exploration
An agent skill that gets AI out of its defaults on open-ended work.
Ask a model to design a landing page, name a product or plan an architecture, and you get the same safe answer it gives everyone. exploration rolls a random seed, reads it for patterns, and uses them to pick a direction for the parts of the task that are genuinely open. Your requirements stay fixed. Then it builds that direction properly.
Connect fal and it also generates the images, video, music and voice each direction needs.
npx skills add bgarcia7/exploration
Same brief, with and without it
Both sets answer the same request: four landing-page directions for a made-up personal assistant called Nova. Each preview is the full page, running live. Select one to open it.
Without the skill
Four pages that are really one. The same sections in nearly the same order, a mock chat or log where imagery should be, three feature blurbs, three steps, a placeholder testimonial and an email box. Swap the colors and they're interchangeable: generic AI slop.
With the skill and fal
Four directions that share nothing but the brief: a linocut almanac, a record sleeve with a spinning disc, a contact sheet of documentary photos, and a 1968 supergraphic around a looping kitchen video. Each has its own structure and voice, and real assets made for it.
Install
1. Add the skill
With the skills CLI, for any agent that reads Agent Skills (Claude Code, Codex, Cursor and others):
npx skills add bgarcia7/exploration
Add -g to install it for all your projects. The CLI needs Node 22 or newer.
- Manually: copy
skills/explorationfrom the repository into your agent's skills folder, for example~/.claude/skills/. - Claude.ai or the Claude desktop app: download
exploration.zipfrom the latest release and upload it under Skills in settings. Code execution needs to be on.
2. Connect fal (recommended)
Without fal the skill still works, but directions that call for photography, illustration or video fall back to code and placeholders. The fal MCP server is free. You pay only for the model runs you trigger.
- Create an API key at fal.ai/dashboard/keys.
- Add the server to your agent, replacing
YOUR_FAL_KEY:
claude mcp add --transport http --scope user fal-ai \
https://mcp.fal.ai/mcp \
--header "Authorization: Bearer YOUR_FAL_KEY"
--scope user makes it available in every project. Run /mcp in Claude Code to check that fal-ai is connected.
Add to ~/.cursor/mcp.json:
{
"mcpServers": {
"fal-ai": {
"url": "https://mcp.fal.ai/mcp",
"headers": { "Authorization": "Bearer YOUR_FAL_KEY" }
}
}
}
Add to ~/.codeium/windsurf/mcp_config.json, then enable fal-ai in Cascade's MCP settings:
{
"mcpServers": {
"fal-ai": {
"serverUrl": "https://mcp.fal.ai/mcp",
"headers": { "Authorization": "Bearer YOUR_FAL_KEY" }
}
}
}
No key needed. Install the fal plugin and connect your fal account when prompted. In the Codex CLI, use /plugins to install it.
Any client that supports Streamable HTTP and a custom header:
URL: https://mcp.fal.ai/mcp
Header: Authorization: Bearer YOUR_FAL_KEY
- Start a new session and ask:
Use fal to search for image generation models. Do not run a model.If you get models back, you're set.
By default the skill uses the highest-quality model for each piece without asking. Mention a budget and it prices the plan and waits for your OK first. More in fal's MCP docs.
Use
Ask in plain words. The skill picks up open-ended requests on its own:
- “Landing page for my bike repair app, x3”
- “Surprise me with names for a tea subscription”
- “Three ways to do offline sync for this app, compared”
- “Rewrite this intro, but try something I wouldn't”
Add a number like x3 for independent directions with a comparison. Ask for a re-roll when you want a new direction rather than a revision. It covers design, architecture, marketing, ideation, writing and editing, naming, research and media.
What's inside
skills/exploration/ ├── SKILL.md the method ├── scripts/seed.py seed generator └── references/ one guide per domain, plus fal
Needs python3 or any POSIX shell. MIT licensed.