Prospect Research

Prospect Research

Turn a name + LinkedIn profile into a short, high-signal brief a human can act on — and (optionally) a personalized first message grounded in that brief. Read-only: it gathers and writes a brief; it never sends a message or modifies the outreach system.

Cost discipline (read first)

The expensive part is raw tool output (LinkedIn profile text, Apollo JSON, scraped pages), not model cleverness.

• When researching more than one person (a batch), run each person in a subagent. The subagent ingests the bulky raw tool output in its context and returns only the compact brief (+ draft). The caller's context stays lean. This is the single biggest cost lever.
• A mid model (Sonnet) is the right default for the gather+synthesize+draft work — capable enough for a non-generic draft, far cheaper than the top model. Don't use a tiny model for the drafting step — that's where quality (the whole point) lives.
• Trim the gather: cap company posts to the top ~3, run at most one targeted web search, and skip low-value calls (e.g. full employee lists) unless the task needs them. Stop gathering once the six sections can be filled with reasonable confidence.

Step 0 — Detect tools + load config

Detect which research tools are present: mcp__linkedin__* (person/company/posts/search), Apollo (apollo_people_match, apollo_organizations_enrich, apollo_organizations_job_postings, apollo_contacts_search), Firecrawl/Exa (web search). Use whatever exists; degrade gracefully. Read optional .claude/prospect-research.md for: the offer/value-prop (what we're selling — sharpens pain-points and hooks), the message voice spec + drafting on/off, model + caps.

Step 1 — Gather (priority order, stop when enough)

1. Person — LinkedIn get_person_profile(slug): current role, tenure, headline, and recent activity/posts (the best personalization hook). If an Apollo contact is supplied, apollo_people_match for title/seniority. If the supplied (Apollo) title disagrees with the LinkedIn headline, flag the drift and lower Quick-Facts confidence — Apollo titles can be stale (e.g. post-acquisition the person still shows their old founder/co-founder headline).
2. Company — LinkedIn get_company_profile + get_company_posts (top ~3); Apollo organizations_enrich for size/funding/industry.
3. Tech / AI-tooling signals — the highest-value signals often live in LinkedIn company posts (get_company_posts): a hosted AI-coding event (a Claude Code / Cursor meetup), a first-party tool built for AI coding agents, an engineer demoing AI-coding workflows. Check these first — they don't depend on web search (which is frequently down). Then Apollo organizations_job_postings (hiring for AI/ML/platform/eng roles = adoption signal) + a Firecrawl/Exa search on the company site/careers/news for stack & AI initiatives.
4. News hooks — one Firecrawl/Exa search: <company> funding OR launch OR AI OR leadership 2026.

Step 2 — Synthesize the brief (6 sections, compact)

1. Quick Facts — name, role, tenure, location, career arc.
2. Company Panel — what they do, size, funding/stage, industry.
3. Tech / AI-tooling signals — stack, AI initiatives, relevant hiring.
4. Recent Activity / News hooks — recent posts, launches, funding, leadership changes (with the specific hook to reference).
5. Connection points + 3 conversation starters + 3 likely pain points — tied to the offer if config provides one.
6. ICP fit verdict — given the project's ICP (from config), is this person/company actually a buyer? Strong / Moderate / Weak / Not-a-fit + one-line rationale + confidence. The Strong/Moderate divider: Strong requires a concrete public signal that their own engineers use the relevant tools — and it does NOT have to come from the prospect personally; a clear company-level signal counts when this person owns that eng org. Qualifying signals: an engineering blog or a leader's post on AI-coding adoption; a stated adoption %/goal; job posts naming the tools; a company-hosted AI-coding event (a Claude Code / Cursor meetup), a first-party product built for AI coding agents, or an engineer publicly demoing AI-coding workflows — these are proof of real internal use, not just marketing. Right company-type + right title + right size but no such signal at all → cap at Moderate, however good it looks. (Don't over-discount a strong company signal just because the named person hasn't personally posted it — that mistake systematically under-rates good fits.) Watch the trap: a company that builds its own AI product is NOT automatically a fit — judge whether their own engineers/users would feel the pain the offer solves, at enough scale to matter, and whether THIS person owns/champions that area. Be willing to say "not a fit" — a clear disqualification is more valuable than a forced pitch.
7. Research confidence — high/medium/low per section, so the message-writer knows how much to trust each field. Never invent facts; mark gaps as "unknown".

Step 3 — Draft, GATED on fit (if config enables it)

Only draft if drafting is on AND the ICP fit verdict (§6) is Moderate or better. If fit is Weak / Not-a-fit, do NOT draft — return the brief with a skip: not ICP recommendation and the one-line reason, so the caller can log it and move on. Don't waste a personalized message (or a real person's attention) on a non-buyer.

When you do draft: write ONE personalized message grounded in the brief — open on a specific hook from §4/§5, follow the voice spec exactly, weave in the configured offer hooks, stay concise. Output the brief + the draft. Do not send.

Step 4 — Return

Return a compact result: the 6-section brief and (if drafted) the message. Nothing is sent and nothing in the outreach system is modified — the caller decides what to do (e.g. route the draft for human approval).

Failure modes

Symptom	Action
LinkedIn session invalid	Note it; proceed with Apollo + web only; lower confidence.
A tool/MCP missing	Use what's present; mark affected sections lower-confidence.
Rate limit hit	Respect the config cap; return a partial brief rather than hammering.
Thin/no public data	Return a short brief marked low-confidence; flag "manual research recommended" and skip the draft.

Reuse / extension

Carries no project specifics. A project supplies .claude/prospect-research.md with the offer, voice spec, drafting toggle, model + caps. Edit /home/vas/projects/aictrl/.claude/prospect-research.md for aictrl — not this skill. Optional upgrade: add the Exa MCP if Firecrawl news relevance proves weak.