Skills

Open Agent Skills — self-contained instruction packages that teach coding agents to do a specific job well.

A skill is Markdown an agent reads at runtime, plus any scripts and reference material it needs. Install one and your agent gains a capability it keeps: it knows when to reach for it, what steps to follow, and what to check before claiming it worked.

Works with Claude Code, Cursor, GitHub Copilot, Codex, Gemini CLI, Windsurf and 20+ other agents.

# see what's here
npx skills add prashant-cr/skills --list

# install just the one you want
npx skills add prashant-cr/skills --skill stock-deep-dive

Skills are independent — take one, take all six. See Installing for the difference between picking one and taking the lot.

Available skills

Web scraping & data extraction

Skill	What it does
scrape-feasibility-audit	Audits a site before you build a scraper — bot-detection vendor, CAPTCHA type, robots.txt rules, server- or client-rendered — then recommends tooling proportionate to what it actually found.
structured-data-extraction	Extracts data without CSS selectors, by reading what a page already publishes: JSON-LD, microdata, Open Graph, embedded state JSON, tables and feeds. Find a field by name, get a JSON path.
fix-broken-scraper	Diagnoses why a scraper broke and repairs it one verified step at a time — separating stale selectors from real blocking, changed routing, and content that moved into embedded JSON.

Together they cover the life of a scraping project: audit before you build, extract without brittle selectors, diagnose when something breaks.

Markets & research

Skill	What it does
news-stock-impact	Turns today's news into a ranked shortlist of listed companies genuinely exposed to it — the event, the transmission mechanism to a revenue or cost line, how much the stock has already moved, and what would invalidate the idea. Any market, any cap.
stock-deep-dive	Full fundamental analysis of one company — business, moat, earnings quality, management, and what the current price already implies — scored 0–10 across six dimensions, with bull/base/bear scenarios and separate verdicts for the long term and the next year.
ipo-rating	Rates an IPO and says whether to apply — scoring the business, valuation against listed peers, and the fresh-issue vs offer-for-sale split. Gives two separate verdicts, listing-day gain and long-term hold, because they often disagree. Reads grey market premium and subscription data, and maps the lock-in expiries.

See one in action

structured-data-extraction surveys what a page publishes, then locates a field by name:

$ python3 extract.py https://github.com/orgs/vercel/repositories --find starsCount

Keys matching 'starsCount':
  json_blocks[5].data.payload.orgReposPageRoute.repositories[0].starsCount = 30801
  json_blocks[5].data.payload.orgReposPageRoute.repositories[1].starsCount = 2264
  json_blocks[5].data.payload.orgReposPageRoute.repositories[2].starsCount = 15996
  json_blocks[5].data.payload.orgReposPageRoute.repositories[3].starsCount = 4163
  json_blocks[5].data.payload.orgReposPageRoute.repositories[4].starsCount = 25874
  json_blocks[5].data.payload.orgReposPageRoute.repositories[5].starsCount = 141153

A JSON path and an exact value, with no DOM inspection. That last one is the argument for this approach: the page renders it as 141k, so scraping the markup silently loses three digits and looks perfectly fine doing it. (Counts drift — that was a real run.)

What a skill here has to earn

Anyone can write instructions an agent will nod along to. These aim higher:

• Runnable tools, not just prose. Deterministic work belongs in a script the agent executes rather than reasoning through. The bundled scripts use the standard library only — nothing to install, nothing to break.
• Claims are verified, not asserted. Where a skill says a site behaves a certain way, that was checked against the live target. Where a claim turned out to be wrong, it was removed rather than softened.
• Tested against real prompts. Each skill carries eval cases in evals/ written the way users actually ask — including prompts whose premise is wrong, because the valuable answer is often the one that corrects the question.
• Honest about scope. A skill that says where it stops is worth more than one that improvises past the edge of what it knows.
• The least powerful tool that works. Recommending a headless browser for a job a plain HTTP request handles is a failure, not caution.

Installing

Just one skill — the common case, since the skills are independent and span different domains:

npx skills add prashant-cr/skills --skill stock-deep-dive

Several at once, comma-separated:

npx skills add prashant-cr/skills --skill stock-deep-dive,news-stock-impact

Everything, to every detected agent, without prompts:

npx skills add prashant-cr/skills --all

One thing worth knowing, because it surprises people: the bare command is not a "browse" command.

npx skills add prashant-cr/skills     # <- installs ALL six when run inside a coding agent

Run in a plain terminal it prompts you to choose. But run inside a coding agent it detects that, prints Agent detected — installing non-interactively, and installs every skill in the repo to every agent it finds — no picker, no confirmation. Use --skill when you want one, and --list when you only want to look.

Skills land in ./.agents/skills/ in the current project, with symlinks into each agent's own directory. Add -g to install user-level in ~/.agents/skills/ instead.

Flag	Effect
-g, --global	Install user-level (~/.agents/skills/) instead of into the current project
-s, --skill <skills>	Install specific skills ('*' for all)
-a, --agent <agents>	Target specific agents ('*' for all)
-l, --list	List the skills in this repo without installing
-y, --yes	Skip confirmation prompts
--copy	Copy files instead of symlinking into agent directories
--all	Shorthand for --skill '*' --agent '*' -y

Try a skill without installing it:

npx skills use prashant-cr/skills --skill <skill-name>

Update and remove:

npx skills update
npx skills remove <skill-name>

Repository layout

skills/
  <skill-name>/           # flat: skills/my-skill/SKILL.md
    SKILL.md              # required: frontmatter + instructions
    scripts/              # optional: executable helpers
    references/           # optional: docs loaded on demand
    evals/                # optional: test cases for the skill
    assets/               # optional: templates, images, fonts
  <category>/<skill-name>/  # or catalogued: skills/web-scraping/my-skill/SKILL.md
template/
  SKILL.md.template       # starting point for a new skill (not itself a skill)
scripts/                  # repo tooling

The CLI discovers skills at exactly those two depths. Anything deeper is invisible without --full-depth.

Two traps worth knowing if you fork this:

• There is deliberately no SKILL.md at the repository root. A shallower SKILL.md wins, so one there makes the CLI treat the whole repo as a single skill and hides everything under skills/.
• The template is named SKILL.md.template, not SKILL.md. Being outside skills/ is not enough to hide it — the CLI scans sibling directories too, and a template/SKILL.md gets published as an installable placeholder skill.

Adding a skill

mkdir -p skills/my-new-skill
cp template/SKILL.md.template skills/my-new-skill/SKILL.md
python3 scripts/validate_skills.py

The name in the frontmatter must match the directory name. The description is the entire triggering mechanism — it is the only part of a skill always resident in an agent's context, so it has to say both what the skill does and the concrete phrases that should summon it. Agents systematically under-trigger skills; a vague description is the usual reason.

Validate

python3 scripts/validate_skills.py            # all skills
python3 scripts/validate_skills.py skills/my-new-skill

No dependencies required. Exits non-zero on any problem.

Publishing

Installability and discoverability are different things, and only the first is automatic.

Push to a public GitHub repo and the skills are immediately installable by owner/repo — the CLI reads GitHub directly, with no registry in between, no submission form and no approval queue.

Appearing in npx skills find or on the skills.sh leaderboard is separate, and is driven purely by anonymous install telemetry from the CLI. Nothing in a repository changes that ranking; only real installs do.

License

MIT