Search1API

Live web research through the s1 command-line tool.

Prerequisites

Every command in this skill runs through s1. Check that it is installed and authenticated:

s1 balance

If the command is not found, install it:

curl -fsSL https://cli.search1api.com/install.sh | bash

Or via npm:

npm install -g search1api-cli

Then authenticate — prefer browser login, so no key is pasted into the conversation:

s1 login

Manual fallback:

s1 config set-key <your-api-key>

You can also set the environment variable SEARCH1API_KEY.

Do not install anything when s1 already works. To update an existing install, the user can run s1 update — only suggest this when a command fails in a version-related way.

If s1 is not found after installing, check that the npm global bin directory is on PATH. npx -y search1api-cli <command> works as a fallback without any install:

npx -y search1api-cli balance

Auth and credit errors are terminal. When a command reports an invalid key or insufficient credits, confirm the state once with s1 balance, then tell the user what is blocking and stop. Retrying the same command returns the same error.

When to use

Read it once, or learn it for good?

These look similar and are not:

• "What does this page say?", "summarize this link", "read these docs and

answer X" → s1 crawl. One page, answer now, nothing written to disk.

• "Learn this URL", "make a skill out of these docs", "记住这个文档" → s1 learn.

Produces an installable skill directory the user's agent can load later.

• Scope follows the request. A single link is one page; a documentation set

("learn these docs", "学一下 XX 文档") is --site pointed at its root — that is the normal case, not an escalation. When it is genuinely unclear, run --discover and show the user what a full run would cover.

Dynamic tuning

Adapt parameters to user intent — don't just use defaults:

• Quick lookup ("search for X", "what is X") → -n 5, no crawl
• Deep research ("research X thoroughly", "comprehensive analysis") → -n 15, then crawl top 3–5 results with separate s1 crawl calls
• User specifies a number ("find 10 articles") → match it with -n
• Recency signals ("latest", "recent", "this week") → -t day or -t month
• Domain-specific ("search on Reddit", "find GitHub repos") → -s reddit, -s github, etc.
• Site-scoped ("only from arxiv.org") → --include arxiv.org
• Chinese queries → use google, bing, wechat, or bilibili according to intent

Commands

search

s1 search "<query>" [options]

Search engines: google, bing, duckduckgo, yahoo, x, reddit, github, youtube, arxiv, wechat, bilibili, imdb, wikipedia

news

s1 news "<query>" [options]

Same options as search. News services: google, bing, duckduckgo, yahoo, hackernews. Default service: bing.

When user asks for breaking/latest news, always add -t day.

crawl

s1 crawl <url>

Extracts clean content from a URL. Use this whenever the user shares a link.

sitemap

s1 sitemap <url>

Returns all discovered links on a URL/domain.

trending

s1 trending <service> [-n <N>]

Services: github, hackernews.

learn

s1 learn <url> --name <name> [--site] [--exclude <paths...>] [--out <dir>]
s1 learn <url> --site --discover          # list what would be learned; crawls nothing
s1 learn --from <dir> --install <global|project> [--name <new-name>]
s1 learn --refresh <dir>                  # relearn using the source and scope it recorded
s1 learn --validate <dir>                 # static checks on a learned directory

Turns a URL into an Agent Skill directory:

<name>/
  SKILL.md                 # routing layer: when to use, what is in references/
  references/*.md          # page content with title/url frontmatter
  references/sources.json  # source URLs, hashes, and any pages that failed

--site learns what lives under the URL you point at: the docs root gets the whole set, /docs/api gets just that section.

One learn produces one skill, and one bundle per site is the default. Do not split a documentation set you are meeting for the first time. Before crawling you know only the URLs, and grouping pages by their slugs is a guess that goes wrong on exactly the pages that matter: on umami's docs, links, pixels and tags read like reporting features and are all instrumentation. What a page actually covers is knowable only after it has been crawled — at which point you already hold one complete bundle, and its routing table does the routing that separate skills would have done.

Point --site at a single section instead of the root only when that section is plainly its own job and you can already tell — a REST API reference under /docs/api, an SDK under /sdk. When in doubt, take the whole set.

--discover crawls nothing; it only reads what the site publishes, so use it to size a job before running it.

Name it after you have seen it

--discover gives you paths, not titles, and titles are what tell you whether a bundle is one job or four. When a site is unfamiliar, learn it under a working name, read the titles in references/sources.json, then rename on install:

s1 learn --from <staged dir> --name <real-name> --install global

That rewrites the folder, the frontmatter and sources.json together, and warns you if the body still mentions the old name.

Naming is yours to choose

s1 learn will not invent a name — it has only the URL, which is the one thing that should not decide the name.

The hard rules are just the shape: lowercase kebab-case, 64 characters or fewer, starting with a letter, and the folder basename must equal the frontmatter name. --validate checks those.

