company-deep-dive（公司深潜）

Company Deep Dive

360° company research powered by Nimble's web data APIs.

User request: $ARGUMENTS

Before running any commands, read references/nimble-playbook.md for Claude Code constraints (no shell state, no &/wait, sub-agent permissions, communication style).

Instructions

Step 0: Preflight

Run the preflight pattern from references/nimble-playbook.md (5 simultaneous Bash calls: date calc, today, CLI check, profile load, index.md load).

From the results:

• CLI missing or API key unset → references/profile-and-onboarding.md, stop
• Profile exists → note it for context (company name helps frame the research). Read ~/.nimble/memory/companies/index.md to check if the target company already has prior research. Follow [[path/entity]] cross-references to load related context.

- Prior research exists: Load it. Run in refresh mode — focus on what's new since the last report date. Tell the user: "I have prior research on [Company] from [date]. Refreshing with latest data." - No prior research: Run in full mode — comprehensive across all dimensions.

• No profile → that's fine. Company deep dive doesn't require onboarding (unlike competitor-intel). Proceed directly to Step 1.

Step 1: Identify Target Company

Parse the target company from $ARGUMENTS or the user's message.

If clear (e.g., "research Stripe", "tell me about Datadog"):

• Extract the company name
• Run two Bash calls simultaneously to confirm identity:

- nimble search --query "[Company] official site" --max-results 3 --search-depth lite - nimble search --query "[Company] company overview" --max-results 5 --search-depth lite

• Confirm briefly: "Researching [Company] ([domain])..."

If ambiguous (e.g., "research Mercury" — could be bank, auto, or other):

• Ask one clarifying question with the top candidates

If missing — ask: "Which company would you like me to research?"

Scope selection — if the user hasn't specified depth, default to full deep dive. If they say "quick overview", "brief", or "summary", run a quick mode that skips the Deep Extraction step and produces a shorter report.

Step 2: WSA Discovery

Discover available WSAs for the target company's domain. Run both searches simultaneously:

nimble agent list --search "{company-domain}" --limit 20

nimble agent list --search "{company-name}" --limit 20

From the results, filter for WSAs with entity_type matching SERP or PDP, and prefer managed_by: "nimble". Validate each with nimble agent get --template-name {name}, then cache discovered WSA names + params for the run. Pass them to dimension agents in Step 3 for enrichment alongside nimble search. If no WSAs found, continue with nimble search alone.

Step 3: Parallel Research Across Dimensions (sub-agents)

Read references/dimension-agent-prompt.md for the full agent prompt template. Follow the sub-agent spawning rules from references/nimble-playbook.md (bypassPermissions, batch max 4, explicit Bash instruction, fallback on failure).

Spawn nimble-researcher agents (agents/nimble-researcher.md) with mode: "bypassPermissions". Each agent researches one dimension of the company. Pass discovered WSA names from Step 2 to each agent so they can use them for enrichment alongside nimble search.

Important: The Nimble API has a 10 req/sec rate limit per API key. With each agent running 4-5 searches in parallel, limit concurrent agents to 2 per batch to stay under the limit. Run overview searches in their own phase, not alongside agent batches.

Call estimation & Scaled Execution: Before launching agents, estimate total API calls: 2 overview searches + ~5 searches per agent × 5 agents = ~27 calls. Each agent should use extract-batch or agent run-batch for 11+ calls instead of individual calls. See the Scaled Execution pattern in references/nimble-playbook.md for tier selection.

Phase A — Overview searches (run directly, before agents):

• nimble search --query "about" --include-domain '["[domain]"]' --max-results 3 --search-depth lite
• nimble search --query "[Company] Wikipedia OR Crunchbase OR Pitchbook" --max-results 5 --search-depth lite

These give foundational context (founding date, HQ, employee count, mission) that frames all dimensional findings.

Phase B — Batch 1 (2 agents simultaneously):

Phase C — Batch 2 (2 agents simultaneously):

Phase D — Batch 3 (1 agent):

Refresh mode adjustment: If prior research exists, pass the known facts to each agent as context so they focus on what's new. Agents should use --start-date to filter for recent data only.

Fallback: If any agent fails or returns empty, run those dimension searches directly from the main context. Don't leave gaps in the report.

Step 4: Deep Extraction

From all agents' results, identify the top 5-8 most informative URLs across dimensions. Prioritize:

• Funding announcements with specific amounts
• Official product/feature pages
• Executive interviews, podcast appearances, or conference talks
• In-depth analyst or journalist profiles
• The company's own about/team page

Make one Bash call per URL, all simultaneously:

nimble extract --url "https://..." --format markdown

For extraction failures, follow the fallback in references/nimble-playbook.md.

