
- Validação automática de vazamento de dados médicos (PHI) em commits
- Handoff estruturado entre agentes planejadores e executores
- Revisão independente de código e cobertura de testes contra especificação
Veredito

Skill altamente especializada para fluxos de pesquisa e dados sensíveis. Excelente para equipes trabalhando com EHR/dados médicos ou handoffs multi-agente, mas com curva de aprendizado íngreme e dependência forte do Claude Code. Não é genérica o suficiente para uso amplo.
README
A tradução para o português deste README ainda está na fila.
research-skills
Personal Claude Code skills, shared in case they're useful. These are slash commands and instructions used across a research workflow.
Contents
commands/— slash commands loaded by Claude Code from~/.claude/commands/explain-plan.md— visual companion toread-plan: generates a self-contained interactive HTML (Mermaid diagrams with pan/zoom, blast-radius map, step accordion, inline feedback widgets) for plans too dense to review in prosereview-plan.md— independent design audit of a plan doc by Codex CLI or a fresh Claude Code subagent, then applies agreed feedbackwrapup.md— end-of-session cleanupnext.md— "where did we leave off?" session-openercommit-review.md— commit workflow with an appropriateness review (catches accidentally-leaked private content) before commit + pushphi-vet.md— hard pre-commit gate for medical-data repos that scans files for PHI leakage (patient/encounter/study identifiers, sample row data, free-text excerpts, image files);commit-reviewescalates to this when working in an OMOP/NeuralFrame/EHR/DICOM context. Machine-aware (fail-closed): inert on machines in the PHI-FREE allowlist (hooks/lib/is-phi-free-machine.sh), active everywhere elsereview-implementation.md— independent implementation audit of uncommitted code against a plan docreview-tests.md— independent test-coverage audit of uncommitted code against a plan docplan-handoff-readiness.md— pre-implementation handoff check: does the plan POINT AT or STATE DIRECTLY everything a fresh agent needs to implement it?vm-handoff.md— author and close the planner-Mac → executor-VM round-trip: renders thedocs/vm-status/<date>-<sha>.mdhandoff doc from the plan's Verification & VM handoff section (the per-step Expected + Stop criteria already reviewed with the plan viareview-plan— derived, not invented), gates them for sign-off (tiered by complexity), then offerscommit-reviewto land; on the VM appends the run results back into the same doc and offersphi-vet→commit-review. Auto-detects author vs readback mode byhostnameperclaude_ops.mdMachine-Aware Operating Mode. Chains with — and reuses, never duplicates —review-plan,plan-handoff-readiness,phi-vet,commit-reviewread-plan.md— opens a plan doc in the user's default.mdappopen.md— opens any file in the user's default app viaopen <path>; resolves the target by obviousness (just-touched file / explicit path → open directly; a description → search then open; ambiguous → confirm). General-purpose sibling ofread-plandebate.md— structured adversarial debate for major design decisions where reasonable people disagree (cross-repo architecture, build system choices, data contract design); two parallel advocates, codebase-grounded arguments, main-agent moderation, synthesized decision doc
hooks/— Claude Code lifecycle hooks (separate from skills: hooks are shell scripts the harness fires automatically at events likePreToolUse/SessionStart, not invoked-by-intent skills)phi-vet-gate.sh—PreToolUsehook that hard-gatesgit commitin medical-data repos untilphi-vethas signed off on the current staged tree (forces both PHI scan and explicit per-doc read-acknowledgement). Self-gates by machine (inert on PHI-free machines, so it's safe to install everywhere). Seehooks/README.mdfor install.lib/is-phi-free-machine.sh— shared fail-closed machine check (PHI-FREE allowlist) consulted by both the hook above and thephi-vetskill, so they never disagree about where PHI tooling is active.
CLAUDE.md— global Claude Code instructions (interaction style, environment notes)claude_ops.md— operating standards shared across research repos: planning workflow, code quality, git practices, pre-commit review pointers, skill composition discipline. Designed to be symlinked into each repo asdocs/claude_ops.mdso all repos reference the same source-of-truth (see Setup → Once per repo below).
Setup
These skills assume Claude Code as the harness. Some reference companion tools (codex CLI, gh); install as needed.
Setup is two parts: once per machine to make the slash commands available, then once per repo to give that repo the operating standards. The mechanics behind each step are in How it works below.
Once per machine
Make the slash commands available globally:
git clone git@github.com:philadamson93/research-skills.git ~/code/research-skills
ln -s /path/to/research-skills/commands ~/.claude/commands
Updating is just git -C ~/code/research-skills pull — symlinks survive Claude Code restarts, no rebuild. Hooks are not installed by these symlinks; they need a ~/.claude/settings.json entry per hook.
Once per repo
Give a repo the operating standards. Two steps — symlink the canonical file in, then import it from the repo's own CLAUDE.md so it loads into every session run from that repo:
cd <repo>
ln -sf /path/to/research-skills/claude_ops.md docs/claude_ops.md # adjust to reach your research-skills clone
Then add to the repo's CLAUDE.md (create one at the repo root if it has none):
## Operating standards
@docs/claude_ops.md
Commit both so collaborators and fresh checkouts pick them up (they only need research-skills cloned alongside):
git add docs/claude_ops.md CLAUDE.md && git commit -m "docs: wire in claude_ops.md operating standards"
How it works
-
The symlink makes the file present; the
@-import makes it load. Claude Code auto-loadsCLAUDE.mdfiles (walking from the cwd up to root) plus~/.claude/CLAUDE.md— but never an arbitrary file underdocs/. So the symlink alone governs nothing; the@docs/claude_ops.mdline is what pulls the standards into the session (@pathinlines that file's contents, and the read follows the symlink to the canonical source). The symlink earns its keep two ways: it gives@docs/claude_ops.mda stable target regardless of whereresearch-skillsis cloned, and it makes theReference: docs/claude_ops.mdbreadcrumb at the top of plan docs resolve to a real file. -
The relative symlink target
../../research-skills/...assumes<repo>andresearch-skillsare siblings (both under~/code/or similar), since the link sits at<repo>/docs/. Adjust the depth ifdocs/lives elsewhere or the repo is checked out somewhere unusual. -
Shortcut if you launch Claude from a parent folder. If you run Claude from a workspace folder that sits above several repos, one
@research-skills/claude_ops.mdimport in that folder'sCLAUDE.mdis inherited by every repo beneath it (CLAUDE.md loads the entire tree from cwd to root), so no per-repo wiring is needed. Convenient at scale, but it assumes the parent-folder launch pattern — the per-repo steps above are the portable default for working inside a single repo. -
Session-start sync:
git -C ~/code/research-skills pull --ff-onlyat the start of a session picks up cross-machine updates.claude_ops.mddocuments this as a Session Start step (it suggests a pull when the local clone is 3+ commits behindorigin/main).
Notes
review-plan.mdlooks for a repo-local checklist at.claude/references/plan-review-checklist.md— if present, it's loaded into the reviewer prompt to ground the audit in repo-specific contracts. Each project should author its own.- Skills evolve; feedback / PRs welcome.