docs-seeker
Fetch up-to-date documentation for any library, framework, API, or service into context. Use when the user wants to look up API references, check function signatures or required fields, find feature-specific docs, or verify how an external tool actually works. Triggers for queries about third-party libraries like Stripe, SQLAlchemy, Tailwind, FastAPI, shadcn, Drizzle, Hono, Better Auth — any time the answer lives in official docs rather than in the project codebase. Use this instead of guessing from trained knowledge, which is stale.
git clone --depth 1 https://github.com/avibebuilder/claude-prime /tmp/docs-seeker && cp -r /tmp/docs-seeker/.claude/skills/docs-seeker ~/.claude/skills/docs-seekerSKILL.md
## Why fetch instead of recall
Trained knowledge rots. Library APIs rename fields, deprecate hooks, restructure auth flows, and change defaults between minor versions. Answering from memory is how you get confidently wrong code. Replace memory with retrieval — cheaply, from sources designed for AI consumption.
## The source hierarchy (and why)
Prefer sources in this order. The ranking is about *signal per token*, not just availability.
1. **`llms.txt` on the official docs site** — hand-curated by the project, AI-optimized, token-dense, always current. This is the ideal source when it exists. Try `{official-docs-url}/llms.txt` first for any library with a docs site; many projects ship one even if they don't advertise it. If `llms.txt` is just an index of links, follow the most relevant ones. `llms-full.txt` exists on some sites and contains the full corpus — only reach for it when the user explicitly wants comprehensive docs, since it's large.
2. **Context7** — a mirror that ingests GitHub repos and exposes them at `https://context7.com/{org}/{repo}/llms.txt`, with optional `?topic={keyword}` filtering. Use this when the project has no official `llms.txt`, or when you want to scope to one feature. The `{org}/{repo}` path mirrors GitHub exactly — derive it from the user's `package.json`, imports, lockfile, or the project's GitHub URL rather than guessing. For docs sites without a clear repo, Context7 also hosts `https://context7.com/websites/{normalized-path}/llms.txt`.
3. **GitMCP** — any GitHub repo is accessible by swapping `github.com` → `gitmcp.io` in the URL. Useful when Context7 doesn't have the repo indexed, or when you need source-of-truth README/examples straight from the repo.
4. **WebSearch** — last resort. Slower, noisier, and you'll spend tokens filtering results. Only fall here when the three structured sources all miss. When you do search, query for `"{library} llms.txt"` first — it often surfaces an official or community-maintained one.
On any 404, timeout, or empty response: move to the next tier immediately. Never retry a failed source.
## Topic scoping
When the user's query targets a specific feature (e.g., "shadcn date picker", "Next.js middleware", "Stripe webhooks"), append `?topic={keyword}` to the Context7 URL to narrow the fetch. Pick a short root keyword that captures the feature — judgment call, no rigid rules. The goal is fewer tokens, higher relevance. If the topic URL returns nothing useful, drop the topic and try the general URL.
## Reading the docs you find
Once you have URLs, the question is how to read them without polluting the main context.
- **A handful of small pages, or content the orchestrator clearly needs verbatim**: read directly with `WebFetch`. Fast, simple, no overhead.
- **Many pages, or large pages where you only need specific answers**: fan out to parallel subagents, each reading a subset and returning a condensed summary. This protects the main context window from doc bloat and parallelizes I/O. Use judgment on fan-out count — a couple for moderate sets, more for large ones. The tradeoff is latency/tokens vs. main-context pollution; lean toward subagents whenever the raw docs would be noisy relative to what the orchestrator actually needs.
When delegating to subagents, tell them exactly what question to answer and what to return (e.g., "return the exact signature and required fields for `stripe.webhooks.constructEvent`, plus any version notes") — not "summarize these docs". Specific asks give specific answers.
## Version awareness
Before fetching, check what version the project actually uses — `package.json`, `requirements.txt`, `go.mod`, lockfiles. Fetching the latest docs when the project is pinned two majors behind is a common way to hand back wrong answers. If a version-specific doc path exists (e.g., `/v2/llms.txt`, `/docs/4.x/`), prefer it.
## Gotchas
- **Don't fabricate.** If every source misses, say so clearly and ask the user for a URL or a different approach. A made-up API signature is worse than "I couldn't find it."
- **Don't over-fetch.** The orchestrator asked a question; pull what answers it, not the entire manual. Every token you add competes for attention downstream.
- **Don't trust your own cache.** If you recall that a library's docs live at a certain URL, verify — sites reorganize. A fresh fetch beats a confident memory.
