Claude Global Skills
BryceEWatson/claude-global-skills
Skills globais para Claude Code: review, busca, Gemini e retrospectivas
Ver no GitHubColeção curada de skills para Claude Code que roda machine-wide com apenas Python e Node.js. Inclui loop de revisão de código multi-agente, busca de histórico local, session end/resume, geração de imagens via Gemini e retrospectivas de transcritos. Sem instalação por projeto — basta dropar na pasta ~/.claude/skills/.
Instalação
npx convez-ai BryceEWatson/claude-global-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 código com multi-agent loop e falsificação de findings
- Buscar exhaustivamente em histórico local de chats entre corpora
- Gerar e editar imagens via Gemini API com diagnostics
- Minerar padrões em transcritos com rigor estatístico (Krippendorff-α)
- Fechar e resumir sessões com registro evidence-grounded
Veredito

Conjunto robusto e opinionado de automações para fluxo de trabalho com Claude Code. Exigente em rigor (falsificação de claims, inter-rater checks), mas excelente para quem já trabalha em escala e quer padrões reproducíveis machine-wide.
README
A tradução para o português deste README ainda está na fila.
claude-global-skills
A curated collection of Claude Code skills:
a multi-agent code-review loop, exhaustive local chat-history search, an
evidence-grounded session end/resume pair, Gemini image generation, and rigorous
transcript retrospectives. They run machine-wide with nothing but python and
node. There's no per-project install: drop a skill into ~/.claude/skills/ and
invoke it as a slash command in any session.
These are global skills, version-controlled here so they can be reviewed, shared,
and deployed to a fresh machine. The repo is the source of truth; each live copy
is a deployed (materialized) copy. Most skills target Claude Code
(~/.claude/skills/). A skill can also declare Codex (~/.codex/skills/) as a
target and deploy to both from one shared source; see
Cross-runtime targets.
What this shows
These are the checks I run on my own software before anyone relies on it:
- A code review that tries to prove each problem it finds wrong before reporting it, runs the project's own tests and build, and, when there's a pull request, records its verdict on the exact version it checked.
- Work notes where every important claim says how I checked it, or says plainly that I didn't.
I bring the same habits to client work through Watson Standard, where I build, fix and harden software before real users rely on it.
Quickstart
Install a single skill by copying its directory into your Claude Code skills folder:
git clone https://github.com/BryceEWatson/claude-global-skills
cp -r claude-global-skills/review-loop ~/.claude/skills/
Then invoke it in any Claude Code session:
/review-loop
That's it for prompt-only skills. A few skills also ship an installer (they wire a
hook into ~/.claude/settings.json) — run node ~/.claude/skills/<name>/install.cjs
after copying. See each skill's section below.
Skills
Core — portable, useful for anyone
| Skill | What it does |
|---|---|
review-loop | Dispatches a multi-agent review team over your session's diff, runs an execution-grounded lint/test/build check, validates each finding through a falsifier stage, and posts a commit-pinned verdict on the PR. Four lens sets by artifact type: code, plans, analytical claims, and finished deliverables (whether the human it's handed to can actually use it). Ships a Stop-hook + installer. |
gemini-image | Generate and edit images via Google's Gemini API from one zero-dependency Python CLI — reference-image input, multi-image output, safety-block diagnostics, best-available-model selection. |
chat-history-search | Exhaustively search your local Claude history across both corpora (Claude Code CLI + Cowork/Desktop) — knows every log location and the false-positive gotchas (task-notifications, TodoWrite items, tool results) that trip up naive grep. |
pattern-retrospective | Mine your transcripts for recurring patterns with real rigor: audit-the-target-first discipline, streaming JSONL parse, 5-tuple extraction with provenance, self-falsification, and Krippendorff-α inter-rater checks. |
session-end | Close out a session into an evidence-grounded record (decisions, claims + verification, assumptions, artifacts, reversals); mid-flight, also emits a ready-to-paste continuation prompt that carries its own reconcile rules. Closes light, with no handoff file, when everything already sits in PRs. |
monitor-agent-thread | Watch a live or recent Claude Code or Codex session from the other product via local session logs, with a safe projection that never exposes hidden reasoning, raw tool arguments, signatures, encrypted content, or secrets. The first dual-target skill (Claude + Codex). |
fable-seat | Operating profile for a session running on Fable (claude-fable-5): default effort high and step DOWN for routine work rather than reaching for xhigh, lead task framing with the outcome, keep the prompt cache warm by farming cheap work to a subagent instead of inline-switching the main model, and budget + compact long runs. A genuine no-op on Opus and every non-Fable model; Fable-mechanics guidance, not a superiority claim. |
session-portal | Pass a short message from a Claude Code session to a Codex session, or back, through a local SQLite queue behind an authenticated MCP server. The recipient picks it up at a safe point and acts under its own permissions; nothing is pushed into a running session. Dual-target (Claude + Codex). |
ship-it | Take a build task from first read to a reviewed pull request: research, a plan that gets reviewed before any code, implementation, tests against the real app, then a review of the PR, looping each step until it's done. Leans on review-loop. |
Personal examples — wired to the author's setup; adapt before use
These show real, working patterns but reference the author's own projects, sites, or companion skills. Read them as reference implementations and adjust the paths, domains, and assumptions to yours.
