Lighthouse Chart — Priya Menon

Investor / Analyst Track | 4 Weeks | 2 sessions/week | 45 min/session

1. Opening / Bearings

You have spent 22 years building a judgment infrastructure that most investors never develop — a proprietary diligence framework earned through hundreds of deals, not borrowed from a template. The question you are bringing here is not "can AI help me?" but the sharper one: "where exactly does it help, and where does it introduce noise I cannot afford?" That is the right question, and it is the one this chart is built to answer.

The outcome you named — wiring AI-assisted research into your standard workflow without losing signal — is precise, and precision is what makes it achievable in four weeks. The risk you are managing is real: AI is genuinely useful on the public-information leg of diligence, and genuinely dangerous when it starts filling gaps with confident-sounding fabrications. The plan below keeps those two zones clearly separated.

By the end of week four, you will have run AI-assisted diligence on four deals from the demo day batch, produced a measurable time-to-memo comparison, and codified a one-page framework that tells you — and eventually anyone you bring into the process — exactly which steps belong to the machine and which belong to you.

2. Track Context

The investor/analyst track is where AI earns its keep fastest and loses trust fastest — often in the same session. The leverage is real: competitive landscape mapping, market sizing synthesis, memo scaffolding, and public-source aggregation are all genuinely compressible with the right prompts. The failure mode is equally real: AI will produce a funding round figure with the same confident tone it uses for a fact it actually knows, and if you are moving fast, you will miss it.

Your existing toolkit — Perplexity for research, Notion for documentation, occasional ChatGPT — already reflects a practitioner's instinct about where AI fits. Perplexity is well-suited to the public-source research leg because it surfaces citations you can verify. Claude is better suited to the synthesis and drafting leg: longer context window, more reliable instruction-following on structured outputs, and — critically for your constraint — a Pro account you already have that keeps your inputs off a shared training pipeline.

The constraint that proprietary deal information cannot go into third-party LLMs is not a limitation to work around — it is a structural boundary that actually clarifies the workflow. Everything before the NDA and the first substantive founder call is public-information territory where AI can work freely. Everything after — cap tables, financial projections, customer names shared in confidence, your own thesis notes — stays offline. The plan below is built around that line, not despite it.

One pattern worth naming early: the investors who get the most out of AI are not the ones who automate the most. They are the ones who use AI to get to a high-quality first draft faster, then apply their judgment to that draft rather than to a blank page. The blank-page problem is where the time goes. That is what this plan targets.

3. Goals & Success Metrics

Priority order:

1. Wire AI into the diligence workflow — specifically the public-information research leg and the memo first-draft scaffold.
2. Cut time-to-first-memo by 40% — measured on the fourth deal against your pre-AI baseline.

What done looks like:

• Three diligence memos drafted with an AI-assisted research leg. Not AI-written memos — memos where AI compressed the research and scaffolding time, and you applied judgment to the output.
• A measurable time-to-memo reduction on the fourth deal. "Measurable" means you have a before number and an after number. The 40% target is ambitious for four weeks; the more important thing is that you have a real comparison, not an impression.
• A one-page framework — the AI-in/out map — that captures which steps are AI-eligible and which are judgment-only. This is the durable artifact. The memos are practice; the framework is the outcome that scales.

What this plan explicitly does not touch:

• The judgment calls that define your fund thesis. Which deals fit your portfolio construction logic, which founders you trust, which markets you believe in — none of that is on the table. AI has no visibility into your LP commitments, your portfolio dynamics, or the pattern recognition you have built across 15 years of venture. It should not.
• Portfolio company reporting. Out of scope for this four-week sprint.

A note on the 40% target: Your current time-to-first-memo is the baseline. If you do not have a precise figure, estimate it before Session 2 — even a rough number ("usually takes me about four hours from first call to draft") gives you something to measure against. The plan builds in a timed benchmark on deal four specifically for this purpose.

4. Baseline & Pacing Note

You already know what a hallucination is and why it matters. You understand token context well enough to know that what you put into a prompt shapes what comes out. You have used Perplexity for research and ChatGPT occasionally, which means you have a working intuition for where AI outputs are reliable and where they need verification. That is a solid foundation.

What this plan builds on top of that foundation: structured prompts that produce consistent, reviewable outputs rather than freeform responses you have to reshape every time. The difference between a useful AI research leg and an unreliable one is almost entirely in prompt structure — specifically, whether you have told the model what format to produce, what sources to cite, and what to flag as uncertain rather than assert as fact.

