Personal CRM for fundraising — founder's relationship management as a private repo

A separate private Git repo (e.g. founder-personal-crm or similar) housing the founder's fundraising relationships. One file per contact, one file per fund/firm, two top-level backlogs (active pipeline vs revisit-later), a tasks file separated from contact files, an inbox-processing rules file, and a small Scripts/ResearchTools/ folder with Perplexity-API-driven research scripts (contact_research.py, company_research.py, general_research.py, linkedin_research.py). Connected via a Telegram bridge so the founder can talk to it walking around. The most important asset is not the structure — it is the single AGENTS.md rule: DO NOT INVENT TASKS.

A founder doing fundraising-grade relationship management without an executive assistant, or with a lean one. Maintains professional CRM-quality discipline across hundreds of VCs, LPs, family offices, advisors, and angels. Natural-language commands ("do web research for [Person]", "do company research for [Fund]") complete in seconds. The repo doubles as personal historical record — after the round closes, you keep the corpus and re-use it the next round.

1. Create a private Git repo with restricted access (just you, or you + one operator). Never inside the team workspace — fundraising data has different exposure rules.
2. Structure under MainCRM/:

• Contacts/{First Last}/ — per-contact folder with profile.md, linkedin.md (Harmonic / Apollo dump), email-thread.md (running paste from Gmail), telegram.md (running paste from Telegram).
• Companies/{Type: Name}/ — e.g. Companies/Fund: Sequoia/, Companies/Firm: Lazard/. Minimal company profile + list of contacts at that firm we know.
• Context/Contacts_Actual.md — your active pipeline backlog (one row per active contact). Sortable, ~70 KB at any moment.
• Context/Contacts_Later.md — future-revisit with explicit re-engage condition ("return when ARR > $5M").
• Tasks/work_tasks.md — this-week to-do list (P0/P1/P2/P3). Strictly tasks YOU explicitly chose to do or that were explicitly agreed in the conversation.
• Inbox/inbox_processing_rules.md — two-stage workflow: collect all inboxes first, then process with full context.
• Scripts/ResearchTools/ — research scripts driven by Perplexity API (or equivalent).
• Specs/ — technical specs per script. Documented so the agent can debug them.
• AGENTS.md — the constitution (see below).

3. Bootstrap the research scripts. Each script wraps a single Perplexity API call with a structured prompt and writes the result into the corresponding Contacts/{Name}/research-YYYY-MM-DD.md or Companies/{Type: Name}/research-YYYY-MM-DD.md. Run from the agent: "do web research for Alex Smith" → agent invokes contact_research.py "Alex Smith".
4. Wire a Telegram bridge (see the mobile Telegram-bridge case below). The Telegram bot accepts natural-language commands and routes them to the CRM repo. Founder walks around, says: "log my call with Sequoia — I told them we'll have a follow-up deck in two weeks; they said they're tracking us; nothing else committed."
5. Set inbox-processing as a daily ritual. Once a day, "collect inbox" pulls everything from Gmail + Telegram + LinkedIn into a flat Inbox/raw-YYYY-MM-DD.md. Then "process inbox" reads it with full context (Contacts_Actual + Strategy folder), drafts updates to the right contact files, and proposes tasks — but never adds tasks unless they were explicitly committed.
6. Strategy folder is loaded for every fundraising comms. Almost all fundraising messages benefit from current company state (ARR, traction, ask). Reference the strategy folder in the master .cursorrules so the agent always knows the latest numbers before drafting.
7. Keep the file naming flat for contacts. One folder per person; PDFs/decks alongside. Mirror the candidate {Name} hiring pattern (no per-contact subfolders by status — that's a different anti-pattern: see #46).

The repo .cursorrules (the constitution — the single most important file):

You manage my personal fundraising CRM. This repo is private. NEVER push to a public remote.

