AI-driven Pipeline Development

I’m continuing my exploration of how to use AI to drive various fundraising tasks. The big one I’m tackling right now is pipeline development. It’s probably the most resource-intensive of all the tools so far.

AI wants to be lazy. A lazy pipeline–a list of obvious billionaires–isn’t useful. It’s a Google search. Al also tends to drift. To lose fidelity and focus over each iteration. That risks the utility, as well.

So I’ve learned, when the AI is doing what I want, to generate a set of instructions and guidance to its future self. I record it elsewhere–usually on this blog–so I can copy it back into the thread to reset the focus.

This is the set of instructions it’s generated for how to build a good pipeline. It’s not a human-focused “how to”. But it carries good lessons and ideas that might be helpful to human researchers, as well.

(Future Me — Full-Fidelity, Token-Safe Process)

Purpose: Build an insider-grade, four-pass donor pipeline at full density without drift or compression, within token limits, using a batch UX that guarantees paste-ready outputs.

(Future Me — Full-Fidelity, Token-Safe Process)

Purpose: Build an insider-grade, four-pass donor pipeline at full density without drift or compression, within token limits, using a batch UX that guarantees paste-ready outputs.

A) Prime Directives (No Shortcuts)

  1. Run all four passes (1→4) every time unless the user explicitly stops. Do not stall after Pass 1.
  2. Full density always. Each donor profile includes:
    • Background: 2–3 full sentences, story-driven.
    • Philanthropy Focus: 2–3 sentences with specifics/examples.
    • Boards/Affiliations: complete, relevant.
    • Capacity: source of wealth + historic gift scale.
    • Tier + Tier Rationale: full sentence rationale.
    • Fit (5 bullets): each bullet = 2–3 sentences (Alignment bridge; Decision trigger; Engagement angle; Risk/differentiator; Local hook/vector).
  3. Surprise > celebrity. Deep cuts first; marquee names second.
  4. Every profile is bespoke. If two Fit bullets sound similar, rewrite both.
  5. Evidence discipline. Anchor claims with verifiable context; tag peer-list sources explicitly in the Fit when applicable.
  6. Never compress to “save tokens.” If output nears limits, stop mid-batch and continue in the next message — do not summarize.

B) Output UX & Delivery Protocol (Token-Safe)

You cannot produce a single downloadable workbook reliably. Therefore:

  1. Batch size: Output 1–2 donors per message at full density. If the user explicitly asks for larger batches, cap at 5 donors per message while preserving full density.
  2. User workflow: After each batch, instruct the user to copy/paste the profiles into their sheet (or doc) and reply “CONTINUE”. Do not proceed until they do.
  3. Progress header: Prepend every batch with a one-line tracker: PROGRESS — Pass: [1|2|3|4] •
  4. Batch: [n] • Donors this batch: [m] • Total delivered (unique): [X]Resume protocol: If interrupted, read the last PROGRESS header and resume at the next donor/pass. Never repeat donors unless the user requests it.
  5. Column mapping: You are delivering prose, not files. Ensure each donor block includes fields in this order so the user can paste to columns.
    • Donor
    • Background
    • Philanthropy Focus
    • Boards/Affiliations
    • Capacity
    • Tier
    • Tier Rationale
    • Fit (Bullets) — 5 bullets, each 2–3 sentences
  6. CSV/TSV option (only if user requests): When asked, emit per-batch TSV with bullets joined by the literal token \n. Tell the user: “Paste into Sheets/Excel, then Find \n → Replace with line break (Excel: Ctrl+J).”

C) Four-Pass Plan (What to Ship in Each Pass)

Pass 1 — Scaffolding (fast, credible)

  • Deliver 10–15 aligned names that establish lane legitimacy (not random billionaires).
  • Include 2–3 validators (funders of closely adjacent work).
  • Batch: 1–2 donors/message (max 5/message if user demands). Continue until the pass total is met.

Pass 2 — Deep Cuts (slow, research-heavy)

  • 15–25 under-the-radar decision-makers: quiet heirs, regional dynasties, second/third-gen wealth, low-profile finance/real estate families, influential program officers/trustees who originate grants.
  • For each: full profile + bespoke Fit.
  • Batch as above; no compression.

Pass 3 — Peer-List Harvest (evidence-driven)

  • Mine 4–6 peer org donor lists; pull names with ≥$100K or multi-year support; tag the source org explicitly in the Fit (e.g., “Source tag: Aspen honor roll”).
  • For each donor: capture where seen, typical gift band, thematic fit, and write full-density profile.
  • Do not replace deep cuts with celebrity peer donors; keep the “surprise” ratio high.

Pass 4 — Clean, Tier, Bespoke (finalize)

  • Deduplicate across passes; remove obvious/no-fit/out-of-scope
  • Assign Tier = Anchor / Network Leverage / Quiet Deep Pockets with a one-sentence Tier Rationale.
  • Rewrite any repeated language so every Fit is unique and actionable.
  • Deliver final batches until the user confirms they have 40–60 total profiles (or a user-specified target).

D) Data Model & Formatting

  • Field order (must match user sheet):
  • Donor
  • Background
  • Philanthropy Focus
  • Boards/Affiliations
  • Capacity
  • Tier
  • Tier Rationale
  • Fit (Bullets) — 5 bullets; each bullet 2–3 sentences; include source tags in Pass 3 entries.
  • Fit bullets content (all 5, full sentences):
    • Alignment bridge (why here, donor-specific).
    • Decision trigger (timing, hook, pledge/portfolio window).
    • Engagement angle (tone, venue, who speaks, what to show).
    • Risk/differentiator (what to avoid, how we differ).
    • Local hook / vector path (who introduces; city/table/board).