Your comfort levels on research and writing with AI are both at 3 — competent but not yet systematic. The goal for four weeks is to move those to 4: not by doing more, but by doing the same things with more deliberate prompts and a repeatable workflow. Building and automation stay low for now, which is exactly right given the constraint on proprietary data and the time budget. Two 45-minute sessions per week is a tight but workable schedule if the sessions are focused. Each session in this plan has a single deliverable and a clear done-when criterion — no session ends with "explore more."

5. Tasks & Projects

Current tasks: where AI plugs in now

Competitive landscape research is the highest-leverage immediate application. This step is almost entirely public-information work — who is in the market, what they have raised, how they are positioned, what signals of momentum or stress are visible. It is also the step that currently takes the most clock time relative to the judgment it requires. The competitive landscape scan prompt in your content assets is ready to use today in Claude.ai or Perplexity Pro. Use Perplexity when you want cited sources you can click through; use Claude when you want a synthesized narrative from notes you have already gathered.

One quality check to build in from the start: after every AI-generated competitive landscape, manually verify the funding figures for the top two or three players. AI training data has a cutoff, and funding rounds are exactly the kind of fact that goes stale. This check takes five minutes and will catch the most common class of error.

Diligence memo writing is the second high-leverage application. The memo scaffold prompt produces a structured first draft from your public-source notes. The key discipline: paste only public information into Claude.ai. Your call notes, anything shared under NDA, your own thesis reasoning — none of that goes in. What goes in is the same material you would have pulled from Crunchbase, the company website, LinkedIn, and press coverage. The AI produces a structured skeleton; you layer in the judgment, the proprietary signals, and the thesis fit.

Initial founder conversations and screening stay entirely human. No AI in the loop during or immediately after the call. Your call notes are proprietary from the moment you write them.

Upcoming project: five deals from the recent demo day

This batch is the practice ground for the entire four-week plan. Structure it as follows:

• Deals 1–3: Use these to develop and refine the workflow. Expect to spend more time than your baseline on deals 1 and 2 — you are building the process, not just running it. Deal 3 should start to feel faster.
• Deal 4: The benchmark. Run the full AI-assisted workflow with a timer. This is where you measure the 40% target.
• Deal 5: Reserve for after the four weeks if the batch runs long, or use it as a stretch session if the workflow clicks early.

The AI-in/out framework — your durable deliverable

By Week 3, Session 5, you will have enough data from three deals to draft the one-page framework that captures which steps are AI-eligible and which are judgment-only. This is not a generic framework — it is specific to your diligence process, your fund thesis, and your quality standards. The fill-in template in your content assets gives you the structure; the deals give you the evidence.

OpenBrain artifact binding (Week 3): The AI-in/out framework, once drafted in Session 5, lives in OpenBrain — not as a file on your laptop, but as a named memory queryable from any Claude surface. By Week 3, you will be able to open Claude on your MacBook or iPad, ask "What does my AI diligence framework say about competitive landscape research?" and get your own framework back. This is the outcome you named — "wire AI-assisted research into the standard diligence workflow" — made durable and portable. The framework does not live in a Notion doc you have to find; it travels with you.

Prompt discipline: the one habit that compounds

The single most important habit to build in four weeks is this: before you run any AI research step, state explicitly in the prompt what you want the model to flag as uncertain. "Flag any funding figures you are not confident are current" and "note when you are inferring rather than stating a known fact" are the two instructions that separate useful AI research output from dangerous AI research output. Both are built into the ready-to-paste prompts in your content assets. Once this becomes automatic, your verification time drops because you are only checking the flagged items, not re-reading the entire output.

6. Setup & Constraints

Devices: MacBook Pro is your primary working surface for Claude Desktop and OpenBrain setup. iPad works well for reviewing AI-generated drafts and running quick Perplexity searches — Claude.ai is fully functional in Safari on iPad, so your prompt library is accessible there too.

Accounts: Claude.ai Pro is the right tool for the memo drafting and synthesis work. Perplexity Pro handles the cited-source research leg. Notion stays as your document home — nothing in this plan requires changing that.

The proprietary data constraint is the structural rule for this entire plan. The workflow is designed around it: public-information steps go to Claude.ai or Perplexity, proprietary information never leaves your local environment. When you are in doubt about whether something is safe to paste, the answer is no. This is not a limitation — it is the boundary that makes AI usable in a fiduciary context.

On running AI locally: Because your constraint prohibits proprietary deal information from entering third-party LLMs, it is worth knowing that local AI is an option for the steps where you want to work with sensitive materials. If that use case ever becomes relevant:

• Osaurus (https://github.com/osaurus-ai/osaurus) is a Mac-native local server optimized for Apple Silicon — runs fast on your MacBook Pro, OpenAI-compatible API, no data leaves the machine.
• Ollama (https://ollama.com) is the cross-platform standard for local AI — works on Mac, Linux, and Windows, broad model support.

