VC Firm Profile Ingest

One instruction — “create a profile for Kalos Ventures” — produces a CRM record, an attached research note, and a wiki page that cross-references the CRM record, with no copy-paste between systems.

This is the two-system sibling of [[crawl-fetch-ingest]]. That skill ingests to files for decks and slides. This one ingests to live systems a client team actually works in.

Prerequisites

Two connectors in the operator’s Claude Desktop (or claude.ai):

Both exist today. When homebase ships ([[Homebase-MCP-One-Connector-Per-Client]]) this collapses to one connector and this skill becomes an MCP prompt — the procedure below does not change.

Client constants (Palmer AI)

⚠️ Collection ambiguity — confirm before writing. A third collection, Firms (0022c7e6-87f5-45f0-b644-25c62ee21ece), also exists and has a better description, but holds one page (Colossus) while Venture Firms holds the five real profiles with revision histories. This skill writes to Venture Firms. If the operator consolidates, update this table — do not guess per run, and do not split profiles across both.

The target state — what “done” means

Three artifacts and one join:

Twenty: company record ──────┐
Twenty: note ──(note_target)─┘   attached to the company
Outline: page in Venture Firms, header carries the Twenty record UUID

The join is the point. A profile that exists in only one system, or in both without the cross-reference, is not done.

Procedure

Step 0 — Dedupe. Always. Both systems.

There are already 98 companies in Twenty and profiles in Outline. Creating a second record for a firm that exists is the most likely failure of this skill and the most annoying to clean up.

find_many_companies  select:["id","name","domainName","description"]
                     name:{ilike:"%<firm>%"}

Also try the distinctive word alone — “Owl” not “Owl Ventures” — and the domain, since a record may exist under a variant name (Andreessen Horowitz (a16z)).

Then search Outline: list_documents with the firm name, and list_documents filtered to the Venture Firms collection.

Branch:

• Neither exists → full create, all steps.
• Twenty record exists, no page → skip Step 3’s create; reuse that UUID.
• Both exist → this is an update, not an ingest. Report what’s already there and ask what to refresh. Never create a parallel page.

Step 1 — Research, in source order

Collect from the open web, preferring sources in this order:

1. The firm’s own site — /team, /portfolio, /about, /for-entrepreneurs. Authoritative for people, titles, stated thesis, screening criteria.
2. LinkedIn company page — headcount sanity check, recent announcements.
3. Named primary sources — a partner’s own fund-announcement post, an SEC filing, the firm’s impact report.
4. Reputable press — for fund size and vintage when the firm doesn’t publish it.

Do not source from VC aggregator profiles (VCSheet and similar). A real example, recorded on the New Markets page: its VCSheet profile is contaminated with an unrelated fund’s content and names the wrong founder. If an aggregator is the only source for a fact, either omit the fact or record it as an open question.

Collect: stage focus, sector focus, thesis in the firm’s own framing, stated screening criteria, team (name / role / LinkedIn / email if published), advisors and network affiliations, notable portfolio, HQ, fund size and vintage, and anything that would change how a partner approaches them.

Step 2 — Stage to a scratch file first

Write findings to a scratch file before touching either system:

<scratchpad>/vc-profile-<firm-slug>.md

Why this is not optional: research is the slow, failure-prone half. If a write fails midway you re-run the writes, not the research. It also gives the operator one artifact to correct before anything reaches a client’s live CRM.

Show the operator the staged file and get a go-ahead before Step 3 when the firm is new. For refreshes of an existing record, proceed.

Step 3 — Create the Twenty company record

create_one_company
  position: "first"                      ← REQUIRED, and easy to forget
  name: "New Markets Venture Partners"
  description: "<dense prose — see conventions below>"
  domainName: { primaryLinkLabel: "newmarketsvp.com",
                primaryLinkUrl:   "https://www.newmarketsvp.com",
                secondaryLinks:   [{url: "...", label: "Portfolio"}] }
  linkedinLink: { primaryLinkLabel: "new-markets-venture-partners",
                  primaryLinkUrl: "https://www.linkedin.com/company/..." }
  address: { addressCity: "Baltimore", addressState: "MD",
             addressCountry: "United States" }

Field conventions observed in the existing 98 records:

• domainName.primaryLinkLabel is the bare domain (a16z.com), the URL is full.
• linkedinLink.primaryLinkLabel is the LinkedIn slug (south-park-commons).
• secondaryLinks carry notable sub-pages — an accelerator programme, a press page, a specific fund-announcement post. Use them; they are how a16z and South Park Commons stay legible.
• address is usually city/state/country only; street is typically empty.
• description is a dense paragraph, not a tagline — thesis, fund history, named partners, portfolio highlights. The best existing records run 100+ words.

