awesome-agent-skills

A small, opinionated set of agent skills I actually use. Each one is mined from a corpus of 200+ real community examples, not generated from scratch.
Built on Claude Code's skill format, works with any agent that reads SKILL.md + references/.

---

What's in here

Three skills. Each is a single folder with a SKILL.md (frontmatter + workflow) plus a references/ directory for deep-dive material. The agent (Claude Code, or any tool that reads the same format) matches the description field against whatever you're asking and loads the skill only when it's relevant. No eager bloat.

Skill	What it does	Corpus mined	Typical trigger
seedance-prompts	Prompt templates for ByteDance Seedance 2.0 video generation. Timecoded 3-shot storyboards, camera-move vocab, 15 category templates, multi-modal references (@图片/@视频/@audio), 12 curated examples.	~195 prompts across 3 awesome-lists	seedance, bytedance video, text-to-video prompt
nano-banana-prompts	Prompt templates for Google Nano Banana and Nano Banana Pro (Gemini 2.5 Flash Image). Two-tier model guidance, 15 fill-in templates (headshot, bento infographic, 3D diorama, split-view render), identity-preservation ranking, structured JSON/YAML/XML shapes, text-in-image rendering, 15 worked examples.	~600 prompts across 5 awesome-lists	nano banana, gemini 2.5 flash image, gemini image gen
awesome-readme	Write GitHub READMEs that land on awesome-lists and still sound human. 8 full templates (library/CLI/webapp/desktop/research/template-repo/monorepo/profile), shields.io badge library, banner/logo/GIF recipes, profile-README widget catalog, anti-patterns, and a dedicated anti-AI-slop audit pass integrating the humanizer skill's Wikipedia Signs of AI writing patterns.	Best-README-Template + ~100 awesome-readme entries + Standard-Readme + Make-a-README + RDD essay	readme, awesome readme, github readme, humanize readme

More coming when I build one that earns its place. I'd rather ship three that work than twenty that almost do.

---

Install

Pick one. Replace <skill> with the skill name (seedance-prompts, nano-banana-prompts, or awesome-readme).

Commands below target Claude Code's default skills dir (~/.claude/skills/). Other agents: swap the destination for whatever path your tool reads (commonly ~/.config/<agent>/skills/ or .agent/skills/ in a project root).

Symlink one skill (recommended — updates follow git pull)

macOS / Linux:

git clone https://github.com/fralapo/awesome-agent-skills.git ~/src/awesome-agent-skills
ln -s ~/src/awesome-agent-skills/skills/<skill> ~/.claude/skills/<skill>

Windows (PowerShell, admin or developer mode):

git clone https://github.com/fralapo/awesome-agent-skills.git $env:USERPROFILE\src\awesome-agent-skills
New-Item -ItemType SymbolicLink `
  -Path "$env:USERPROFILE\.claude\skills\<skill>" `
  -Target "$env:USERPROFILE\src\awesome-agent-skills\skills\<skill>"

Copy (static snapshot, no auto-update)

macOS / Linux:

git clone https://github.com/fralapo/awesome-agent-skills.git
cp -r awesome-agent-skills/skills/<skill> ~/.claude/skills/

Windows:

git clone https://github.com/fralapo/awesome-agent-skills.git
xcopy /E /I awesome-agent-skills\skills\<skill> $env:USERPROFILE\.claude\skills\<skill>

Symlink every skill at once (macOS / Linux)

git clone https://github.com/fralapo/awesome-agent-skills.git ~/src/awesome-agent-skills
for d in ~/src/awesome-agent-skills/skills/*/; do
  ln -s "$d" ~/.claude/skills/
done

Restart the agent after install. Most tools load skills at startup.

---

How a skill is organized

skills/<skill-name>/
├── SKILL.md              # required — frontmatter (name, description) + workflow
└── references/           # optional — deep-dive files loaded only when relevant
    ├── patterns.md
    ├── templates.md
    └── examples.md

The frontmatter description field is the matcher. Keep it specific and keyword-dense; the agent uses it to decide whether to load the skill at all.

---

How I build them

The pattern is consistent across all three skills in this repo:

1. Mine a real corpus. Community awesome-lists, vendor docs, internal notes. Several hundred examples minimum. I don't write skills from imagination.
2. Cluster what actually recurs. Templates, anti-patterns, category-specific moves. Things a real prompter or writer reaches for ten times a week.
3. Write SKILL.md short. Frontmatter, when-to-use, quick rules, pointers into references. The agent reads this at load time; every extra paragraph costs context.
4. Push depth into references/. Topical files. Loaded on demand when the agent follows a link.
5. Cite sources. When a template or pattern came from a specific public repo, it gets named in examples.md or the frontmatter analyzed_repos list.

If you're writing your own agent skills, those five steps are the only ones that have mattered for me. Everything else is decoration.

---

Contributing

Personal repo, but useful PRs get in. The bar:

• New skill: must match the structure above and cite its corpus
• Skill improvement: cite the specific pattern that motivated the change
• Bug fix: the smallest diff that fixes the thing

Open an issue before a large PR so neither of us wastes time.

---

License

MIT.

(back to top)