| Skill | What it does |
|---|---|
transcript-analysis | Single-project transcript miner → proposes that project's CLAUDE.md candidates. The single-project sibling of global-review-loop. |
seo-index-validation | Probe a deployed site's crawl/index health (status codes, redirects, soft-404, sitemap, GSC) and diagnose why pages aren't indexed. A no-auth bash+curl script plus a playbook. |
global-review-loop | Mine your whole fleet's history for friction that recurs across projects, then propose global ~/.claude changes — reconciled against what already ships and self-validated by an adversarial claim loop. (Wired to a project registry; see its SKILL.md.) |
chat-arch-thrash-detect | A PostToolUse hook that nudges when a session falls into edit-thrash / read-loop / test-loop / tool-flail spirals. Hook host (not slash-invoked); ships its installer. |
weekly-work-log | Build a public weekly work-log page from session-end handoffs + git, with every number re-verified. Wired to the author's site as a worked example. |
email-editorial-pass | An editorial pass that runs before an outbound email draft is created: no em dashes, no hype, first person, the point up front. Wired to the author's published voice profile as a worked example. |
session-handoffis a thin alias that routes tosession-end(the skill was renamed);/session-handoffstill works if it's installed.
Privacy & safety
Several skills (chat-history-search, transcript-analysis,
pattern-retrospective, global-review-loop) read your private local Claude
chat history, and they don't all handle what they write the same way:
global-review-loopis the only one with a write guard. Its scripts that take an output path check it with a fail-closed guard that refuses to write into your~/.claudeconfig or into any git working tree. Its own git-ignored.local-state/folder is the one exception (the guard only warns if that folder sits inside a git working tree). Its proposals ledger is written to a fixed file in that folder by a separate check, which refuses outright if the folder sits inside a git working tree.pattern-retrospectivehas no guard. Its instructions put studies and extracted data underresearch/studies/in the current project, its finding-registry scripts write toreports/_data/there, and its coder script writes wherever you point--output, so its output can land in a tracked tree.transcript-analysisis instructions only. It tells the session to write its report toresearch/studies/in the current project, or to the project root if that folder doesn't exist. It also keeps a findings ledger with quotes from your sessions at.claude/transcript-analysis/findings-ledger.jsonin the project, and may leave extracted transcript text underdata/. Any of these can land in a tracked tree.chat-history-searchis instructions only. Its example commands keep lists of matching log files in/tmp, and it reports what it finds in the session rather than in a report file.
Check where a skill will write before you run it, and keep mined output out of
anything you commit or publish. Skills also run as executable code (some install
hooks), so a skill is a code-execution surface. Read SECURITY.md
before installing or contributing, and review a skill's code before you deploy it.
How this repo is maintained
The repo is canonical: skills are authored here and deployed (materialized) to
each live product tree. The engine is scripts/sync.py (stdlib-only):
python scripts/sync.py --deploy # repo -> live, for each skill's declared targets
python scripts/sync.py --check # report drift (live vs the repo-materialized output)
python scripts/sync.py --capture --target claude # pull a live edit back into the repo
--check/--deploy with no --target act on each skill's declared targets whose
live home exists, so --deploy then --check round-trips cleanly and a machine
without Codex is never touched. --capture requires an explicit --target and
refuses to silently resolve a divergence between two live copies. Contributors don't
need the maintainer's live tree — see CONTRIBUTING.md for the
outside-contributor path, the test commands, and the merge gate.
Optional drift nudge. hooks/skills_drift_hook.py reminds you to reconcile a
skill you edited in a live tree. It's a PostToolUse hook — register it in
~/.claude/settings.json (point CLAUDE_GLOBAL_SKILLS_REPO at your checkout):
{ "hooks": { "PostToolUse": [ { "matcher": "Edit|Write|MultiEdit|NotebookEdit",
"hooks": [ { "type": "command",
"command": "python \"$HOME/Projects/claude-global-skills/hooks/skills_drift_hook.py\"" } ] } ] } }
It's non-blocking (always exits 0), target-aware (Claude or Codex), and silenced with
CLAUDE_SKILLS_DRIFT_HOOK=0.
Cross-runtime targets (Claude Code + Codex)
A skill declares where it installs with an optional top-level targets: key in its
SKILL.md frontmatter:
targets: [claude, codex] # dual-target; omit the key entirely for Claude-only
- Absent key →
[claude](every existing skill is unchanged — backward compatible). - Shared content lives once. Files unique to one product go under
<skill>/overlays/<target>/and are added on top of the shared files at deploy time (additive-only). This is how Codex'sagents/openai.yamlreaches~/.codex/skills/but is never installed into Claude Code. - One
SKILL.md, product-correct commands. A{{SKILL_HOME}}token in shared text expands to$HOME/.claude/skills/<name>or$HOME/.codex/skills/<name>per target, so each install shows a runnable, product-correct path.
monitor-agent-thread is the reference dual-target skill. Homes are env-overridable
(CLAUDE_SKILLS_DIR / CODEX_SKILLS_DIR), which is how the test suite runs
hermetically. Full contract: SKILL-SPEC.md; migration from the
previous Claude-only workflow: CONTRIBUTING.md.
Documentation
CONTRIBUTING.md— how to add or change a skill, run the tests, and the review gateSKILL-SPEC.md— theSKILL.mdcontract (frontmatter, directory layout, dependency + privacy rules)SECURITY.md— data-handling model, secret scanning, and how to report a vulnerabilityCODE_OF_CONDUCT.md·CHANGELOG.md
License
MIT © 2026 Bryce Watson.
These skills run autonomously, and several read your local Claude chat history. The license's "AS IS / NO WARRANTY" terms apply — review each skill before you install it.