For the current four-week plan, the public/proprietary split handles the constraint without requiring local setup. But if you ever want to run AI against NDA materials or call notes, local is the path.

7. OpenBrain — Foundational Setup

OpenBrain is a personal knowledge layer that travels with you across Claude.ai, Claude Desktop, Claude mobile, and any other Claude surface you use. Think of it as a persistent memory for your working context — your diligence workflow, your prompt library, your in/out framework — that you set up once and then draw on in every subsequent session without re-explaining yourself. It is designed to be set up by a non-technical user in under 30 minutes.

Setup on your MacBook Pro:

1. Install Claude Desktop from claude.ai/download if you have not already.
2. Go to https://github.com/NateBJones-Projects/OB1 and follow the README — it walks through adding the OpenBrain MCP server to Claude Desktop. The process involves editing one configuration file; the README shows exactly what to paste.
3. Once installed, open Claude Desktop and verify the OpenBrain tools are available by typing "What tools do you have access to?" — you should see OpenBrain listed.
4. Your first memory to store: the diligence workflow map you will produce in Session 1. Paste it in and ask Claude to "remember this as my standard diligence workflow."

Three usage patterns for the investor/analyst track:

• Competitive landscape context: After running a landscape scan on a new deal, ask Claude to "learn this competitive map for [sector]." Future sessions in the same sector can start by asking "What do I already know about the [sector] competitive landscape?" — you are building a cumulative knowledge base, not starting from scratch each time.
• Prompt library retrieval: Store your three ready-to-paste prompts in OpenBrain by name. On your iPad, instead of finding a saved document, you ask Claude "Give me my competitive landscape scan prompt" and it returns it immediately.
• Framework evolution: Each time you update the AI-in/out framework after a new deal, save the revision to OpenBrain. Ask "What does my current AI diligence framework say about [step]?" to retrieve the latest version from any device.

The learn session loop: At the end of any working session where you made a meaningful decision, refined a prompt, or updated your framework, close the session by asking Claude to "learn this session." Describe what you did and what you concluded in two or three sentences. This builds a running log of your process evolution that future sessions can draw on — so by Week 4, Claude has context on everything you have learned across the four weeks, not just the current conversation.

8. Resources & Links

5 minutes

OpenBrain setup README https://github.com/NateBJones-Projects/OB1 Start here before Session 3. The README is short and the setup is a single configuration file edit.

Anthropic prompt engineering overview https://docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/overview Skim the "be clear and direct" and "use structured output" sections — these two principles account for most of the quality difference between useful and unreliable AI research output.

30 minutes

Perplexity Pro for investment research — search operator guide https://www.perplexity.ai/hub/blog/perplexity-pro Understanding how to scope Perplexity searches to recent results and specific source types (academic, news, company sites) meaningfully improves the quality of competitive landscape scans. Worth 20 minutes before Session 2.

Anthropic model documentation — Claude capabilities and context window https://docs.anthropic.com/en/docs/about-claude/models/overview Relevant specifically for understanding how much you can safely paste into a single session — useful when you are deciding whether to run a multi-company landscape scan in one prompt or split it.

Deep dive

Osaurus — Mac-native local AI server (Apple Silicon) https://github.com/osaurus-ai/osaurus Relevant if you ever want to run AI against NDA materials or proprietary deal notes without any data leaving your machine. Apple Silicon optimized — fast on your MacBook Pro. Not required for the current plan, but worth bookmarking.

Ollama — cross-platform local AI https://ollama.com The cross-platform standard for running AI models locally. Works on Mac, Linux, and Windows. Broad model support. Same use case as Osaurus — local processing of sensitive materials — with wider platform coverage.

9. Closing / Signals from the Keeper

Four weeks is enough time to know whether AI belongs in your diligence process — and exactly where. The plan is designed so that by the end of Week 4, you have evidence, not impressions: three memos, a timed benchmark, and a framework that reflects your actual experience rather than a vendor's promise.

The one thing to watch for: the pull toward doing more with AI than the workflow calls for. The value in this plan is not in maximizing AI involvement — it is in finding the precise steps where AI compresses time without introducing noise you have to spend time correcting. That line is different for every investor and every fund. The framework you build in Sessions 5 and 8 is how you make your line explicit.

The path exists. The first session starts with a blank workflow map and ends with a marked-up one. That is enough for week one.

Start with Session 1 this week.