Persona-modeled investor reviewer — deep-research → deck-generator → reviewer loop

Three-agent loop for any deliverable that needs investor-grade review. Agent 1 — Deep-research: ingests the target market, competitive landscape, and unit economics, producing a ~100k-word raw research dump. Agent 2 — Deck-generator: uses claude.ai/design to turn the research into a 4- or 8-slide deck with multiple internal self-validation passes (numbers cross-check, source attribution). Agent 3 — Persona-modeled reviewer: a synthesized "personality" built from public artifacts of the specific investor the deck is going to — their podcasts, tweets, LinkedIn posts, blog, conference talks. The reviewer reads the generated deck and asks the uncomfortable questions that investor is known for. The generator iterates against the reviewer until the reviewer approves. Reported result: usable shipped output by iteration 15.

The mechanic is not "AI feedback on your deck" — that already exists and is generic. The signal is persona-modeled adversarial review: the agent isn't asking generic-VC questions, it's asking the questions the specific partner you're pitching asks, in their tone, against their actual decision criteria, with their format preferences. The output is a deck that survives the specific meeting you're walking into, not one that survives "a VC review" in the abstract. The same pattern generalizes to client deliverables — synthesize the client's CMO, run them as the reviewer, save the rework cycle.

1. Pick a target investor (or reviewer persona) deliberately. The whole loop is calibrated to one human. Don't do this for "VCs in general" — do it for the person on the calendar.
2. Scrape their public artifacts. Podcasts, conference talks (transcripts via YouTube or Granola), tweets, LinkedIn posts, blog. Anything where they discuss what they look for in a deck, decision criteria, what makes them say no.
3. Synthesize the persona. Compress the corpus into a reviewer spec: what they care about, what they ask, their decision criteria, their format preferences, the kinds of numbers that trigger skepticism. This becomes the system prompt for the reviewer agent.
4. Wire the 3-agent loop.

• Deep-research agent → ~100k-word raw data dump (market sizing, competitive map, unit economics, regulatory context). Enrich with Statista, SimilarWeb, Sensortower, AppMagic, Google Ads search-volume API where the category supports it.
• Deck-generator agent → uses claude.ai/design for layout; runs internal self-validation passes on numbers (catches multiplication errors before the reviewer does); produces a 4-slide exec version + an 8-slide full version.
• Persona-reviewer agent → reads the deck, asks the uncomfortable questions, returns a pass/fail with specific gaps.

5. Iterate to convergence. One operator reports 15 iterations before the reviewer stopped flagging gaps. Each iteration: reviewer flags → generator revises → re-submit. Don't shortcut: the early iterations catch the obvious gaps; the late iterations catch the ones that would have killed the meeting.
6. Run a dedicated data-validation test suite separately. An LLM-generated deck will at some point ship a multiplication error (one operator caught a 180×240 arithmetic mistake in a deck — "anyone who notices this once never trusts the calculation again"). Build a deterministic numerical-consistency check that runs on the final draft. The generator's internal self-validation is necessary but not sufficient.
7. Optional: enrich with external data APIs. Statista, SimilarWeb, Sensortower, AppMagic for market sizing; Google Ads search-volume API for category demand signals. These are not in the base loop — add them when the reviewer's questions keep landing on "show me category traction".

Persona-modeling kernel — the rationale behind the reviewer:

Theoretically what I built is a synthesis of a real person's personality
from publicly available data — understanding what's important to them,
what they look at in a pitch deck. Everyone has their own particularities
in how they process data, so I tried to adapt, to personalize the
presentation for a specific person. A bit of social engineering, but with
good intent.

Adversarial-reviewer pattern+ #12698 (one operator's earlier version uses ChatGPT directly for the same effect):

"I usually ask ChatGPT to write a prompt for a strict investor and have it
 ask hard questions that founders are afraid of and find difficult to
 answer. Then I send the prompt with my deck and let it think for ~20
 minutes — usually enough."

System-prompt skeleton for the persona-reviewer agent:

You are <Reviewer-Persona-Name>, a synthesized model of an investor based on
public artifacts. Your decision style:

- <criterion 1, extracted from their podcasts/posts — e.g. "always asks about
 retention before revenue">
- <criterion 2 — e.g. "skeptical of TAM slides; wants serviceable obtainable
 market">
- <criterion 3 — e.g. "rejects decks that lead with 'AI-powered' as a verb">
- <format preference — e.g. "8 slides max; one chart per slide">
- <numerical-skepticism patterns — e.g. "will multiply revenue × retention ×
 CAC payback in their head and reject if math is off">

Your job is NOT to write the deck. Your job is to read it as <Reviewer> would
and ask the 3-5 questions that would derail the meeting. Be specific. Cite
the slide. Where a number doesn't reconcile, point to which numbers.

Return:
1. PASS / NEEDS_REWORK
2. Specific gaps (slide-cited)
3. The single question that, if unanswered, kills the deal

• LLM-generated decks read as "no soul, too many numbers". Multiple reviewers in source flag the same failure mode: generated decks feel like research dumps, not investor pitches. The persona-reviewer catches some of this if you've trained on a reviewer who values story arc — but not all. Have a human do the final story-arc pass after the agent loop converges.
• The "I'll prove I won't lose your money" vibe is a real failure mode. One reviewer's specific critique on a generated deck: it reads as defensive ("see, I won't lose your money") rather than offensive ("here's why this becomes a $1B outcome"). If your persona-reviewer doesn't have this critique in their criteria, generated decks will fall into it.
• At pre-seed, the deck barely matters — they fund people and traction. One reviewer's caveat: "below $3-4M, one check, it doesn't matter what deck you bring; they fund the person, then traction." Don't over-optimize the deck loop for a stage where the deck isn't the decision artifact.
• Numerical errors compound trust collapse. A multiplication mistake (180 × 240 = the wrong number) in one of the iterations destroys confidence in every number on every slide. The data-validation test suite is not optional once you're at investor-grade output.
• Don't run the reviewer on a generic-VC persona. The whole value is specificity. Generic-VC reviewers produce generic-VC feedback — you could have done that yourself in a chat window.