E) Style & Density Rules (Enforce)

  • Density: Every bullet is a mini-paragraph (2–3 sentences).
  • No filler stems (“values-aligned,” “innovative”) unless tied to concrete evidence.
  • Peer tone: Write as if prepping a senior colleague.
  • Zero repetition: If a phrase repeats across donors, rewrite for variety and precision.

F) Research Sources (Use All That Apply)

  • Peer org artifacts (annual reports, honor rolls, gala chairs, anniversary lists).
  • 990-PF / charity registers (trustees, outflows, rhythm).
  • Board rosters & minutes (museums, hospitals, universities, federations).
  • Press & business journals (pledges, succession, liquidity, obits for lineage).
  • Regional philanthropy lists (Crain’s, museum patron circles, symphony guilds).
  • Hidden wealth sectors (logistics, waste, aggregates, regional banks, dealerships, construction, timber/paper, food distribution, agribusiness).
  • Political bundling disclosures (finance committees → philanthropic tables).
  • Program officers: include gatekeepers with real discretion.

Verification: For each donor, anchor at least two of: grant mention, board listing, press piece, 990-PF line, annual report. If uncertain, omit or label as inference (e.g., “Signal: consistent $500K+ to X suggests comfort at $1M+”).

G) Heuristics for “How to Engage” (by donor type)

  • Tech/VC: Brevity, KPIs, cost-to-impact math, scale path, platform play.
  • Political/campaigners: Stakes, momentum, coalition leverage, co-chair framing, time-boxed wins.
  • Family/legacy: Lineage & continuity; “extend your family’s impact into X”; visible leadership, minimal busywork.
  • Arts/culture/media: Story, representation, place-making; convene tastemakers; lift fellows’ voices.
  • Quiet finance/real estate: Discreet impact, community stability, workforce pipeline; minimal publicity.
  • STEM/science: Evidence, persistence, talent diversification; partner with labs & internships.
  • Program officers: Mirror portfolio thesis & vocabulary; propose pilot with learning agenda + evaluation.

Always specify tone calibration (e.g., “peer-to-peer strategist,” “quiet/off-press,” “campaign-style energy”).

H) Anti-Patterns (Never Do)

  • ❌ One-size “Fit” bullets; if two read alike, rewrite both.
  • ❌ Laundry lists of celebrities; keep 1–2 for legitimacy max.
  • ❌ Hollow claims without evidence bridges.
  • ❌ Four-bullet Fits; must be 5.
  • ❌ Walls of prose without field structure.
  • ❌ Shipping without Donor + Fit mapping intact.

I) Tiering & Prioritization

  • Anchor (6–10): legitimacy + capacity + time-horizon fit; “funds 3 peer orgs at $1M+; chair potential.”
  • Network Leverage (8–12): conveners/validators who open rooms and unlock co-funding.
  • Quiet Deep Pockets (8–12): low-profile, high-throughput; emphasize discreet approach.

Include a one-line Tier Rationale for each.

J) Regionalization (Adapt to Org)

  • Chicago/Midwest (civic families; federations; hospital/higher-ed; bundlers; quiet real estate/distribution).
  • Bay Area (tech founders/second-gen; low-profile FOs; community foundations; Stanford/Berkeley boards).
  • NYC (hedge/PE; arts patrons; hospitals/universities; legacy families; DAFs).
  • Texas/Houston/Dallas (energy/logistics heirs; regional banks; sports owners).
  • Miami/South Florida (LatAm family capital; real estate dynasties; arts patrons).
  • Always localize hooks in Fit (neighborhoods, federations, flagship institutions, alma maters).

K) Time & Pass Accounting (Progress, not promises)

  • Track counts, not minutes.
  • Pass 1: deliver 10–15 donors.
  • Pass 2: deliver 15–25 donors.
  • Pass 3: deliver peer-list donors from 4–6 orgs (≥$100K/multi-year), source-tag each in Fit.
  • Pass 4: dedupe, tier, bespoke rewrites; continue batches until 40–60 total (or user-specified target).
  • Prepend every batch with the PROGRESS header.

L) Recovery & Resume Protocol

  • If interrupted or the user disappears, do not guess. When the user returns, read the last PROGRESS header and resume.
  • If a batch cut mid-donor, finish that donor first in the next message.
  • Never recap past donors unless the user asks; keep moving forward.

M) Human-in-the-Loop Tasks (Make them explicit)

  • After each batch, instruct the user to paste into their sheet and reply “CONTINUE.”
  • If the user wants TSV/CSV, provide TSV per batch and remind them to convert \n to line breaks (Excel: Find \n → Replace Ctrl+J).
  • If the user asks for edits per donor, apply edits, then continue the queue.

N) Escalation Triggers (Tighten Quality)

  • Too many marquee names → replace ≥50% with deep cuts.
  • Repeated language → rewrite for voice variety/specifics.
  • Thin evidence → pause and pull one primary source (peer list, 990-PF).
  • Misalignment risk flagged → add a Fit bullet differentiating from peer programs.

O) Final Principle

If an insider reads a batch and says, “How did you even find these people?”, we did it right. If they say, “I’ve seen this list before,” start over.

Execute now using this UX:

  • Batch 1–2 donors/message at full density.
  • Wait for “CONTINUE.”
  • Maintain PROGRESS header.Complete Passes 1→4 to the target count without compression.

Leave a comment