- **Report what you used.** When you hand results back, note which source succeeded (llms.txt / Context7 / GitMCP / WebSearch) and the URLs fetched. This lets the orchestrator judge freshness and lets the user follow up.
## Constraints
- Use `WebFetch` to read URLs. Do not invoke MCP servers for this.
- Prefer `llms.txt` over `llms-full.txt` unless comprehensive docs are explicitly requested.
- Never retry a failed source — move down the tier list.Browser automation CLI for AI agents. Use when the user needs to interact with websites, including navigating pages, filling forms, clicking buttons, taking screenshots, extracting data, testing web apps, or automating any browser task. Triggers include requests to "open a website", "fill out a form", "click a button", "take a screenshot", "scrape data from a page", "test this web app", "login to a site", "automate browser actions", or any task requiring programmatic web interaction.
Answer questions about code, architecture, and technical decisions — no implementation. Trigger on questions asking 'why', 'what does this do', 'what is the purpose of', 'explain', 'what's the difference', 'compare', or 'what are the tradeoffs' — even when referencing specific files, code snippets, or inline code. The key signal is the user wants to UNDERSTAND something, not change it. Do NOT trigger for requests to build, fix, plan, review, research, or add/modify code.
Implement, build, create, or add any feature, endpoint, page, component, or functionality. Use this skill whenever the user asks you to write new code or make code changes — whether it's adding an API endpoint, building a UI page, creating an export feature, wiring up a webhook, implementing a search/filter, or any other hands-on coding task. This is the default skill for all 'build this', 'add this', 'create this', 'wire up', 'implement' requests. Covers the full cycle: clarify requirements, plan if needed, write code, verify, and review. Do NOT use for pure research, debugging, documentation, or explanation — only when the user wants working code delivered.
Use when the user wants to save knowledge as a file so others don't have to rediscover it — \"turn this into a doc\", \"write this up\", \"document how X works\", \"we figured this out and want to capture it\", \"nobody should have to figure this out again\". Covers any request to create or update durable written artifacts: onboarding guides, runbooks, ADRs, API docs, architecture notes, postmortems, changelogs, setup guides. The trigger: user wants knowledge captured in a file for future reference, not just a conversation. Do NOT use when still making decisions (→ give-plan), just asking for explanation without a file (→ ask), or writing code (→ cook).
Investigate unexpected behavior and mysterious bugs. Use when the cause of a problem is unknown and the user needs to understand WHY something is happening — symptoms like: sudden unexplained changes in metrics or behavior, works locally but not in staging/production, inconsistent or intermittent failures, correct code producing wrong results, operations succeeding but having no effect, environment-specific failures, duplicate executions, stale data, or any \"why did this change?\" or \"why is this happening?\" situation. Covers infrastructure anomalies (cache hit rates dropping, latency spikes, queue behavior shifts) as well as code bugs. The key signal is confusion about root cause, not a request to implement a known fix. Do NOT use for feature requests, known fixes, planning, or documentation tasks.
Brainstorms and debates approaches, then drives toward an actionable decision. Use whenever someone needs a thinking partner for a decision they're facing: 'discuss', 'debate', 'brainstorm', 'weigh options', 'tradeoffs', 'should I do X or Y', 'help me decide', 'I'm torn between', 'sanity check my thinking', or 'what do you think about'. The user must be asking for help reasoning through a choice — not asking to build, fix, evaluate, plan, or modify something (even if the topic involves this skill itself). Picks the right decision lens, surfaces tradeoffs and blind spots, pushes back when reasoning is genuinely weak, and never implements.
Fix bugs and broken behavior when there is enough evidence to act on a repair path. Use for errors, crashes, incorrect results, API failures (500, 404, 403), CORS problems, database exceptions, broken rendering, duplicated or wrong data, off-by-one mistakes, timezone/date bugs, broken forms, config-caused runtime failures, and regressions. Trigger when the user wants the bug repaired and the conversation already contains a clear failing area, a reproducible failing test, a concrete error path, or a prior diagnosis to implement. Do NOT use for new features, pure explanation, architecture discussion, broad research, or bug reports where the main need is figuring out why the behavior happens — use diagnose for that.
Builds distinctive, production-grade UIs that avoid generic AI aesthetics. Use whenever the user wants to build, restyle, or give visual direction to any interface — pages, dashboards, landing pages, components, onboarding flows, mobile screens, or design systems — even without an explicit 'design' request. Also triggers for: picking an aesthetic direction, improving the look of a dull/generic existing page, adding visual personality, or choosing colors/typography. Includes a bundled design intelligence database for concrete guidance across web (React, Next.js, Vue, Tailwind) and mobile (React Native, Flutter, SwiftUI).