Nimble Web Expert

Web extraction, search, and URL discovery using the Nimble CLI. Returns clean structured data from any website.

User request: $ARGUMENTS

Core principles

• Route by intent first (see [Analyze & Route](#analyze--route) for the full decision model). Named site with a matching Extraction Template + a direct item to look up → run the template. Site with no template, or a need that requires discovery/reasoning across pages → a Web Search Agent. One-off single URL → nimble extract. Raw results to work from ("find pages/articles about…") → nimble search; a synthesized deliverable (report, brief, comparison, recommendation) → a Web Search Agent. Discover/crawl URLs → nimble map or nimble crawl.
• Web Search Agent runs: pick a run mode before building the command. Default to named create-or-reuse — nimble agents run --agent-name <stable-name> — so a repeat session lands on the same agent. agents:runs create is the explicit-agent-ID route only and requires --agent-id. references/nimble-agents/reference.md has the mode table, use_case locking, and the one-time skill override.
• One command → present results → done. Run once, show the data immediately as a table. Do NOT experiment, loop, or write Python to parse output.
• Multiple inputs → always parallel. 2+ URLs/keywords/ASINs → &+wait. 6–20 → xargs -P. 20+ → Python asyncio script. See references/batch-patterns.md.
• Escalate render tiers silently on empty or truncated content. Tier 1 → 2 → 3 → … without asking. Surface a decision only when all tiers fail and investigation tools are needed. An access barrier is a different outcome, not a tier to climb — see Guardrails.
• Never answer from training data. Live prices, current news, today's listings → always fetch via Nimble. If unavailable, say so.
• AskUserQuestion at every meaningful choice. Header ≤12 chars, 2–4 options, label 1–5 words, recommended option first. Never present choices as numbered prose.
• Save all outputs to .nimble/. Never leave extraction results in memory only.
• Verify the connection BEFORE working — don't fire a data call and react to the error. With bash, nimble --version + NIMBLEAPIKEY confirms the CLI path; otherwise run one read-only mcp__pluginnimblenimble__nimbleagentslist probe. Success = connected; an auth/not-connected error or a response containing an OAuth authorization URL = not connected.
• No working CLI and no connected MCP → stop. Do not fall back to WebFetch, WebSearch, curl, or dangerouslyDisableSandbox. If the plugin is installed but the connector isn't connected (typical Cowork / claude.ai), surface the verbatim connect steps from rules/setup.md and stop; if no plugin at all, follow the install flow in rules/setup.md.
• If a tool hands back an OAuth "Authorize" link instead of data, present it exactly as given and stop. Never invent a "paste the URL back" / "I'll complete the connection" step — none exists — and never claim tools "will activate" then call them in the same turn. Wait for the user to authorize, then retry or re-probe.

Capabilities

One skill, one taxonomy. Use Nimble's own product names precisely — never paraphrase them.

Extraction Templates and Web Search Agents are distinct — an Extraction Template is a fixed, site-specific parser; a Web Search Agent reasons across sources. Never call an Extraction Template a "WSA" or a "legacy WSA," and never route a template use to agents (or vice-versa) by name alone. Building new templates/agents is out of scope here — use existing ones (point users to the Nimble app to build new).

Interactive UX

• Use AskUserQuestion at every meaningful choice — never guess, never ask in prose.
• Ambiguous request (no URL, vague topic): ask before running — "What would you like to do?" → Research & report / Search / Fetch URL / Discover URLs
• Gate B landed on a Web Search Agent at high+ effort: offer the cost/latency fork — Researched report / Quick scan (see [Analyze & Route](#analyze--route))
• Before running a search (if task maps to a specific focus mode): offer focus mode — General / News / Coding / Shopping / Academic / Social
• After all tiers fail: check investigation tools (which browser-use, python3 -c "from playwright.sync_api...") and ask whether to investigate with browser-use, Playwright, or skip.
• After presenting results, always close with: "Were these results what you needed?" → Looks great! / Mostly good / Not quite / Skip feedback

Prerequisites

Pick CLI or MCP at session start — same skill, two transports. Once a transport is selected, stick with it for the session and don't re-probe on every command.

nimble --version && echo "${NIMBLE_API_KEY:+API key: set}"        # CLI path
# OR (fallback when shell isn't available)
claude mcp list 2>/dev/null | grep -q "nimble" && echo "MCP: ok"  # plugin MCP

• CLI ready (version + API key both print) → proceed to [Step 0](#analyze--route), use nimble ... commands.
• MCP connected (no CLI, but plugin is installed) → proceed to [Step 0](#analyze--route), use mcp__pluginnimblenimble__* tools instead.
• Neither → load rules/setup.md for the environment-aware install flow. Any Claude product (Code, Cowork, claude.ai) → /plugin install nimble. Codex or other terminal-only agents → npm i -g @nimble-way/nimble-cli. Cursor / VS Code / generic MCP clients → paste the mcp.json snippet.

If bash is denied: you're in a Cowork-like / MCP-only host. Use mcp__pluginnimblenimble__* tools, but verify the connection first with one read-only nimbleagentslist probe. If the probe fails with an auth/not-connected error or returns an OAuth authorization URL, the connector isn't connected — surface the connection steps from [Core principles](#core-principles) and stop (and never invent an auth-completion flow). Never substitute WebFetch, WebSearch, curl, or any other tool for Nimble operations.

Analyze & Route

Two gates, in order. Gate A asks where the data lives; Gate B asks what the user wants back. Most mis-routes come from skipping Gate B — a request with no location signal is not automatically a search.

Gate A — do I know where the data lives?

