Imported from karthik-pj/careset2 (.local/secondary_skills/ai-sdr/SKILL.md). Install upstream with npx skills add karthik-pj/careset2 --skill ai-sdr. Copyright stays with the author.

AI SDR

Find relevant companies and leads for B2B sales. Define ideal customer profiles, identify target accounts, qualify prospects, and organize research for outreach.

When to Use

• User wants to find companies that match their ICP
• User needs to build a prospect list for sales outreach
• User wants to research target accounts
• User needs lead qualification analysis

When NOT to Use

• Writing outreach sequences (use sdr-outreach or cold-email-writer skills)
• Recruiting candidates (use ai-recruiter skill)
• General market research (use deep-research skill)

Methodology

Step 0: Interview the User About Their Business

Before doing any research, interview the user. Most users won't give you enough context unprompted — they'll say "find me customers" without explaining what they sell or who buys it. Use multi-option quizzes to make this fast and low-friction. Ask one question at a time.

Question flow:

1. What do you sell?

   • A) Software / SaaS
   • B) Physical product
   • C) Professional services / consulting
   • D) Marketplace / platform
   • E) Something else (describe)

2. Who buys it?

   • A) Other businesses (B2B)
   • B) Consumers (B2C / DTC)
   • C) Both

3. (If B2B) What size companies?

   • A) Startups / small teams (1-50 people)
   • B) Mid-market (50-500)
   • C) Enterprise (500+)
   • D) Not sure yet

4. What's your price point?

   • A) Under $100/mo
   • B) $100-$1,000/mo
   • C) $1,000-$10,000/mo
   • D) $10,000+/mo or custom pricing
   • E) One-time purchase

5. Who inside the company makes the buying decision?

   • A) Engineering / technical (CTO, VP Eng, developers)
   • B) Marketing (CMO, growth, content)
   • C) Sales / revenue (CRO, VP Sales, RevOps)
   • D) Operations / finance (COO, CFO)
   • E) Founder / CEO directly
   • F) Not sure

6. Do you have existing customers?

   • A) Yes, paying customers — I can describe who they are
   • B) A few early users / pilots
   • C) No customers yet

7. (If they have customers) What do your best customers have in common? (free text — this is the most valuable answer)

8. Any industries or verticals you're focused on? (free text or skip)

If the user provides a detailed prompt upfront, skip the questions they already answered. Don't re-ask what's obvious. But if key info is missing (who buys, what size, what price), ask before proceeding — the ICP will be wrong without it.

Step 1: Define ICP (Ideal Customer Profile)

An ICP describes accounts where three things are true: they can buy (budget/size fit), they will buy (pain exists now), and they will stay (retention profile). Pick 6-10 attributes — more than 10 and nothing qualifies.

ICP vs Buyer Persona: ICP = which company. Persona = which human inside it. A Series B fintech (ICP) has a VP Eng who cares about velocity and a CFO who cares about cloud spend (two personas, different messaging).

Step 2: Source Accounts

Free sources (use webSearch + webFetch):

• Crunchbase — site:crunchbase.com "series a" fintech 2025 for funding events
• LinkedIn — site:linkedin.com/company [industry] "11-50 employees" (headcount filter leaks into page text)
• BuiltWith / Wappalyzer lookups — webFetch a prospect's homepage, then scan source for tech signatures (Segment snippet, Intercom widget, Shopify checkout)
• Job boards — site:linkedin.com/jobs "[target company]" "data engineer" reveals what they're building; site:greenhouse.io and site:lever.co for startup hiring
• G2 / Capterra category pages — companies reviewing competitors are in-market
• GitHub orgs — public repos reveal tech stack and eng team size for dev-tool ICPs
• SEC EDGAR (public cos) — 10-K "Risk Factors" sections list the exact problems they're worried about

Paid sources the user likely has (shape output for these):

• Apollo (~210M contacts, $49+/mo) — best value for SMB/mid-market, filters on headcount growth + job postings + intent
• LinkedIn Sales Navigator (~1B profiles) — most accurate job-change data, but no email export
• ZoomInfo — strongest US enterprise coverage + intent data (tracks content consumption across the web)
• Clay ($134+/mo) — waterfall enrichment: chains Apollo → Hunter → Cognism to maximize match rate. Best for teams with RevOps capacity.
• Cognism — best EU/UK data + phone-verified mobiles (GDPR-compliant)

Step 3: Buying Signals (Trigger Events)

Prospects with an active trigger convert 3-5x higher. Rank by signal strength:

Step 4: Qualify & Tier

Fast disqualification (do this first): Before researching, kill accounts that fail any hard constraint — wrong geo, below headcount floor, competitor customer under contract, recent layoffs (no budget).