Gotcha — link fields are composite objects. Passing domainName: "a16z.com" as a plain string does not error usefully; it just fails to populate. Same for linkedinLink.

Gotcha — position is required on create_one_company, create_one_note, and create_one_note_target. Use "first".

Step 4 — Attach a research note (two calls, not one)

Twenty does not let you attach a note to a company in one call. Notes are free-floating until a note_target links them.

1. create_one_note        position:"first"
                          title: "Research — <firm> (<YYYY-MM-DD>)"
                          bodyV2: { markdown: "<the findings>" }
   → returns noteId

2. create_one_note_target position:"first"
                          noteId: <from step 1>
                          targetCompanyId: <the company id>

Skipping call 2 is the classic failure — the note exists, looks fine in a tool response, and is invisible on the company record.

The note carries the research narrative and its sources; the company description carries the durable summary. Put source URLs in the note.

Step 5 — Create the Outline page

create_document with collectionId = Venture Firms, title = the firm’s name exactly as in Twenty, publish: true, and text = the template below.

Step 6 — Write the join

The page’s first line carries the Twenty record UUID. This is what makes the two systems one system:

**Website:** https://www.newmarketsvp.com **Twenty CRM record:** `0c757f61-…` **Last reviewed:** 2026-07-27 (sourced from firm website)

Step 7 — Verify, don’t assume

• find_many_companies by id → the record has description and links populated
• The note appears on the company, not just in the notes list
• fetch the new document → the Twenty UUID is present and correct
• Report the Outline URL and the Twenty record id to the operator

The page template

Reproduce this structure. It is the house shape, derived from the five existing profiles — not an invention.

**Website:** <url> **Twenty CRM record:** `<uuid>` **Last reviewed:** <YYYY-MM-DD> (sourced from <source>)

## Snapshot

|     |     |
|-----|-----|
| Stage focus | |
| Sector focus | |
| Framing | |
| Track record | |
| Reported reach | |
| Team size | |

## Thesis

<Prose: what they believe and why, in their own framing. Then the firm's stated
investment beliefs as bullets.>

**Sourcing edge.** <Where their deal flow actually comes from, and whether it is
genuinely differentiated.>

### What they screen for

<Published criteria — useful as a comparison point for our own scorecard work.>

## Key people

| Name | Role | LinkedIn | Email |
|------|------|----------|-------|

## Notable portfolio

## Portfolio overlap

*HumainVC-specific — to fill in.*

## Relationship history

| Date | What happened | Who |
|------|---------------|-----|
|      |               |     |

## Working notes

<Anything an operator should know before a meeting. Include data-hygiene warnings
here — e.g. "the VCSheet profile is contaminated, do not source from it".>

## Open questions

- [ ] Current fund size and vintage
- [ ] Check size range, and whether they lead
- [ ] Confirm HQ location
- [ ] Primary point of contact for HumainVC

The open-questions checklist is load-bearing. Every existing profile carries 4–5 unchecked boxes, and Outline surfaces them as task counts. An honest profile names what it does not know. Never fabricate a fund size to empty the list.

Voice

Match the existing pages: analytical, skeptical, useful to a partner walking into a meeting. They read like a competent associate’s memo, not marketing copy.

• Say what is claimed vs. what is verified — “reported reach 75M+ learners”
• Flag contradictions rather than smoothing them — “reconcile ‘early and growth stage’ vs. newer growth-stage site language: has the mandate moved later?”
• Note what is unusual and why it matters — “the site publishes direct email for every team member, in <first-initial><lastname>@ form. Unusually open.”

Anti-patterns

• Creating a duplicate because Step 0 was skipped or matched too narrowly.
• Note without a note_target — invisible on the record.
• Plain strings in domainName / linkedinLink — silently unpopulated.
• Writing to Firms instead of Venture Firms — see the ambiguity warning.
• A page with no Twenty UUID — two disconnected systems, which is the problem this skill exists to solve.
• Filling every field to look complete. Unknowns belong in Open questions.
• Sourcing from aggregator profiles without a primary-source check.
• Writing to a client’s live CRM before the operator has seen the staged file on a first-time ingest.

Extending to other clients

Only the constants are Palmer-specific. To use this for another client, swap the collection IDs and the Twenty connector, and confirm the target collection’s page template — a client with a different profile shape needs its own template block, not this one bent to fit.

Related

• [[crawl-fetch-ingest]] — the file-based sibling for decks and slides
• [[Homebase-MCP-One-Connector-Per-Client]] — where this becomes one connector and an MCP prompt
• client-stacks/palmer-ai/stack.md — the Palmer AI stack of record