convez.ai
Skill
dragonfly-thinking

fluency-agents-and-skills

dragonfly-thinking/fluency-agents-and-skills

Ver no GitHub
InicialPluginScore · de 0 a 100

Fluency Agents and Skills — a working Claude Code + Codex starter kit of agents and skills from the Dragonfly Thinking AI Fluency course (Extending session).

Instalação

$npx convez-ai dragonfly-thinking/fluency-agents-and-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
1
Forks
0
Acompanhando
0
Saúde
37%
Último commit
há 1 mês
Versão
—
Issues abertas
0
Contribuidores
2

README

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

Fluency Agents and Skills

Everything you take home from the Dragonfly Thinking AI Fluency course: the agents and skills we built and used, setup guides for the external connections we covered, and the key points from each session. Open it in Claude Code or OpenAI Codex and you have a real multi-agent setup — to use as-is, and to make your own.

A skill is a verb you invoke ("proofread this", "build me a deck"). A subagent is a specialist a skill can hand work to. Several skills delegate to the base agents below.

Fresh from the course? After installing, start with course-notes/ — the key points from all four sessions, with ready-to-paste prompts for putting them into action.

Install (let your agent do it)

No GitHub account needed — this page is public. Copy this page's link, paste it into Claude Code or Codex, and say:

"Read the AGENTS.md at this link and install the kit for me."

Your agent does the rest: it downloads the kit, follows AGENTS.md, and copies the 6 agents and 15 skills into your setup (~/.claude/ or ~/.codex/) so they're available in every session. When it's done, it'll tell you what to try first.

Already downloaded the kit (the green Code → Download ZIP button)? Same idea — tell your agent: "I've downloaded the fluency kit to my Downloads folder — install it for me." (Took the course earlier and still have last time's copy? Just paste the link above instead — your agent grabs the current version, since the kit keeps improving.)

From inside the downloaded/cloned folder:

# Claude Code
mkdir -p ~/.claude/agents ~/.claude/skills
cp -R .claude/agents/* ~/.claude/agents/ && cp -R .claude/skills/* ~/.claude/skills/

# Codex (then merge .codex/config.toml's [agents.*] blocks into ~/.codex/config.toml)
mkdir -p ~/.codex/agents ~/.codex/skills
cp -R .codex/agents/* ~/.codex/agents/ && cp -R .codex/skills/* ~/.codex/skills/

Start a new session and the skills/agents are live. See AGENTS.md for verification and runtime notes.

The agents (specialists)

AgentWhat it does
critical-friendPressure-tests an argument or plan — pushbacks, steel-manned counter-position, blind spots
fact-checkerVerifies factual/statistical claims against authoritative primary sources via web search
writing-editorHeavy editorial pass — clarity, structure, voice, cuts — without replacing your voice
project-plannerTurns a goal into milestones, tasks, dependencies, and honest estimates
vault-librarianReads your local notes/vault and surfaces what's relevant to the task
web-searcherRoutes a query to the best backend — cited search, papers, public stats, or live X/social — and returns a sourced answer with inline citations

The skills (verbs you invoke)

SkillWhat it doesDelegates to
setup-workspaceSets up your CLAUDE.md / AGENTS.md + context/ + projects/, writing the files for you. Smart-detects what you've got: builds from scratch if there's nothing, fills the gaps if your setup is half-done, or adds a project / refreshes context / adds a guardrail if it's complete.new-project / discovery-interview
new-projectInterviews you to find and shape your next project — offers ideas if you're not sure what to work on — then scaffolds it as a tracked project: overview.md + plan.md + progress.md + a router entry.discovery-interview / project-planner
proofreadClarity / grammar / structure / tone passwriting-editor
critical-reviewStress-test an argument and fact-check its claims, in parallelcritical-friend + fact-checker
research-briefSourced briefing on a topicweb-searcher
discovery-interviewInterviews you to turn a vague idea into a specproject-planner
premortemSurfaces how a plan could fail before you commitproject-planner
weekly-reviewPulls the week together into a reviewvault-librarian + project-planner
daily-briefA morning brief from your notes and the webvault-librarian + web-searcher
visual-explainerTurns content into a shareable HTML one-pager—
slidesBuilds an HTML slide deck—
canvas-designDesigns canvas/poster-style visual layouts—
pdf-createProduces a polished PDF—
here-nowPublishes a file/folder to a live {slug}.here.now URL—
generate-imageGenerates an image from a prompt (Nano Banana via one OpenRouter key)—

