A finance agent should not predict.
All briefs
Wall Street Prompt · LinkedIn one-page
Build An AI Research Agent Without Making It A Stock Picker
LinkedIn one-page
A finance agent should not predict. It should retrieve, verify, challenge, and escalate to a human.
LinkedIn1080 × 1350
References & validationValidated
Grounded in a real captured creator post (linked below).
Reference creator: AI Automation / Uppit AI (YouTube)
Post content
Dense one-page content — sections, tables and frameworks
The Mistake Most People Build First
They wire an LLM to a price feed and ask it what to buy. That is not a research agent. That is a slot machine with a citation problem.
| Common Build | Failure Mode | Disciplined Build |
|---|---|---|
| Agent outputs a buy/sell call | Confident guess with no provenance | Agent outputs evidence + open questions |
| Predicts next quarter's number | Hallucinated precision, no error bars | Extracts reported numbers, flags estimates |
| Summarizes one source fast | Single-source bias, stale data | Cross-checks 2+ filings before asserting |
| Optimizes for a clean answer | Hides uncertainty to look smart | Surfaces conflicts and forces escalation |
Why it matters: a model rewarded for sounding decisive will fabricate decisiveness. You want it rewarded for being checkable.
The Four Jobs (Retrieve, Verify, Challenge, Escalate)
Scope the agent to four jobs. None of them is prediction. Each job has a hard boundary and a defined output.
| Stage | What It Does | Hard Boundary |
|---|---|---|
| Retrieve | Pull 10-Ks, 10-Qs, earnings call transcripts, comps tables into context | Cite source + filing date on every fact |
| Verify | Confirm each number against the primary filing, not a summary | No claim ships without a matched source line |
| Challenge | Argue the bear case against its own summary | Must list 3 ways the thesis breaks |
| Escalate | Route conflicts, gaps, and judgment calls to a human | Never resolves a valuation call on its own |
Key: each stage narrows what the agent is allowed to say. The agent assembles the file. The human signs the IC memo.
The Verification Layer (Where Trust Is Earned)
Retrieval without verification is just faster hallucination. Build a checking pass that runs before any output reaches you.
- Every numeric claim links to a filing line, page, and date. No link, no claim.
- Two-source rule for any figure that moves a thesis (revenue, margin, debt, share count).
- Recency gate: if the newest source is older than the last earnings call, the agent flags it as stale.
- Unit and period checks: TTM vs fiscal year, GAAP vs adjusted, gross vs net. Mismatches get raised, not smoothed over.
Heads up: most bad agent output is not a lie, it is a mixed-period comparison stated as fact. Catch it in the verify pass.
Challenge And Escalate (Designing The Friction)
A good research agent makes its own case harder, not easier. Bake in disagreement and a clean handoff.
| Mechanism | Agent Action | Human Action |
|---|---|---|
| Red-team pass | Writes the strongest bear case from the same filings | Decides if the bear case is disqualifying |
| Confidence tag | Labels each claim: Reported / Estimated / Inferred | Discounts Inferred claims by default |
| Conflict ticket | Opens a flag when two sources disagree | Resolves which source governs |
| Judgment escalation | Stops at valuation, sizing, timing | Owns the call, the memo, and the risk |
Why it matters: Human judges. AI builds. The moment the agent decides position sizing, you have rebuilt the thing you were trying to avoid.
Scope Guardrails (What It Must Never Do)
Write these as hard rules in the system prompt and the output schema. Treat a violation as a failed run, not a quirk.
| Banned Output | Allowed Substitute |
|---|---|
| Buy / sell / hold rating | Structured evidence summary with sources |
| Price target or fair value number | Range of reported comps and the math behind them |
| Predicted return or timing call | List of catalysts and the dates they resolve |
| Personalized recommendation | Standardized research file for human review |
| Confident single-source claim | Claim tagged with source count and recency |
Tip: enforce this in the schema, not just the prompt. If the field for a rating does not exist, the model cannot fill it.
Caption
LinkedIn post copy
Visual design notes
- Near-black forest-green background (#0B1410 to #0E1A14 gradient). ONE green accent only (#1FB67A) for the title keyword, table header rules, and stage numbers. No second accent color.
- Heavy condensed sans headline (e.g. Anton or Archivo Black) for the title; emphasize the word 'Research' in the green accent, rest in off-white.
- Teal subtitle banner: a single thin horizontal bar under the title holding the one-liner, left-aligned, small-caps tracking.
- Dense 2 and 3-column tables with hairline row dividers in muted green, alternating row tint at 4% opacity. Left-aligned text, monospace numerals where figures appear.
- Section headings numbered 01-05 in the green accent with a short vertical tick mark to the left, creating a clear scan rhythm.
- Small horizontal flow diagram near the top or between sections: Retrieve > Verify > Challenge > Escalate as four linked nodes, arrows in accent green, human icon at the Escalate node.
- 'Why it matters' / 'Key' / 'Tip' notes set in a thin-bordered callout box, italic-light, slightly indented, never colored fill so they stay quiet under the tables.
- Generous left margin, tight top-to-bottom packing. Footer pinned to a thin baseline rule at the bottom edge. Everything left-aligned, no centered text.
Production checklist
- ☐Design the 1080x1350 one-pager in the WSP template: near-black forest bg, single green accent, condensed headline with 'Research' emphasized, teal subtitle banner.
- ☐Build the four data tables (Mistake/Failure/Disciplined, Four Jobs, Challenge-Escalate, Banned/Allowed) with hairline dividers, 4% row tint, and monospace numerals.
- ☐Draw the Retrieve > Verify > Challenge > Escalate flow diagram with a human icon at the Escalate node in accent green.
- ☐Add numbered section headers 01-05, 'Why it matters/Key/Tip' callout boxes, and verify zero em dashes and zero buy/sell/hold language across all copy.
- ☐Add header and footer branding (Dave Wang · Wall Street Prompt) and the 'Educational only. Not investment advice.' line.
- ☐Export PNG at 1080x1350 for the feed and a matching PDF for DM delivery; spell-check and confirm every table fact reads as evidence, not advice.
Follow Dave Wang · Wall Street Prompt · For more AI investing workflows · COMMENT & REPOST
CTA
Save this before you wire your next agent to a price feed. Then comment RESEARCH and I'll send the one-page verification-layer checklist.