Qualification frameworks:

• BANT (Budget / Authority / Need / Timeline) — fine for transactional/SMB
• MEDDPICC (Metrics / Economic buyer / Decision criteria / Decision process / Paper process / Identify pain / Champion / Competition) — use for enterprise deals >$50k. The extra P's (paper process, competition) matter because enterprise deals die in legal/procurement, not in the pitch.

Tiering:

• Tier 1 — ICP match + active trigger in last 30 days → full personalization, multi-channel, SDR owns it
• Tier 2 — ICP match, no trigger → lighter-touch automated sequence, monitor for triggers
• Tier 3 — Partial fit → newsletter/nurture, revisit quarterly

Step 5: Scale with Parallel Agents

Target minimum 40 prospects. A single sequential search won't get there fast enough. Use startAsyncSubagent to run 5 parallel research agents, each focused on a different search angle:

1. Industry/vertical search — companies in the target vertical via Crunchbase, G2 category pages
2. Funding/growth search — recently funded companies matching the ICP
3. Hiring signal search — companies hiring for roles that indicate they need the user's product
4. Competitor customer search — companies using competitors or reviewing them on G2/Capterra
5. Lookalike search — competitors and alternatives to any strong-fit companies already found

Each agent should return 10-15 prospects with all columns filled in. Use the wait_for_background_tasks tool to wait for them, then deduplicate and merge into the final spreadsheet.

Step 6: Output as a Spreadsheet

Build a real spreadsheet using the excel-generator skill or write a CSV file — not a markdown table. The output should be something the user can import directly into a CRM, Clay, or Apollo.

Columns:

| Company | Domain | Headcount | Fit Score (1-5) | Trigger | Trigger Date | Target Contact Name | Target Title | LinkedIn URL | Email (if found) | Why Now (1 sentence) |

• LinkedIn URL — search for the target persona at the company: site:linkedin.com/in "[company]" "[title]". Include direct profile links.
• Email — look for email patterns via webSearch("[company] email format" OR "[name] [company] email"). Common patterns: first@company.com, first.last@company.com. If not found, leave blank and note the likely format.
• Why Now — the most valuable column. It's the first line of the cold email.

The spreadsheet should be downloadable and ready to import — not just displayed as text.

Deep Research for Complex ICPs

For industries or markets you don't know well, pull in the deep-research skill to build context before prospecting. This is especially useful when:

• The user sells into a niche vertical you don't have strong priors on (e.g., "construction tech", "veterinary SaaS")
• You need to understand market landscape, key players, and buyer behavior before defining the ICP
• The user wants competitor analysis as part of the prospecting process

Use deep-research to gather industry context, then return here to build the prospect list with that knowledge.

Agent Tactics

Tech stack detection: Use webSearch("[company] tech stack" OR "[company] built with") to find BuiltWith/Wappalyzer/StackShare profiles. Common signatures to look for in search results:

• Segment, Intercom, Stripe, Shopify, Google Analytics
• Job postings often reveal stack: webSearch("[company] careers engineering")

Note: webFetch returns markdown content, not raw HTML — script tags and asset URLs are stripped. Use search-based detection rather than HTML source scanning.

Waterfall search pattern: If one query returns nothing, don't stop — try synonym variants. "VP Engineering" OR "Head of Engineering" OR "Engineering Lead" OR "CTO" all map to the same persona at different company sizes.

Lookalike expansion: Once you find 5 good-fit accounts, search for their direct competitors — webSearch: "[good-fit company] vs" or "[good-fit company] alternatives" surfaces the category.

Best Practices

1. 50 researched > 500 sprayed — reply rates on researched lists run 3-5x higher
2. Disqualify before you qualify — negative filters are cheaper to check
3. Trigger freshness decays fast — a funding round is a 60-day window, a job change is a 90-day window
4. Enrich once, cache the result — don't re-research the same account every sequence
5. B2B data decays ~30%/year — any list older than 6 months needs re-verification

Limitations

• Cannot log into LinkedIn Sales Navigator, Apollo, ZoomInfo, or Clay — builds search strategies the user executes
• Cannot verify email deliverability (user should run through NeverBounce/ZeroBounce before sending)
• Cannot detect intent data (Bombora/6sense-style content consumption signals require paid platforms)
• Company headcount/revenue estimates from public web are approximate — private company data is inherently fuzzy