External connections (mcp/)

Agent-followable setup guides — point your agent at one and say "follow this and set it up for me":

GuideWhat it unlocksKey?
mcp/paper-search.mdAcademic literature — arXiv, PubMed, Semantic Scholar and ~20 moreNone
mcp/data-commons.mdPublic statistics — World Bank, WHO, UN, ABS and ~240 datasetsFree key
mcp/openrouter.mdImage generation (generate-image), live cited search, X/social searchOne paid key (~$10 credit)

Together these fill out all of the web-searcher agent's lanes — it routes queries to whichever source fits.

What's inside

.
├── .claude/                   # Claude Code kit
│   ├── agents/                #   6 subagents (.md)
│   └── skills/                #   15 skills (SKILL.md per folder)
├── .codex/                    # Codex kit (same capabilities, Codex-native)
│   ├── AGENTS.md
│   ├── config.toml            #   registers the agent roles (multi_agent = true)
│   ├── agents/                #   6 agent-role personas (.toml, via config_file)
│   └── skills/                #   the same 15 skills, SKILL.md format
├── course-notes/              # Key points from the 4 sessions + put-into-action prompts
└── mcp/                       # Setup guides for external connections (paper-search, …)

Claude Code vs. Codex

Both runtimes have skills and subagents, configured differently:

Claude CodeCodex
Skills.claude/skills/<name>/SKILL.md.codex/skills/<name>/SKILL.md (same format)
Subagents.claude/agents/<name>.md (markdown).codex/agents/<name>.toml, registered in .codex/config.toml
Subagent routingautomaticexplicit — "use the web_searcher agent to…"

See .codex/AGENTS.md for the full Codex mapping and setup notes.

How to kick things off

Once the kit is installed (see Install above), here's how to use it:

  • First time? Run setup-workspace. It'll interview you for ~5 minutes and create your personalised CLAUDE.md / AGENTS.md + context/ + projects/ for you — nothing to copy or save by hand. Coming back after the course? Run it again — it'll check what you've already got, fill in anything missing, and leave the rest alone. (If your workspace is already complete, it offers to add a project, refresh context, or add a guardrail instead.)
  • Starting something new — or not sure what to start? Run new-project. It opens by asking what you'd like to work on, helps you find the project if you want ideas, then scaffolds it as a tracked project (overview.md / plan.md / progress.md) so a future session can pick up exactly where you left off.
  • Have an idea you can't quite write down? Run discovery-interview. It asks the non-obvious questions until the shape of what you want is clear, then writes the spec.
  • Working on something concrete? Invoke any skill on the real work — "proofread this draft", "build me slides on X", "research-brief on Y", "run a premortem on this plan", "publish this with here-now".
  • Not sure what to do with any of this? Paste this into your agent:

Read the course notes in course-notes/, then look at what's actually set up on my computer. What from the course am I not using yet? Suggest three things worth putting into action this week — and walk me through the first one.

Make it yours

This is a starter, not a product. The whole point of the course was that you can shape these tools — so shape them:

  • Improve a skill after it stumbles: "that wasn't quite right — update the proofread skill so it keeps my heading style next time."
  • Add a skill when you notice a repeated task: "I keep doing X by hand — turn it into a skill." (Tip from Session 3: add a line to your CLAUDE.md / AGENTS.md asking your agent to suggest these moments.)
  • Trim and tailor: delete skills you never use, adjust defaults, add examples of your own documents to a skill's folder so outputs come out in your style.

Your agent can do all of this for you — just ask.


Part of the AI Fluency course by Dragonfly Thinking. This is the AI Fluency tier — a starter scaffold, not a methodology-license deployment.

Ver o original no GitHub

Bom para

Relacionados