Everything else is judgement, and the user's judgement wins. Reuse the prefix of related skills they already have (ls ~/.agents/skills). Think about what else might live alongside it: umami-docs is a good name precisely when a separate skill about analytics practice could exist for the same product, and hono-build-api is a good name when nothing else will. Ask the user if two readings are both defensible — they know what their skill library looks like.

Writing the SKILL.md

The generated SKILL.md is a deterministic skeleton — no model writes it. Replace the placeholder description with real trigger wording, and turn the page list into a task-to-page routing table so the reader knows which file answers which question. Until that description is rewritten the skill will not be selected by an agent, and --validate fails on it.

Read the pages you need. Deciding what a section contains, what to exclude, what to call the skill, and what to put in the routing table all require knowing what is actually in the files — the URLs alone will mislead you. What to avoid is dumping the corpus into the conversation: read the pages that inform a decision, not all of them.

State no fact you have not read. The moment you write a concrete claim — a header name, a base URL, a flag, a limit — open that reference file first and cite it (Source: references/<file>.md). Writing API details from memory into a skill whose whole purpose is to be sourced is the one failure this command cannot survive.

Relearning the same URL rewrites references/, reports what changed, and keeps a SKILL.md you have edited while refreshing its reference table in place.

balance

s1 balance

Shows remaining API credits.

Workflows

Deep research

1. s1 search "<topic>" -n 15 for broad results
2. s1 crawl <url> on the top 3–5 relevant pages
3. Synthesize all gathered content into a coherent answer with source citations

Learning a URL into a skill

Nothing lands on disk until the user confirms, and nothing is installed until it is worth installing. Work in this order — scope decides coverage, coverage decides the name and the description, and all of that has to be settled before anything is installed.

Plan

1. Settle the scope. "Read this link" is one page. "Learn these docs", "学一下

XX 文档" is the set: --site, pointed at the docs root. Point at a single section instead only when the user asked for that section, or when that section is plainly its own job.

1. s1 learn <url> --site --discover --json — crawls nothing, returns the

section breakdown and every URL it would learn.

1. Decide what to leave out. Most doc sets carry a section that is dead weight

for half their readers: hosted versus self-hosted, one platform's deploy guide out of twenty, contributor and governance pages, changelogs. Ask the user which flavour applies. A section is not automatically homogeneous. Before excluding one, look inside: s1 learn <url-of-that-section> --site --discover. umami's /docs/guides is 33 pages of which 21 are hosting guides and 11 are core how-tos for tracking outbound links, form submissions, SPAs and server-side events — excluding it wholesale silently drops those. When a section is mixed, either keep it or exclude the individual pages.

1. ls ~/.agents/skills before naming. If something for this product already

exists, compare the job, not the name: the same job means refresh that skill instead (step 10), a different job means the two names must say which is which. Then choose the name (see naming above) and report the plan to the user: name, sections, page count, and what you are excluding.

Learn and author

1. s1 learn <url> --site --name <name> [--exclude <paths...>] --json. It writes

to a staging directory and installs nothing.

1. Author SKILL.md in the staging directory, before installing:

- a description that names the topics and trigger phrasing, and states what was excluded, so the skill declares its own gaps; - a routing table by task, not by directory, so the reader does not have to understand the site's layout; - facts only from files you have read, cited.

1. s1 learn --validate <staging dir> and fix what it reports.

Install

1. Tell the user what it captured and where it is, then ask: globally

(~/.agents/skills/), in this project (./.agents/skills/), or leave it staged for now.

1. Only after they say yes: `s1 learn --from <staging dir> --install

<global|project> — no further requests. Add --name <name> here to settle on a different name; it rewrites the folder, the frontmatter and sources.json together. Installing also links the skill into the agent directories that already exist beside the store, and the output names them — read that line, because a skill in the store that no agent is linked to is invisible. Run s1 learn --validate <installed dir>` afterwards too — if you renamed on install, the installed copy is the only place the folder name and the frontmatter can be checked against each other.

Maintain

1. Later, s1 learn --refresh <dir> relearns from the source, mode and

--exclude it recorded — no arguments to remember. Pass --exclude to change the scope, and it is recorded for next time. A refresh rewrites references/, keeps a SKILL.md you have edited while refreshing its reference table in place, and reports what was added, changed or removed. If the scope changed, update the description to match — a description that still claims a section was excluded after you added it back is worse than no description.

1. Pages that failed after their retry are listed in the output and recorded in

references/sources.json. Name them to the user rather than letting the skill look complete, and note that --refresh will try them again.

URL summarization

1. s1 crawl <url> → get the page content
2. Summarize or answer questions based on the content

Trending deep dive

1. s1 trending github -n 10 → discover hot topics
2. s1 search "<interesting topic>" -t day → search for details
3. s1 crawl <url> → read full article if needed

Output handling

• Commands produce human-readable output by default; add --json for

programmatic processing.

• After retrieving results, summarize and synthesize the information instead

of dumping raw output.

• Preserve source URLs and cite the relevant pages in the final answer.
• Distinguish facts found in sources from your own inference.

References

• [Usage examples](reference/examples.md) — read for additional patterns