CRITICAL — DO NOT INVENT TASKS
- A task lives in Tasks/work_tasks.md only if:
 (a) I explicitly said "add task: X", OR
 (b) the contact and I explicitly agreed it in a message thread I show you.
- Do NOT generate "obvious next steps". Do NOT add follow-up tasks because it
 "would be a good idea". Do NOT pre-draft "as agreed" tasks from a meeting I
 haven't yet shared with you.
- If you think a task belongs, output a "Suggested tasks (NOT ADDED)" section.
 I will decide.

CONTACT WRITES
- When updating a contact file: append, never overwrite. Each update gets a YYYY-MM-DD
 H:MM dateline. Cite source: [gmail/<thread>] or [tg/<chat>:<msg>] or [call/<file>].
- For new contacts, run Scripts/ResearchTools/contact_research.py first. Save under
 Contacts/{First Last}/research-YYYY-MM-DD.md.

STRATEGY CONTEXT
- Before drafting any fundraising message, read Docs/Strategy/Company-Overview.md
 and Docs/Strategy/Current-Round-Ask.md. Those are canonical.
- Cite numbers as [CANONICAL] or [REF: file.md#section]. Never extrapolate ARR or
 traction figures.

WHAT YOU MAY NOT DO
- Generate speculative "AS-IF-AGREED" recap texts (e.g. "as discussed, here's the deck"
 when no such agreement exists). Past failure: 2026-03 — sent a 'follow-up' to an LP
 for a meeting that hadn't happened yet.
- Auto-classify contacts as "cold" or "passed". Status lives only on my explicit say-so.

COMMANDS YOU UNDERSTAND
- "do web research for [Name]" → run contact_research.py
- "do company research for [Fund/Firm]" → run company_research.py
- "do linkedin research for [Name]" → run linkedin_research.py (requires session)
- "research: [topic]" → general_research.py
- "log [verb] [Name] [Co]: [what was said]" → append to contact file with full citation
- "update CRM" → git pull
- "commit" → git add -A; git commit; git push (private remote only)

Sample contact research script call (verbatim signature):

python3 Scripts/ResearchTools/contact_research.py \
 --name "Alex Smith" \
 --firm "Sequoia Capital" \
 --output "Contacts/Alex Smith/research-2026-05-12.md" \
 --depth deep

The script wraps a Perplexity API call with a structured prompt: "Research [Name], current role at [Firm], recent investments, recent public commentary, prior funds, education, public social presence. Cite all sources. Output as Markdown with sections: role, recent activity (last 12 months), prior investments by sector, public quotes, public social handles, possible connectors to me (if information is available)."

• DO NOT INVENT TASKS. This is the rule that justifies the case. The single hardest part of an agentic CRM is suppressing the agent's helpful-but-fabricating instinct. Encode it in .cursorrules and reinforce it every weekly tune-up.
• No per-contact subfolders by status (mirror of the hiring anti-pattern). One folder per contact, status in frontmatter. Status moves are frontmatter edits, not file moves. (Distributed anti-pattern lesson, 2026-05-05.)
• Don't share the CRM repo with the team. Fundraising relationships have different confidentiality posture than the team workspace. Some VCs assume they're "in stealth" with you; visible cross-pollination breaks trust.
• Don't bridge to Telegram without strong auth. Anyone with the bot token has access to your fundraising data. Per-user bot, restricted to your Telegram user ID only.
• Don't auto-classify "cold" or "passed". Status is a deliberate human act. The agent only proposes.

• Lighter version (pre-fundraise): drop the Perplexity scripts. Just file structure + the .cursorrules. Useful for any founder managing >50 high-value relationships (partnerships, advisors, hires).
• Heavier version (with a chief of staff): add an inbound-routing layer: every email/Telegram message to the founder gets pre-classified into "needs me / can be handled / FYI", with handle-able items routed to the chief of staff via a separate cos/ repo (see #211). Founder reviews summary once a day.
• Investor side: flip the schema — Investments/, Pipeline/, Watchlist/, with the same constitution rule that the agent must never auto-progress a deal.
• Live-event capture variant — drop a phone screenshot of a referral; OCR plus voice prompt creates a vault card before the conversation ends.