Research
Earnings summaries, precedent search, and comp screening with source-citation discipline, so the model shows its work instead of asserting figures from nowhere.
Armory · Prompt Vault
Prompts that survive contact with a live model.
Battle-tested prompts for real IB tasks, comps, precedents, CIM extraction, IC-memo drafts, each one shipped with the failure modes it is designed to dodge.
Why it matters
A prompt is an instruction set, not a magic word.
On a live model the difference between a good instruction set and a careless one is a comps page you can defend versus one quietly seeded with a hallucinated figure. Every prompt inside has already been refined on the exact task you are doing, and every prompt names the specific way the model will try to fool you.
The return is measurable in hours: document extraction runs at roughly 90 percent or better on standard filings, and pure document review alone can free more than 200 hours per analyst per year. The Vault is how you capture that time instead of leaving it on the table.
Source: Knowledge / AI Capability Assessment 2026 · Perplexity research, 2026
/Skip the trial and error.
Start from a prompt that already works on the task in front of you, not a guess you debug under deadline.
/Get the catch baked in.
Each prompt names what the model tends to get wrong on that task, so you review with your eyes open.
/Build a compounding asset.
Save, adapt, and reuse. A maintained prompt library grows in value instead of being a one-off.
/Stay the operator the tools need.
The model drafts at speed; you frame, direct, and decide. That is the role banks now screen for.
What's inside
The four moves you make on almost every task.
Extraction
Turn a 10-K or CIM into structured Excel. The prompt pins down format, units, and edge cases and forces a confidence flag on anything ambiguous.
Drafting
Pitchbook sections and IC-memo first drafts. The model scaffolds structure; you own the thesis, the risk weighting, and the narrative.
Verification
Prompts that make the model check its own work and surface its assumptions, because confident hallucination is real. A verification pass is not optional.
Proof and credibility
The leading models hallucinate figures, break on circular references, misapply GAAP and IFRS, and stay blind to EBITDA normalization. None of that is a reason to avoid the tools; it is the reason to use disciplined prompts. Every prompt is written to constrain exactly those failure modes: leverage with a seatbelt.
Source: Knowledge / AI Capability Assessment 2026
Straight answers
Before you ask.
Can't I just write my own prompts?
You can, and should keep getting better. What the Vault saves is the iteration tax: each prompt encodes lessons learned on real tasks, so you start at version ten instead of version one.
Will prompts go stale as models change?
Model behavior shifts quietly, which is why the Vault is maintained and versioned. When behavior moves, affected prompts get revisited, and you see the version and date on every card.
Is it safe with deal data?
Each prompt states its data-handling assumptions. The hard rule: never route raw MNPI through public tools without the right controls. We flag the boundary on every prompt that touches sensitive material.
Free vs paid?
Students get free sample prompts. The full, maintained, versioned Vault is part of the professional subscription, which keeps the student tier free.