The most common overlap — a site with no Extraction Template. It looks like a choice between a raw extract (which dumps parsing work on the user) or building a template (out of scope). Neither is right: fall through to Gate B, which will land on a Web Search Agent — it configures fresh for any site and reasons about structure without a maintained template. This isn't a question to put to the user; when no template fits, the answer is the same every time.

Step 0 — Extraction Template check (when a site + direct item is named)

Templates return clean structured data with zero selector work. Always check first.

Always verbalize — never silently:

1. Announce: "Let me check if there's a Nimble Extraction Template for [site]..."
2. Report: "Found <templatename> — using it now." or "No template for [site] — using a Web Search Agent instead."_

Lookup order:

1. ~/.claude/skills/nimble-web-expert/learned/examples.json → learned templates
2. nimble extract:templates list --limit 100 → filter by site/domain client-side; confirm the match
3. Inspect the schema before running: nimble extract:templates get --extract-template-name <name>
4. No match → route to a Web Search Agent (per the overlap rule above)

nimble extract:templates run --template <name> --params '{"key": "value"}'

--params is a JSON/YAML mapping matching the template's inputschema. The response is the records defined by the template's outputschema (array for list/SERP-style, object for detail/PDP-style) — read the schema from get to know the shape. See references/nimble-extract-templates/reference.md.

⚠️ For finding information, use nimble search, not a SERP-analysis template. SERP templates are for rank/SEO analysis, not general retrieval.

Gate B — what does the user want back?

nimble search returns raw material to skim. A Web Search Agent returns a finished, cited answer. The prompt's deliverable noun decides it — route on that, not on how open-ended the topic sounds.

Structured rows about many entities → Web Search Agent with enrichment or dataset_building. See references/nimble-agents/reference.md.

Offer the fork when the answer is the expensive one

A Web Search Agent at high effort takes minutes and costs more; a search takes seconds. That's a real trade-off, so surface it — but only when Gate B lands on a Web Search Agent AND the recommended effort is high or above. One AskUserQuestion, recommended option first:

• Researched report — Web Search Agent, a few minutes, every claim cited
• Quick scan — nimble search, seconds, raw links you skim yourself

Below high, don't ask — just run the Web Search Agent. Never ask when Gate A already resolved the route.

Dataset requests always clear the threshold. "Build a list of…" → dataset_building, which runs at high or above by definition, so the fork always applies. Enrichment has no such floor — judge "enrich these rows" on the normal effort rule and skip the prompt when a small, well-specified fill-in lands below high.

Before starting any Web Search Agent run, say how long it will take, then narrate at phase transitions. On MCP, progress comes from bounded status polling rather than a live stream — poll and report each step, because an un-narrated multi-minute run reads as a hang.

Workflow

For the full extract waterfall (tiers, flags, browser actions, network capture), see references/nimble-extract/reference.md.

Response shapes

Read the template's output_schema (from extract:templates get) before parsing — a list/SERP-style template returns an array, a detail/PDP-style template returns an object. Web Search Agent runs are async: poll agents:runs get to a terminal state, then fetch result.

Output & Organization

mkdir -p .nimble   # save all outputs here

Naming: .nimble/<site>-<task>.md (e.g. .nimble/amazon-airpods.md, .nimble/yelp-sf-italian.json)

Working with saved files:

wc -l .nimble/page.md && head -100 .nimble/page.md
grep -n "price\|rating" .nimble/page.md | head -30

End every response with: Source: [URL] — fetched live via Nimble CLI

Self-Improvement

The skill maintains ~/.claude/skills/nimble-web-expert/learned/examples.json.

• At task start: read the file, scan good[] for url_pattern matches → use documented command/tier as starting point. Scan bad[] → avoid documented pitfalls.
• After presenting results: ask "Were these results what you needed?" → on positive feedback, append to good[] with urlpattern, task, command, tier, notes. On negative feedback, ask "What went wrong?" and append to bad[] with urlpattern, task, issue, avoid, better.
• Keep entries concise — 5–10 per site. Only write on real feedback, never speculatively.

Guardrails

• NEVER answer from training data for live prices, current news, or real-time data. If Nimble is unavailable, say so.
• NEVER skip Step 0 silently. Even if certain there's no template, announce the check before falling back to a Web Search Agent or extract/search.
• NEVER answer a synthesis deliverable with raw search results. "Report", "brief", "compare", "best X", "which should I" → Gate B routes to a Web Search Agent. Handing back a list of links and calling it a report is the most common mis-route.
• Distinguish Extraction Templates from Web Search Agents. Never call a template a "WSA"/"legacy WSA," and never route a template use to agents by name alone (or the reverse). Building new templates/agents is out of scope — use existing ones.
• When a run comes back empty, partial, or clearly wrong, say so plainly — a domain that returned nothing, a template that matched poorly, a search with no relevant hits are real outcomes, not something to present as success. Suggest an obvious next step (broader source, a different capability) where one exists.
• NEVER retry the same render tier. If a tier returns empty or truncated content, escalate — do not re-run.
• NEVER escalate at an access barrier. A CAPTCHA, a human-verification page, or a sign-in wall in place of the target is a real outcome — report it plainly and stop. Where a supported alternative exists, take it: --focus social search for social profiles, public search results for gated articles.
• NEVER substitute WebFetch, WebSearch, curl, or wget for nimble operations. They're not in allowed-tools — if a Nimble transport isn't available, stop and follow the guidance in the no-transport branch of Core principles. Don't try to work around it.
• NEVER load reference files speculatively. Only read a reference when the current task explicitly needs it.
• Task agents MUST use runinbackground=False.
• Hard retry limit. On error (not empty content): retry at most 2 times with different flags. After 2 errors, report and stop.
• Hard 429 rule. On rate-limit error: stop immediately. Do not retry or switch tiers.

Reference files

Load only when needed: