convez.ai
Skill
chenlu-hung

my-skills

chenlu-hung/my-skills

Ver no GitHub
InicialPluginScore · de 0 a 100

Personal collection of Claude Code Agent Skills (handoff, grill-me, caveman)

Instalação

$npx convez-ai chenlu-hung/my-skills

1 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

Estrelas
5
Forks
0
Acompanhando
0
Saúde
28%
Último commit
há 19 horas
Versão
—
Issues abertas
0
Contribuidores
1

README

A tradução para o português deste README ainda está na fila.

my-skills

Personal collection of Claude Code Agent Skills.

SkillPurpose
handoffContext transfer between AI coding sessions — creates a compact handoff doc so a fresh agent can resume. Includes a SessionStart hook (check-handoff.sh) that auto-detects handoff files. Integrates with project-map (below) to keep docs lean and cut resume-time exploration.
grill-meStress-tests a plan or design by interviewing you relentlessly until the decision tree is resolved.
cavemanUltra-compressed communication mode — cuts token usage ~75% while keeping technical accuracy.
project-mapBuilds a committed, on-demand .projectmap/ index (ctags symbols + short module summaries) so agents remember the codebase and grep the map instead of re-scanning the repo. Includes build-map.py; requires universal-ctags.
llm-councilConvenes a multi-model council — Codex (ChatGPT sub), Gemini (Antigravity agy), Claude (claude -p), and DeepSeek (opencode, free) — to answer a question, cross-review each other anonymously, then this session chairs the synthesis. Inspired by karpathy/llm-council; every member runs through its own subscription/sign-in CLI, no API keys. Includes council.py (parallel dispatch).
dispatchDelegates coding tasks to external agent CLIs as writable worker subagents — Codex (ChatGPT sub), Gemini (Antigravity agy), and DeepSeek (opencode, free). Splits work into self-contained briefs, runs workers in parallel (optionally in isolated git worktrees), then this session reviews and merges. Each worker runs through its own subscription/sign-in CLI, no API keys. Includes dispatch.py, plus an opt-in auto-dispatch layer: a UserPromptSubmit hook (dispatch-nudge.sh) that, on bulk/parallel/mechanical prompts, forces a one-line DISPATCH-or-SELF verdict before any code is written, governed by dispatch-policy.md.

Tools

Not a skill, but lives here too:

ToolPurpose
autocontinueAuto-resumes a Claude Code session after a usage-limit reset. A StopFailure hook queues the interrupted session; a launchd agent (every 5 min) resumes it headlessly once the limit resets, reusing the original permission mode and stopping after a configurable chain limit. Install once via autocontinue/install.sh — applies to all sessions. macOS only.

How handoff + project-map compose

They attack the two halves of resume cost. project-map carries the codebase structure (where things are); handoff carries the session state (decisions, dead ends, next steps). During a handoff, the skill runs project-map's read-only status check:

  • On resume — if the map is missing or stale, it offers to build/update so the fresh session reads the map instead of re-scanning the repo.
  • On create — it records a rebuild as a next step rather than spending build tokens at wrap-up.

A current map also lets handoff docs link .projectmap/ARCHITECTURE.md instead of re-describing structure, so the doc the next session reads back stays small.

Install

Copy any skill folder into ~/.claude/skills/:

cp -R handoff ~/.claude/skills/

Note: handoff's SessionStart hook must be registered separately in ~/.claude/settings.json and references ~/.claude/skills/handoff/check-handoff.sh.

Note: dispatch's auto-dispatch layer (nudge hook + policy) installs via dispatch/install.sh — it copies the skill/hook/policy, registers the UserPromptSubmit hook in ~/.claude/settings.json, and @-includes the policy from ~/.claude/CLAUDE.md. Safe to re-run to update.

Ver o original no GitHub

Bom para

Relacionados