Quick mode: Skip this step entirely. Report from search snippets only.

WSA enrichment: If WSAs were discovered in Step 2, use them here for richer extraction on key URLs before falling back to nimble extract.

Step 5: Synthesize Report

Structure the output as a 360° Company Report:

# [Company Name] — Deep Dive
*As of [today's date]*

## Quick Assessment
[2-3 sentence verdict: what this company is, where they stand, and the one thing
that matters most right now. This is the "if you read nothing else" paragraph.]

## Company Overview
- Founded: [year] | HQ: [location] | Employees: [estimate]
- Domain: [domain] | Industry: [industry]
- Mission/focus: [one line]

## Funding & Financials
[Latest round, total raised, key investors, valuation signals, revenue indicators.
Every claim dated and sourced.]

## Leadership & Team
[Founders, C-suite, notable recent hires or departures. Executive perspectives
on company direction — direct quotes when available from interviews or talks.]

## Product & Technology
[Core products, recent launches, tech stack signals, engineering culture,
open-source contributions. What they're building and how.]

## Market Position
[Key competitors, differentiation, market share signals, analyst perspectives,
customer sentiment from reviews (G2/Capterra/Reddit).]

## Recent News & Events
[Chronological, most recent first. Each entry dated with source.]

## Strategic Outlook
[Synthesis across all dimensions: where the company is heading, key risks,
growth signals, and strategic bets. This is insight, not summary.]

## Sources
[Numbered list of all URLs cited in the report]

Core rules:

• Every factual claim must have a date and source URL.
• Lead with the Quick Assessment — most readers stop there.
• Say "no public data found" for a dimension rather than speculating.
• Distinguish between confirmed facts and inferred signals.
• Executive quotes add credibility — include direct quotes from interviews, earnings calls, or conference talks when found.
• In refresh mode: lead with "What's New Since [last date]" before the full sections.

Step 6: Save to Memory

Make all Write calls simultaneously:

• Report → ~/.nimble/memory/reports/company-deep-dive-[date].md
• Company profile → ~/.nimble/memory/companies/[company-name-slug].md (use the format in references/memory-and-distribution.md). Add [[path/entity]] cross-references for key people discovered (e.g., [[people/jane-smith]]), competitors in the same space (e.g., [[competitors/widgetco]]), and any other related entities.
• If profile exists → update last_runs.company-deep-dive in ~/.nimble/business-profile.json
• Follow the wiki update pattern from references/memory-and-distribution.md: update index.md rows for all affected entity files, append a log.md entry for this run.

The company profile in companies/ should contain structured key facts (overview, financials, leadership, products) that can be loaded by future runs of any skill that needs context on this company.

Step 7: Share & Distribute

Always offer distribution — do not skip this step. Follow references/memory-and-distribution.md for connector detection, sharing flow, and source links enforcement.

Step 8: Follow-ups

• Go deeper on a dimension → focused searches on that topic
• Compare with another company → side-by-side analysis
• "What about [specific topic]?" → targeted search + extraction
• "Looks good" → done

Sibling skill suggestions:

Next steps: - Run competitor-intel to track this company as a competitor over time - Run meeting-prep if you're meeting with someone at this company - Run competitor-positioning to analyze their messaging vs yours

Agent Teams Mode (Dual-Mode)

Check at startup: echo $CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS

Team mode (flag set): Spawn 3 teammates instead of sub-agents. Each teammate covers related dimensions and can message the others to cross-check findings.

Lead (you): Create shared tasks, wait for all teammates to complete, then synthesize the final report. When a teammate finds a claim that another should verify (e.g., a funding amount that implies a valuation), it posts a task for the relevant teammate.

Solo mode (flag not set): Standard sub-agent flow from Step 3.

What This Skill Is NOT

• Not competitor monitoring. For tracking multiple competitors over time, use competitor-intel. This skill goes deep on ONE company.
• Not meeting prep. For researching people you're meeting with, use meeting-prep. This skill researches companies, not individuals.
• Not financial advice. This is intelligence gathering from public sources, not investment analysis or due diligence certification.
• Not real-time monitoring. This produces a point-in-time report. For ongoing tracking, run it again later or use competitor-intel with the company added.

Error Handling

See references/nimble-playbook.md for the standard error table (missing API key, 429, 401, empty results, extraction garbage). Skill-specific errors:

• Search 500: Retry once without --focus flag. If still failing, retry with a simplified query (shorter terms, no date filter). Log the failure but don't skip the dimension.
• Search timeout: Retry once, then skip that call and continue — consistent with the playbook's timeout policy.
• Company not found: Retry with domain, alternative names, or parent company
• Empty results for a dimension: Note "No public data found" — don't speculate